I

Data Storage
 Data Storage

Edited by
Prof. Florin Balasa

In-Tech
intechweb.org
Published by In-Teh

In-Teh
Olajnica 19/2, 32000 Vukovar, Croatia

Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in any
publication of which they are an author or editor, and the make other personal use of the work.

© 2010 In-teh
www.intechweb.org
Additional copies can be obtained from:
[email protected]

First published April 2010

Printed in India

Technical Editor: Maja Jakobovic

Cover designed by Dino Smrekar

Data Storage,
Edited by Prof. Florin Balasa

p. cm.
ISBN 978-953-307-063-6
 V

Preface

Many different forms of storage, based on various natural phenomena, has been invented.
So far, no practical universal storage medium exists, and all forms of storage have some
drawbacks. Therefore, a computer system usually contains several kinds of storage, each with
an individual purpose.
Traditionally, the most important part of a digital computer is the central processing unit
(CPU or, simply, a processor), because it actually operates on data, performs calculations, and
controls all the other components. Without a memory, a computer would merely be able to
perform fixed operations and immediately output the result. It would have to be reconfigured
to change its behavior. This is acceptable for devices such as desk calculators or simple digital
signal processors. Von Neumann machines differ in that they have a memory in which they
store their operating instructions and data. Such computers are more versatile in that they
do not need to have their hardware reconfigured for each new program, but can simply be
reprogrammed with new in-memory instructions. Most modern computers are von Neumann
machines.
In practice, almost all computers use a variety of memory types, organized in a storage
hierarchy around the CPU, as a trade-off between performance and cost. Generally, the lower
a storage is in the hierarchy, the lesser its bandwidth (the amount of transferred data per time
unit) and the greater its latency (the time to access a particular storage location) is from the
CPU. This traditional division of storage to primary, secondary, tertiary and off-line storage
is also guided by cost per bit.
Primary storage (or main memory, or internal memory), often referred to simply as memory,
is the only one directly accessible to the CPU. The CPU continuously reads instructions stored
there and executes them as required. Any data actively operated on is also stored there in
uniform manner.
As the random-access memory (RAM) types used for primary storage are volatile (i.e., they
lose the information when not powered), a computer containing only such storage would not
have a source to read instructions from, in order to start the computer. Hence, non-volatile
primary storage containing a small startup program is used to bootstrap the computer, that is,
to read a larger program from non-volatile secondary storage to RAM and start to execute it.
Secondary storage (or external memory) differs from primary storage in that it is not directly
accessible by the CPU. The computer usually uses its input/output channels to access
secondary storage and transfers the desired data using intermediate area in the primary
storage. The secondary storage does not lose the data when the device is powered down – it is
VI

non-volatile. Per unit, it is typically also two orders of magnitude less expensive than primary
storage. In modern computers, hard disk drives are usually used as secondary storage. The
time taken to access a given byte of information stored on a hard disk is typically a few
milliseconds. By contrast, the time taken to access a given byte of information stored in a
RAM is measured in nanoseconds. This illustrates the very significant access-time difference
which distinguishes solid-state memory from rotating magnetic storage devices: hard disks
are typically about a million times slower than memory. Rotating optical storage devices, such
as CD and DVD drives, have even longer access times. Some other examples of secondary
storage technologies are: flash memory (e.g. USB flash drives), floppy disks, magnetic tape,
paper tape, punched cards, stand-alone RAM disks, and Zip drives.
Tertiary storage or tertiary memory provides a third level of storage. Typically it involves
a robotic mechanism which will mount (insert) and dismount removable mass storage
media into a storage device according to the system’s demands; this data is often copied to
secondary storage before use. It is primarily used for archival of rarely accessed information
since it is much slower than secondary storage (e.g. tens of seconds). This is primarily useful
for extraordinarily large data stores, accessed without human operators. Typical examples
include tape libraries and optical jukeboxes.
Off-line storage is a computer data storage on a medium or a device that is not under the
control of a processing unit. In modern personal computers, most secondary and tertiary
storage media are also used for off-line storage. Optical discs and flash memory devices are
most popular, while in enterprise uses, magnetic tape is predominant. This book presents
several advances in different research areas related to data storage, from the design of a
hierarchical memory subsystem in embedded signal processing systems for data-intensive
applications, to data representation in flash memories, to the data recording and retrieval
in conventional optical data storage systems and the more recent holographic systems, to
applications in medicine requiring massive image databases.
In optical storage systems, sensitive stored patterns can cause failure in data retrieval and
decrease the system reliability. Modulation codes play the role of shaping the characteristics
of stored data patterns. In conventional optical data storage systems, information is recorded
in a one-dimensional spiral stream. The major concern of modulation codes for these systems
is to separate the binary 1’s by a number of binary 0’s.
The holographic data storage systems are regarded as the next-generation optical data
storage due to an extremely high capacity and ultra-fast transfer rate. In holographic systems,
information is stored as pixels on two-dimensional pages. Different from the conventional
optical data storage, the additional dimension inevitably brings new considerations to the
design of modulation codes. The primary concern is that interferences between pixels are
omni-directional. Moreover, interferences between pixels are imbalanced: since pixels carry
different intensities to represent different bits of information, pixels with higher intensities
tend to corrupt the signal fidelity of pixels with lower intensities more than the other way
around.
Chapter 1 analyzes different modulation codes for optical data storage. It first addresses types
of constraints of modulation codes. Afterwards, the one-dimensional modulation codes,
adopted in the current optical storage systems (like EFM for CD’s, EFMPlus for DVD’s, 17PP
for Blu-ray discs), are presented. Next, the chapter focuses on two-dimensional modulation
codes for holographic data storage systems – the block codes and the strip codes. It further
 VII

discusses the advantages and disadvantages of variable-length modulation codes in contrast

to fixed-length modulation codes.
Chapter 2 continues the discussion on holographic data storage systems, giving an overview
on the processing of retrieved signals. Even with modulation codes, two channel defects
– inter-pixel interferences and misalignments – are major causes for degradation of the
retrieved image and, thus, degradation in detection performance. Misalignments severely
distort the retrieved image and several solutions for compensating misalignment – itera- 2
tive cancellation by decision feedback, oversampling with re-sampling, interpolation with
rate conversion – are presented in the chapter. Equalization and detection are the signal
processing operations for the final data detection in the holographic data storage reading
procedure, restoring the stored information from the interferenceinflicted signal. In contrast
to the classical de-convolution method based on linear minimum mean squared error
(LMMSE) equalization, that suffers from a model mismatch due to the inherent nonlinearity
of the channel, the chapter presents two nonlinear detection algorithms that achieve a better
performance than the classical LMMSE at the cost of a higher complexity.
Chapter 3 describes recent advances towards three-dimensional (3D) optical data storage
systems. One of the methods for 3D optical data storage is based on volume holography.
The physical mechanism is photochromism, which is a reversible transformation of a single
chemical species between two states having different absorption spectra and refractive
indices. This allows for holographic multiplexing recording and reading. Another technique
for 3D optical data storage is the bit-by-bit memory at the nanoscale, which is based on the
confinement of multiphoton absorption to a very small volume because of its nonlinear
dependence of excitation intensity.
Both organic and anorganic materials are convenient for 3D data storage. These materials
must satisfy several norms for storing and reading: resistance to aging due to temperature
and recurrent reading, high-speed response for high-rate data transfer, no optically scattering
for multilayer storage. The chapter presents experiments with two particular media: a
photosensitive (zinc phosphate) glass containing silver and a spin state transition material.
Flash memories have become the dominating member in the family of non-volatile memories.
Compared to magnetic and optical recording, flash memories are more suitable for mobile,
embedded, and mass-storage applications. The reasons include their high speed, physical
robustness, and easy integration with other circuits. In a flash memory, cells (floating-
gate transistors) are organized into blocks. While it is relatively easy to inject charge into
a cell (operation called writing or programming), to remove charge from a cell (operation
called erasing), the whole block containing it must be erased to the ground level and then
reprogrammed. This block erasure operation not only significantly reduces the speed, but
also reduces the lifetime of the flash memory.
Chapter 4 analyzes coding schemes for rewriting data in flash memories. The interest to this
problem comes from the fact that if data are stored in the conventional way, even to change
one bit in the data may necessitate to lower some cell’s charge level, which would lead to the
costly block erasure operation. This study discusses the Write Asymmetric Memory (WAM)
model for flash memories, where the cells’ charge levels are changing monotonically before
each erasure operation. The chapter also presents a data representation scheme called Rank
Modulation whose aim is to eliminate both the problem of overshooting (injecting a higher
VIII

charge level than desired) while programming cells, and also to tolerate better asymmetric
shifts of the cells’ charge levels.
Chapter 5 deals with data compression which is becoming an essential component of
high-speed data communication and storage. Lossless data compression is the process of
encoding (“compressing”) a body of data into a smaller body of data, which can, at a later
time, be uniquely decoded (“decompressed”) back to the original data. In contrast, lossy data
compression yields by decompression only some approximation of the original data. Several
lossless data compression techniques have been proposed in the past – starting with Huffman
code. This chapter focuses on a more recent lossless compression approach – the Lempel-Ziv
(LZ) algorithm – whose princi- ple is to find the longest match between a recently received
string stored in the input buffer and an incoming string; once the match is located, the
incoming string is represented with a position tag and a length variable, linking it to the old
existing one, thus achieving a more concise representation than the input data. The chapter
presents an area- and speed-efficient systolic array implementation of the LZ compression
algorithm. The systolic array can operate at a higher clock rate than other architectures (due
to the nearest-neighbor communication) and can be easily implemented and tested due to
regularity and homogeneity.
Although the CPU performance has considerably improved, the speed of file system
management of huge information is considered as the main factor that affects computer system
performance. The I/O bandwidth is limited by magnetic disks whose rotation speed and seek
time has improved very slowly (although the capacity and cost per megabyte has increased
much faster). Chapter 6 discusses problems related to the reliability of a redundant array of
independent disks (RAID) system, which is a generally used solution since 1988 referring to
the parallel access of data separated on several disks. A RAID system can be configured in
various ways to get a fair compromise between data access speed, system reliability, and size
of storage. The general trade-off is to increase data access speed by writing the data into more
places, hence increasing the amount of storage. On the other hand, more disks cause a lower
reliability: this, together with the data redundancy, cause a need for additional algorithms
to enhance the reliability of valuable data. The chapter presents recent solutions for the use
of Reed- Solomon code in a RAID system in order to correct single random error and double
erasures. The proposed RAID system is expandable in terms of correction capabilities and
presents an integrated concept: the modules at the disk level mainly deal with burst or random
errors in disks, while the control level does the corrections for multiple failures of the system.
A grid is a collection of computers and storage resources, addressing collaboration, data
sharing, and other patterns of interaction that involve distributed resources. Since the grid
typically consists nowadays of hundreds or even thousands of storage and computing nodes,
key challenges faced by high-performance storage systems are manageability, scalable
administration, and monitoring of system state. Chapter 7 presents a scalable distributed
system consisting of an administration module that manages virtual storage resources
according to their workloads, based on the information collected by a monitoring module.
Since the major concern in a conventional data storage distributed system is the data access
performance, the focus was on static performance parameters not related to the nodes load.
In contrast, the current system is designed based on the principle that the whole system’s
performance is affected not only by the behavior of each individual application, but also by
the execution of different applications combined together. The new system takes into account
the utilization percentage of system resources, like CPU load, disk load, network load. It
 IX

offers a flexible and simple model that collects information on the nodes’ state and uses its
monitoring knowledge together with a prediction model to efficiently place data during
runtime execution in order to improve the overall data access performance.
Many multidimensional signal processing systems, particularly in the areas of multimedia
and telecommunications, are synthesized to execute data-intensive applications, the data
transfer and storage having a significant impact on both the system performance and the
major cost parameters – power and area. In particular, the memory subsystem is, typically,
a major contributor to the overall energy budget of the entire system. The dynamic energy
consumption is caused by memory accesses, whereas the static energy consumption is due
to leakage currents. Savings of dynamic energy can be potentially obtained by accessing
frequently used data from smaller on-chip memories rather than from the large off-chip main
memory, the problem being how to optimally assign the data to the memory layers. As on-
chip storage, the scratch-pad memories (SPM’s) – compiler-controlled static RAM’s, more
energy-efficient than the hardware-managed caches – are widely used in embedded systems,
where caches incur a significant penalty in aspects like area cost, energy consumption, and hit
latency. Chapter 8 presents a power-aware memory allocation methodology. Starting from
the high-level behavioral specification of a given application, where the code is organized
in sequences of loop nests and the main data structures are multidimensional arrays, this
framework performs the assignment of the multidimensional signals to the memory
layers – the on-chip scratch-pad memory and the off-chip main memory – the goal being
the reduction of the dynamic energy consumption in the memory subsystem. Based on the
assignment results, the framework subsequently performs the mapping of signals into the
memory layers such that the overall amount of data storage be reduced. This software system
yields a complete allocation solution: the exact storage amount on each memory layer, the
mapping functions that determine the exact locations for any array element (scalar signal)
in the specification, metrics of quality for the allocation solution, and also an estimation of
the dynamic energy consumption in the memory subsystem using the CACTI power model.
Network-on-a-Chip (NoC) is a new approach to design the communication subsystem
of System-on-a-Chip (SoC) – paradigm referring to the integration of all components of a
computer or other electronic system into a single integrated circuit, implementing digital,
analog, mixed-signal, and sometimes radio-frequency functions, on a single chip substrate.
In a NoC system, modules such as processor cores, memories, and specialized intellectual
property (IP) blocks exchange data using a network that brings notable improvements over
the conventional bus system. An NoC is constructed from multiple point-to-point data links
interconnected by temporary storage elements called switches or routers, such that messages
can be relayed from any source module to any destination module over several links, by
making routing decisions at the switches. The wires in the links of NoC are shared by many
signals, different from the traditional integrated circuits that have been designed with
dedicated point-to-point connections – with one wire dedicated to each signal. In this way,
a high level of parallelism is achieved, because all links in NoC can operate simultaneously
on different data packets. As the complexity of integrated circuits keeps growing, NoC
provides enhanced throughput and scalability in comparison with previous communication
architectures (dedicated signal wires, shared buses, segmented buses with bridges), reducing
also the dynamic power dissipation (as signal propagation in wires across the chip require
multiple clock cycles).
X

Chapter 9 discusses different trade-offs in the design of efficient NoC, including both
elements of the network – interconnects and storage elements. This chapter introduces a high-
throughput architecture, which is applied to different NoC topologies: butterfly fat-free (BFT),
a mesh interconnect topology called CLICH´E , Octagon, and SPIN. A power dissipation
analysis for all these high-throughput architectures is presented, followed by a discussion on
the throughput improvement and an overhead analysis.
Data acquisition and data storage are requirements for many applications in the area of sensor
networks. Chapter 10 discusses different protocols for interfacing smart sensors and mobile
devices with non-volatile memories.
A “smart sensor” is a transducer that provides functions to generate a correct representation
of a sensed or controlled quantity. Networks of smart sensors have an inbuilt ability to sense
and process information, and also to send selected information to external receivers (including
other sensors). The nodes of such networks – the smart sensors – require memory capabilities
in order to store data, either temporarily or permanently. The non-volatile memories (whose
content need not be periodically refreshed) used in the architecture of smart sensors include
all forms of read-only memories (ROM’s) – such as programmable read-only memories
(PROM’s), erasable programmable read-only memories (EPROM’s), electrically erasable
programmable read-only memories (EEPROM’s) – and flash memories. Sometimes, random
access memories powered by batteries are used as well.
The communication between the sensor processing unit and the non-volatile memory units is
done using different communications protocols. For instance, the 1-wire interface protocol –
developed by Dallas Semiconductor – permits digital communications through twisted-pair
cables and 1-wire components over a 1-wire network; the 2-wire interface protocol – developed
by Philips – performs communication functions between intelligent control devices, general-
purpose (including memories) and application-specific circuits along a bi-directional 2-wire
bus (called the Inter-Integrated Circuit or the I2C-bus).
The chapter also presents memory interface protocols for mobile devices, like the CompactFlash
(CF) memory protocol – introduced by SanDisk Corporation – used in flash memory card
applications where small form factor, low-power dissipation, and ease-of-design are crucial
considerations; or the Secure Digital (SD) memory protocol – the result of a collaboration
between Toshiba, SanDisk, and MEI (Panasonic) – specifically designed to meet the security,
capacity, performance, and environment requirements inherent to the newly-emerging audio
and video consumer electronic devices, as well as smart sensing networks (that include smart
mobile devices, such as smart phones and PDA’s).
The next chapters present complex applications from different fields where data storage plays
a significant part in the system implementation.
Chapter 11 presents a complex application of an array of sensors, called the electronic nose.
This system is used for gas analysis when exposed to a gas mixture and water vapour in
a wide range of temperatures. The large amount of data collected by the electronic nose is
able to provide high-level information, to make characterizations of different gases. The
electronic nose utilizes an efficient and affordable sensor array that shows autonomous and
intelligent capabilities. An application of this system is the separation of butane and propane
– (normally) gaseous hydrocarbons derived from natural gas or refinery gas streams. A neural
network with feed forward back propagation training algorithm is used to detect each gas
 XI

with different sensors. Fuzzy logic is also used because it enhances discrimination techniques
among sensed gases. A fuzzy controller is used to detect the concentration of each gas in the
mixture based on three parameters: temperature, output voltage of the microcontroller, and
the variable resistance related to each sensor.
With the steady progress of the technology of digital imaging and the rapid increase of
the data storage capacity, the capability to manipulate massive image databases over the
Internet is used nowadays both in clinical medicine – to assist diagnosis – and in medical
research and teaching. The traditional image has been replaced by Computer Radiography
or Digital Radiography derived imagery, Computed Tomography, Magnetic Resonance
Imaging (MRI), and Digital Subtraction Angiography. This advance reduces the space and
cost associated with X-ray films, speeds up hospital management procedures, improving
also the quality of the medical imagery-based diagnosis. Chapter 12 presents a complex
system where a grid-distributed visual medical knowledge and medical image database was
integrated with radiology reports and clinical information on patients in order to support
the diagnostic decision making, therapeutic strategies, and to reduce the human, legal, and
financing consequences of medical errors. This system started being used on the treatment of
dementia – a brain generative disease or a group of symptoms caused by various diseases and
conditions – to establish a prototype system integrating content-based image retrieval (CBIR)
and clinical information from electronic medical records (EMR).
Chapter 13 describes a data storage application from the food industry. The food legislation
in the European Union imposes strict regulations on food traceability in all the stages of
production, transformation, and distribution. The back or suppliers traceability refers to
the capability of having knowledge of the products that enter a food enterprise and their
origin and suppliers. The internal or process traceability refers to the information about what
is made, how and when it is made, and the identification of the product. The forward or
client traceability refers to the capability of knowing the products delivered, when and to
whom they have been supplied. The chapter describes a traceability system called the radio-
frequency identification (RFID). This is a contactless method for data transfer, carried out
through electromagnetic waves. RFID tags (transponders), consisting of a microchip and an
antenna, represent the physical support for storing the information required to perform a
complete traceability of the products, as well as facilitate the collection of data required by
the technical experts of regulatory boards in charge of quality certification. The RFID read/
write device creates a weak electromagnetic field. When a RFID tag passes through this field,
the microchip of the transponder wakes up and can send/receive data without any contact
to the reader. The communication gets interrupted when the tag leaves the field, but the data
on the tag remains stored.

Prof. Florin Balasa

Dept. Computer Science and Information Systems
Southern Utah University
XII
 XIII

Contents

Preface V

1. Modulation Codes for Optical Data Storage 001

Tzi-Dar Chiueh and Chi-Yun Chen

2. Signal Processing in Holographic Data Storage 018

Tzi-Dar Chiueh and Chi-Yun Chen

3. Optical data storage in photosensitive glasses and spin

state transition compounds 034
Matthieu Bellec, Lionel Canioni, Arnaud Royon, Bruno Bousquet,
Jérôme Degert and Eric Freysz

4. Data Representation for Flash Memories 053

Anxiao (Andrew) Jiang and Jehoshua Bruck

5. Design and Implementation of FPGA- based Systolic Array

for LZ Data Compression 075
Mohamed A. Abd El Ghany, Magdy A. El-Moursy and Aly E. Salama

6. Design of Simple and High Speed Scheme to Protect Mass Storages 093
Ming-Haw Jing, Yan-Haw Chen, Zih-Heng Chen,
Jian-Hong Chen and Yaotsu Chang

7. Administration and Monitoring Service for Storage Virtualization

in Grid Environments 107
Salam Traboulsi

8. Power-Aware Memory Allocation for Embedded

Data-Intensive Signal Processing Applications 117
Florin Balasa, Ilie I. Luican, Hongwei Zhu and Doru V. Nasui

9. High Throughput Architecture for High Performance NoC 133

Mohamed A. Abd El Ghany, Magdy A. El-Moursy and Mohammed

10. Non-volatile memory interface protocols for smart sensor

networks and mobile devices 157
Octavian Postolache, Pedro Silva Girão and José Miguel Dias Pereira
XIV

11. Electronic Nose System and Artificial Intelligent Techniques

for Gases Identification 176
Iman Morsi

12. Medical Image Intelligent Access Integrated with Electronic

Medical Records System for Brain Degenerative Disease 202
Mei-Ju Sua, Po-Hsun Chengb, Sao-Jie Chena, Chung-Yi Yangd,
Ping-Kung Yipf, Daniel Racoceanugh and Heng-Shuen Chence

13. Use of RFID tags for data storage on quality control

in cheese industries 214
Raquel Pérez-Aloe, José M. Valverde, Antonio Lara, Fernando Castaño,
Juan M. Carrillo, José González and Isidro Roa
Modulation Codes for Optical Data Storage 1

X1

Modulation Codes for Optical Data Storage

Tzi-Dar Chiueh and Chi-Yun Chen
National Taiwan University
Taipei, Taiwan

1. Introduction
In optical storage systems, sensitive stored patterns can cause failure in data retrieval and
decrease the system reliability. Modulation codes play the role of shaping the characteristics
of stored data patterns in optical storage systems. Among various optical storage systems,
holographic data storage is regarded as a promising candidate for next-generation optical
data storage due to its extremely high capacity and ultra-fast data transfer rate. In this
chapter we will cover modulation codes for optical data storage, especially on those
designed for holographic data storage.
In conventional optical data storage systems, information is recorded in a one-dimensional
spiral stream. The major concern of modulation codes for these optical data storage systems
is to separate binary ones by a number of binary zeroes, i.e., run-length-limited codes.
Examples are the eight-to-fourteen modulation (EFM) for CD (Immink et al., 1985), EFMPlus
for DVD (Immink, 1997), and 17 parity preserve-prohibit repeated minimum run-length
transition (17PP) for Blu-ray disc (Blu-ray Disc Association, 2006). Setting constraint on
minimum and maximum runs of binary zeros results in several advantages, including
increased data density, improved time recovery and gain control and depressed interference
between bits.
In holographic data storage systems, information is stored as pixels on two-dimensional (2-
D) pages. Different from conventional optical data storage, the additional dimension
inevitably brings new consideration to the design of modulation codes. The primary concern
is that interferences between pixels are omni-directional. Besides, since pixels carry different
intensities to represent different information bits, pixels with higher intensities intrinsically
corrupt the signal fidelity of those with lower intensities more than the other way around,
i.e., interferences among pixels are imbalanced. In addition to preventing vulnerable
patterns suffering from possible interferences, some modulation codes also focus on decoder
complexity, and yet others focus on achieving high code rate. It is desirable to consider all
aspects but trade-off is matter-of-course. Different priorities in design consideration result in
various modulation codes.
In this chapter, we will first introduce several modulation code constraints. Next, one-
dimensional modulation codes adopted in prevalent optical data storage systems are
discussed. Then we turn to the modulation codes designed for holographic data storage.
These modulation codes are classified according to the coding methods, i.e., block codes vs.
strip codes. For block codes, code blocks are independently produced and then tiled to form a
2 Data Storage

whole page. This guarantees a one-to-one relationship between the information bits and the
associated code blocks. On the contrary, strip codes produce code blocks by considering the
current group of information bits as well as other code blocks. This type of coding
complicates the encoding procedure but can ensure that the constraints be satisfied across
block boundary. We will further discuss variable-length modulation codes, which is a
contrast to fixed-length modulation codes. Variable-length modulation codes have more
freedom in the design of code blocks. With a given code rate, variable-length modulation
codes can provide better modulated pages when compared to fixed-length modulation
codes. However, variable-length modulation codes can suffer from the error propagation
problem where a decoding error of one code block can lead to several ensuing decoding
errors.

2. Constraints
Generally speaking, constraints of modulation codes are designed according to the channel
characteristics of the storage system. There are also other considerations such as decoder
complexity and code rate. In conventional optical data storage systems, information carried
in a binary data stream is recorded by creating marks on the disk with variable lengths and
spaces between them. On the other hand, information is stored in 2-D data pages consisting
of ON pixels and OFF pixels in holographic data storage system. The holographic
modulation codes encode one-dimensional information streams into 2-D code blocks. The
modulated pages created by tiling code blocks comply with certain constraints, aiming at
reducing the risk of corrupting signal fidelity during writing and retrieving processes. Due
to the additional dimension, other considerations are required when designing constraints
for holographic modulation codes. In the following some commonly adopted constraints are
introduced.

2.1 Run-Length Limited Constraint

The run-length limited constraint is widely adopted in optical storage systems. Examples
are the eight-to-fourteen modulation (EFM) in CD, EFMPlus in DVD, and the 17 parity
preserve-prohibit repeated minimum run-length transition (17PP) in Blu-ray disc. Due to the
different reflectivity states, peak detection is the most common receiver scheme. To reliably
detect peaks, separation on the order of 1.5-2 mark diameters is required between marks
(McLaughlin, 1998). The run-length limited constraint thus sets limits to the frequency of
ones in the data stream to avoid the case where large run-length of zeros causes difficulty of
timing recovery and streams with small run-length of zeros have significant high-frequency
components that can be severely attenuated during readout.
From one-dimensional track-oriented to 2-D page-based technology, the run-length between
“1”s (or ON pixels) has to be extended to 2-D. Two run-length limited constraints for 2-D
patterns have been proposed. One constraint sets upper and lower bounds to run-length of
zeros in both horizontal and vertical directions (Kamabe, 2007). The other one sets upper
and lower bounds to 2-D spatial distance between any two ON pixels (Malki et al., 2008;
Roth et al., 2001). See Fig. 1 for illustration of this 2-D run-length limited constraint.
Modulation Codes for Optical Data Storage 3

2.2 Conservative Constraint

The conservative constraint (Vardy et al., 1996) requires at least a prescribed number of
transitions, i.e., 10 or 01, in each row and column in order to avoid long periodic
stretches of contiguous light or dark pixels. This is because a large area of ON pixels results
in a situation similar to over-exposure in photography. The diffracted light will illuminate
the dark region and lead to false detection. An example of such detrimental pattern is
shown in Fig. 2.

Fig. 1. Illustration of the 2-D run-length limited constraint based on 2-D spatial distance with
lower and upper bounds of two and four, respectively.

Fig. 2. A pattern forbidden by the conservative constraint (Blaum et al., 1996).

2.3 Low-Pass Constraint

The low-pass constraint (Ashley & Marcus, 1998; Vadde & Vijaya Kumar, 2000) excludes
code blocks with high spatial frequency components, which are sensitive to inter-pixel
interference induced by holographic data storage channels. For example, an ON pixel tends
to be incorrectly detected as “0” when it is surrounded by OFF pixels and similarly an OFF
pixel tends to be incorrectly detected as “1” when surrounded by ON pixels. Therefore, such
patterns are forbidden under the low-pass constraint.
In fact, the low-pass constraints can be quite strict so that the legal code patterns have good
protection against high-frequency cutoff. This is, however, achieved at the cost of lower
code rate because few code blocks satisfy this constraint. Table 1 lists five low-pass
constraints and examples that violate these constraints.
4 Data Storage

Constraints Examples of forbidden patterns

At least one ON/OFF pixel exists among

eight nearest neighboring pixels of an
ON/OFF pixel.
,

At least one ON/OFF pixel exists among

four nearest neighboring pixels of an
ON/OFF pixel.
,
At least two ON/OFF pixels exist among
four nearest neighboring pixels of an
ON/OFF pixel. , , etc.

No ON/OFF pixels are isolated either

horizontally or vertically.
, , ,

No ON/OFF pixels are isolated either , , , ,

horizontally, vertically or diagonally.

, , etc.
Table 1. Low-pass constraints and examples of forbidden patterns.

2.4 Constant-Weight Constraint

The weight of a page is defined as the ratio of the number of ON pixels over the number of
all pixels. With a constant-weight page, low-complexity correlation detection can be
efficiently implemented (Coufal et al., 2000; Burr & Marcus, 1999; Burr et al., 1997). Two
kinds of weight distribution garner interests of researchers. One is balanced weight, which
makes the numbers of ON pixels and OFF pixels the same. The other is sparse weight,
which makes the number of ON pixels less than that of OFF pixels. Balanced-weight
modulation codes have higher code rates than sparse-weight modulation codes with the
same code block size. However, due to imbalanced interference between ON pixels and OFF
pixels, OFF pixels are favored as they cause lower level of interference. Therefore, sparse
codes enable more data pages that can be superimposed at the same location on the
recording medium by reducing optical exposure and increasing diffraction efficiency (the
ratio of the power of the diffracted beam to the incident power of that beam) per pixel
(Daiber et al., 2003). In addition, the probability of undesirable patterns can be reduced by
decreasing ON pixel density. To be more exact, up to 15% improvement of total memory
capacity can be achieved by sparse codes with the page weight decreased to 25% (King &
Neifeld, 2000).
Modulation Codes for Optical Data Storage 5

2.5 Summary
We have introduced four types of constraints for modulation codes commonly adopted in
optical data storage: run-length limited, conservative, low-pass and constant-weight
constraints. The run-length limited constraint focuses on the density of ON pixels. The
conservative constraint considers the frequency of transitions between ON and OFF pixels.
The low-pass constraint helps avoid those patterns vulnerable to inter-pixel interference
effects in the holographic data storage systems. As for the constant-weight constraint, it
enables a simple decoding scheme by sorting pixel intensities. Sparse modulation codes
further decrease the probability of vulnerable patterns and increase the number of
recordable pages.

3. One-Dimensional Modulation Codes

The modulation codes adopted by current optical data storage systems, i.e., CD, DVD, and
Blu-ray disc, are developed according to the run-length limited constraint. As we mentioned
previously, short runs result in small read-out signal power which tends to cause errors
while long runs cause failure in time recovery. Therefore, a run-length limited code in the
non-return-to-zero (NRZ) notation requires the number of “0”s between two “1”s to be at
least d and at most k. The real bits recorded on the disc are then transformed from NRZ to
non-return-to-zero inverted (NRZI) format consisting of sequence of runs which changes
polarity when a “1” appears in the NRZ bit stream, as shown in Fig. 3.
Under the EFM rule adopted by CD, 8-bit information sequences are transformed into 14-bit
codewords using a table. The 14-bit codewords satisfy the (2, 10)-run-length limited
constraint in the NRZ notation, that is, two “1”s are always separated by at least two “0”s
and at most ten “0”s. To make sure bits across two adjacent codewords also comply with the
run-length limited constraint, three emerging bits precede every 14-bit codeword. Although
two merging bits are sufficient to meet this requirement, one additional bit is used in order
to do DC-control, that is, make the running digital sum, which is the sum of bit value (±1)
from the start of the disc up to the specified position, close to zero as much as possible.
Therefore, total 17 bits are needed to store eight information bit in the EFM code.
The EFMPlus code used in DVD translates sequences of eight information bit into 16-bit
codewords satisfying the (2, 10)-run-length limited constraint. Unlike EFM, which is simply
realized by a look-up table, EFMPlus code exploits a finite state machine with four states.
No merging bits are required to promise no violation across codewords. In addition, to
circumvent error propagation, the encoder is further designed to avoid state-dependent
decoding.
Blu-ray disc boosts data capacity by further shrinking physical bit size recorded on the disc,
leading to severer interference than CD and DVD. In Blu-ray disc, the 17PP modulation code
follows the (1, 7)-run-length limited constraint and exploits a new DC-control mechanism.
The DC-control is realized by inserting DC-bits at certain DC-control points in the
information stream. The polarity of the corresponding NRZI bit stream is flipped if a DC-bit
is set as “1” while the polarity remains if a DC-bit is set as “0”. Therefore, the best DC-free
bit stream that makes the running digital sum closest to zero can be obtained by properly
setting the DC-bits. The DC-control mechanism is illustrated in Fig. 4.
6 Data Storage

Fig. 3. Actual patterns recorded on the disc through a change from the NRZ format to the
NRZI format.

Fig. 4. Illustration of DC control in the 17PP modulation code (Blu-ray Disc Association,
2006).

4. Block Codes for Holographic Data Storage

A block code used in holographic data storage maps a sequence of information bits into a 2-
D code block in a one-to-one manner. The encoded blocks are tiled to create a whole page.
The ratio of the number of information bits to the number of block pixels is called code rate.
For example, a simple block code, named differential code, uses two pixels to represent one bit
and achieves a code rate of 1/2. Its code blocks are illustrated in Fig. 5. Basically, higher
code rate is preferred since less redundant pixels are included and more information is
carried in the final modulated page. For this purpose, enlarging block size or changing the
average intensity is applied (Kume et al., 2001). Fig. 6 illustrates a 5:9 block code with a code
rate of 5/9 and a 6:9 block code with a code rate of 2/3. With two of nine pixels in a code
block turned ON, there exist C29  36 possible code blocks. Therefore, these coded blocks are
sufficient to represent five information bits, giving a 5:9 block code. Similarly, with three of
nine pixels in a code block turned ON, there are C39  84 possible coded blocks, thus
achieving a 6:9 block code.
However, a larger block size may deteriorate decoding performance due to biased intensity
throughout the block. Higher intensity also degrades the quality of reconstructed images
according to the sparse code principle. On the other hand, increasing the number of possible
code blocks (and thus code rate) often leads to more complex decoding. Hence, modulation
Modulation Codes for Optical Data Storage 7

code design is a trade-off among higher code rate, simpler decoder and satisfactory BER
performance.
Block codes indeed provide a simple mapping in encoding and decoding. However, the risk
of violating constraints becomes significant when tiling multiple blocks together since illegal
patterns may occur across the boundary of neighboring blocks. To circumvent this problem,
additional constraints may be required. Unfortunately, more constraints can eliminate some
patterns that would have been legitimate. To maintain the code rate, a larger block size is
called for.

Fig. 5. Differential code.

Fig. 6. (a) A code block in the 5:9 block code and (b) a code block in the 6:9 block code.

4.1 6:8 Balanced Block Code

The 6:8 balanced block code (Burr et al., 1997) is one of the most common block modulation
codes used in the holographic data storage systems. A group of six information bits is
encoded into a 2×4 block with exactly four ON pixels and four OFF pixels, satisfying the
8
balanced constraint. Since C4  70 is larger than 26=64, we choose 64 patterns from 70 code
blocks and assign Gray code to these blocks . Correlation detection (Burr et al., 1997) is also
proposed to decode such block code. In this method the correlations between the retrieved
block and all code blocks are computed and the code block with the maximum correlation to
the received block is declared the stored pattern.

4.2 6:8 Variable-Weight Block Code

The constant-weight constraint can have negative impacts on the achievable code rate of a
modulation code. For instance, if we apply a sparse constant-weight constraint to a 2×4
block code and require two ON pixels and six OFF pixels in every code block, then there are
only 28 legal code blocks, giving four information bits and a low code rate of 1/2.
The 6:8 variable-weight modulation code (Chen & Chiueh, 2007) is a sparse block
modulation code with variable weight and relatively high code rate. Similar to the 6:8
balanced block code, the variable-weight code encodes a group of six data bits into a 2×4
block. There are C18  8 codewords with one ON pixel and C38  56 codewords with three ON
pixels, enough to represent 6-bit information. This design decreases the weight of a
modulated page from 50% to 34.38%. Therefore, the variable-weight code provides better
BER performance with the same coding scheme as the 6:8 balanced code performance due to
lower interference level from fewer ON pixels.
8 Data Storage

4.3 Codeword Complementing Block Modulation Code

The codeword complementing block modulation code (Hwang et al., 2003) adopts an
indirect encoding scheme so as to satisfy the pseudo-balanced constraint and thus achieve
uniform spectrum of the modulated pages. The coding procedure consists of two steps:
constant weight encoding and selective complementary encoding. The constant weight encoder
maps input information bits into one-dimensional constant-weight codewords. Each one-
dimensional N-bit codeword forms a row in an (N+1)×N code matrix. Then, some rows are
flipped according to the elements in the last row of the current code matrix. If a pixel in the
last row is “1”, the selective complementary encoder flips the corresponding row. In this
way, the final N×N modulated block has constant weight and relatively uniform spectrum
in the Fourier plane. An example of the codeword complementing block modulation code
encoding process with N=11 is illustrated in Fig. 7.
The decoding procedure is simple. By calculating the average intensity of each row of a
retrieved block, rows that have been complemented by the selective complementary encoder
are easily identified and the polarities of the pixels are reversed. Moreover, the last row in
the code matrix can be decided. Decoding the constant weight encoded blocks is achieved
by first sorting the intensity of the received pixels in each codeword and marking proper
number of pixels with higher intensity as “ON” and the other pixels of the codeword as
“OFF.” Then information bits are obtained by inverse mapping of constant weight encoding.

Fig. 7. Illustration of codeword complementing block modulation code encoding process

(N=11). (a) Input information block, (b) constant weight encoded block and (c) selective
complementary encoded block (Hwang et al., 2003).

4.4 Conservative Codes

Conservative codes are designed to satisfy the conservative constraint discussed previously.
The coding scheme consists of a set of modulation blocks which is used as masks for the
pixel-wise exclusive-OR operation with input information blocks. The encoded output
blocks will be conservative of strength t, meaning that there exist no less than t transitions in
each row and column. A method based on an error correcting code having minimum
Hamming distance d and d≥2t-3 is proposed in (Vardy et al., 1996).
First define a mapping φ: FnEn as follows:
 x   x   x  (1)
Modulation Codes for Optical Data Storage 9

where Fn is the set of all binary column vectors and En is the subset of Fn consisting of all
even-weight vectors; σ(x) is the 1-bit cyclic downward shift of a column vector x. Given two
sets of codewords, C1 and C2, produced by two (t-2)-error correcting codes of length m and
n, respectively, two sets of column vectors {a1, a2,…, an} and {b1, b2, …, bm} are obtained by
inverse-mapping of φ-1(C1) and φ-1(C2). A theorem in (Vardy et al., 1996) states that for any
x  ai
vector x in Fn, if has less than t transitions, x  a j will have has at least t transitions
for all j≠i.
Next one constructs matrices A1,A2,An and B1,B2,Bm all with the size of m×n, from the
aforementioned sets of column vectors, {a0, a1,…, an} and {b0, b1, …, bm}. Ai is formed by
tiling ai n times horizontally while Bj is formed by tiling transposed bj m times vertically.
A i  a i a i  a i , i  1, ..., n (2)
b Tj 
 T
b
B j   j , j  1, ..., m (3)
  
 T
b j 
Finally, a modulation block is obtained by component-wise exclusive-OR of an Ai and a Bj.
On the total, there are (m+1)×(n+1) such modulation blocks. With such modulation block
construction, it can be shown that pixel-wise exclusive-OR operation of an input
information block with a modulation block yields an modulated block that satisfies the
conservative constraint with strength t. To indicate which modulation block is applied,
several bits are inserted as an extra row or column to each recorded page.

4.5 Block Code with 2-D Run-Length-Limited Constraint

In 2-D run-length limited block codes, only the minimum distance (dmin) between ON pixels
is enforced. Codebook design involves an exhaustive search of all possible binary blocks and
prunes any blocks containing two ON pixels with a square distance smaller than dmin. A
mapping between segments of information bits and these code blocks is then developed.

Fig. 8. An example of 8×8 block with three all-ON sub-blocks (Malki et al., 2008).

In (Malki et al., 2008), a fixed number of all-ON sub-blocks are present in each code block.
This scheme lets the proposed code comply with the constant-weight constraint
automatically. An example of three 2×2 all-ON sub-blocks is shown in Fig. 8. For an 8×8
block without any constraints, there are (8-1)×(8-1)=49 possible positions for each sub-block,
10 Data Storage

denoted by filled circles. The value of dmin determines the available number of legal code
blocks.

5. Strip Codes for Holographic Data Storage

In contrast to block codes, which create modulated pages by tiling code blocks
independently, strip codes exploit a finite state machine to encode information streams
taking adjacent blocks into consideration. The name of “strip” comes from dividing a whole
page into several non-overlapping strips and encoding/decoding within each strip
independently. To decode such modulation codes, a sliding block decoder (Coufal et al., 2000)
is required. Such a decoder has a decoding window containing preceding (memory), current
and subsequent (anticipation) blocks.
Strip codes provide an advantage of ensuring that the modulated patterns across borders of
blocks can still comply with the prescribed constraints if the coding scheme is well designed.
However, additional constraints for patterns across borders of strips are still required. These
additional constraints are called strip constraints (Ashley & Marcus, 1998; Coufal et al., 2000).
Another advantage of strip codes is their better error-rate performance due to the fact that
adjacent modulated blocks are used to help decode the current block.
In terms of the decoding procedure, strip codes need decoders with higher complexity when
compared with block codes. Another concern for strip codes is that using a finite state
machine as the encoder may fail to achieve the 2-D capacity C (Ashley & Marcus, 1998)
defined as
log 2 N h, w
C  lim (4)
h , w hw ,
where N(h,w) is the number of legal code blocks on an h×w block.
Strip codes solve problems of violating constraints across block borders by considering
adjacent modulated blocks during encoding/decoding operation of the current block. As
such, better conformance to the constraints and BER performance are attained at the cost of
higher encoding/decoding complexity.

5.1 8:12 Balanced Strip Code

The 8:12 strip modulation code proposed in (Burr et al.,1997) is a typical balanced strip code.
A page is first divided into non-overlapping strips with height of two pixels. Balanced code
blocks with size of 2×6 are prepared in advance with minimum Hamming distance of four.
During encoding, a finite state machine with four states is used to map 8-bit information
into a 2×6 block based on the current state and input information bits, achieving a code rate
of 2/3. The process goes on until a complete modulated page is obtained. As for decoding,
the Viterbi decoder can be adopted for better performance. The 8:12 modulation code is
block-decodable and can be decoded using the correlation detection. However, the error rate
performance will not be as good as those using Viterbi decoders.

5.2 9:12 Pseudo-Balanced Strip Code

The 9:12 pseudo-balanced code (Hwang et al., 2002) is a strip code which maps 9-bit
information into a 12-bit codeword with a code rate of ¾. It is similar to the 8:12 balanced
Modulation Codes for Optical Data Storage 11

modulation code, but with higher code rate by relaxing the balanced constraint to allow
certain degree of codeword imbalance. The 9:12 pseudo-balanced code maps information
bits to codewords according to a trellis with eight diverging and eight merging states. Each
branch represents a set of 64 pseudo-balanced 12-bit codewords with minimum Hamming
distance of four. These codewords are divided into 16 groups. The most-significant three
information bits of the current data word and the most-significant three bits of the previous
data word indicate the states in the previous stage and the current stage, respectively. The
codeword set is determined uniquely by these two states. The least-significant six bits of the
current information word are used to select a codeword within the set.
The decoding is similar to Viterbi decoder for trellis codes (Proakis, 2001). First, the best
codeword with respect to the acquired codeword for each codeword set is found. The index
of that best codeword and the associate distance are recorded. Then the Viterbi algorithm
finds the best state sequence using those aforementioned distances as branch metrics. In
(Hwang et al., 2002), it is shown that the 9:12 pseudo-balanced code, though with a higher
code rate, provides similar performance as the 8:12 balanced strip code.

5.3 2-D Low-Pass Codes

Previously, several low-pass constraints which prohibit some particular patterns with high
spatial frequency components are introduced. Strip codes are better suited for compliance
with these constraints since they can control patterns across block borders more easily.
Current Input information bits
state 000 001 010 011 100 101 110 111
00 10 01 11 11 01 00 11
1 1 1 2 2 2 2 3 4
00 10 01 11 01 11 11 00
00 10 01 11 00 10 00 11
2 1 1 2 2 1 1 3 4
00 10 01 11 10 00 11 00
00 10 01 11 00 01 00 11
3 1 1 2 2 1 2 3 4
00 10 01 11 10 11 11 00
00 10 01 11 11 10 00 11
4 1 1 2 2 2 1 3 4
00 10 01 11 01 00 11 00
Table 2. Encoder for the third constraint in Table 1 (Ashley & Marcus, 1998).

00 10 01 11 00 11 10 01 00 11
Code block
00 10 01 11 10 01 00 11 11 00
Information
000 001 010 011 100 100 101 101 110 111
bits
Table 3. Decoder for the third constraint in Table 1 (Ashley & Marcus, 1998).
Besides, the high-frequency patterns may not only appear within strips but also across
strips. Strip constraint is therefore very important when we require a low-pass property over
the whole modulated page. Below is an example of the strip constraint applying to a strip
code that satisfies the low-pass constraint. Given a forbidden pattern in Table 1 with height
L, a strip constraint banning the top/bottom m top/bottom rows of this forbidden pattern to
appear in the top/bottom of the current strip, where m is between L/2 and L. With this extra
strip constraint, it is guaranteed that the low-pass constraint will be satisfied across strip
boundary. Table 2 and Table 3 are an example of encoder/decoder designed for the third
12 Data Storage

constraint in Table 1. Using a finite state machine with four states, the encoder transforms
three information bits into a 2×2 code block based on the current state. The code block and
the next state are designated in the Table 2. The three information bits are easily decoded
from the single retrieved block. The mapping manner is shown in Table 3.

6. Variable-Length Modulation Codes for Holographic Data Storage

In Sections 4 and 5, we have introduced block codes and strip codes for holographic data
storage, respectively. The code rates of the modulation codes we discussed are generally
fixed. This kind of modulation codes is called fixed-length since they map a fixed number of
information bits into a fixed-size code block. To allow for more flexibility in modulation
codes, variable-length modulation codes were proposed. With such modulation codes, a
variable number of information bits are mapped into a variable-size code block. Despite its
advantage of more freedom in choosing legal code blocks that comply with certain
constraints, variable-length modulation codes need a more elaborate decoding scheme to
correctly extract code blocks without causing error propagation.
A variable-length modulation code designed to obey the low-pass constraint is proposed in
(Pansatiankul & Sawchuk, 2003). It eliminates patterns which contain high spatial frequency
not only within blocks but also across block borders. With this strict constraint and a given
code rate, variable-length modulation codes produce better modulated pages than fixed-
length modulation codes. Instead of preventing forbidden patterns such as those listed in
Table 1, (Pansatiankul & Sawchuk, 2003) sets maximum numbers of ON pixels surrounding
an OFF pixel. The notation of this variable-length modulation code is ([m1, m2, …], [n1, n2,
…]; [k1, k2, …]; α, β), where mi and ni are the size of 2-D blocks, ki is the length of one-
dimensional input information bits, α is a maximum number of ON pixels in the four
connected neighbor positions surrounding an OFF pixel and β is another maximum number
of ON pixels in the rest four of eight nearest neighbor positions surrounding an OFF pixel.
Parts of the modulated page produced by (4,[1,2,3];[2,4,6];2,3) and (1,[4,6,8]; [2,3,4];3,0)
variable-length modulation codes are given in Fig. 9. Rectangles indicate 3×3 blocks with
highest allowable number of ON-pixels around an OFF-pixel according to the respective (α,
β) constraints.
The encoding process is the same as those in conventional block codes, giving a one-to-one
manner between information bits and code blocks. In contrast to the straightforward
encoding scheme, the decoding has the challenge of correctly determining the size of the
current code block to do inverse-mapping. A decoding scheme for the (4,[1,2,3];[2,4,6];2,3)
variable-length modulation code is provided in (Pansatiankul & Sawchuk, 2003). We first
grab a 4×1 retrieved block and compare it to all 4×1 code blocks corresponding to 2-bit
information. Note that in (4,[1,2,3];[2,4,6];2,3) variable-length modulation code, all 4×2 and
4×3 coded patterns are designed to have an all-OFF last column. If we cannot find a 4×1
code block that is close to the retrieved block and the next column contains only OFF pixels,
we enlarge the retrieved block size from 4×1 to 4×2 and compare it to all 4×2 code blocks
corresponding to 4-bit information. If still we cannot find any good 4×2 code block, then an
error is declared. Similarly, we can check the second column to the right of the current 4×2
retrieved block for all OFF pixels and enlarge the retrieved block to 4×3 if the check turns
out positive. Then compare the extended retrieved block to all 4×3 code blocks
Modulation Codes for Optical Data Storage 13

corresponding to 6-bit information. Similarly, an error is declared when no code block

matches the retrieved block.

Fig. 9. Sample modulated page produced by two variable-length modulation codes

(Pansatiankul & Sawchuk, 2003).

7. Conclusion
In this chapter, modulation codes for optical data storage have been discussed. At first, four
types of constraints are introduced, including run-length limited, conservative, low-pass
and constant-weight constraints. Since high spatial frequency components tend to be
attenuated during recording/reading procedures and long runs of OFF pixels increase
difficulty in tracking, the former three types of constraints are proposed to avoid these
adverse situations as much as possible. On the other hand, the constant-weight constraint
gives modulated pages that are easier to decode. In addition, experiments indicate that
better performance can be obtained for modulation codes that have sparse weight.
Based on the constraints, several modulation codes are discussed. The one-dimensional
modulation codes adopted in current optical storage systems, i.e., EFM for CD, EFMPlus for
DVD and 17PP for Blu-ray disc, are first introduced. All of these modulation codes are
developed for the run-length limited constraint.
Next, we focus on 2-D modulation codes for holographic data storage systems. They are
classified into block codes and strip codes. Information bits and code blocks have a one-to-
one relationship in block codes whose encoder/decoder can be simply realized by look-up
tables. However, block codes cannot guarantee that patterns across block borders comply
with the required constraints. This shortcoming can be circumvented by strip codes, which
produce code blocks based on not only the input information bits but also neighboring
modulated blocks. A finite state machine and a Viterbi decoder are typical schemes for the
encoding and decoding of the strip codes, respectively.
Variable-length modulation codes, in contrast to fixed-length modulation codes, do not fix
the number of input information bits or the code block size. The relaxed design increases the
number of legal patterns and provides better performance than the fixed-length modulation
codes with the same code rate. However, error propagation problems necessitate a more
elaborated decoder scheme.
Finally, comparisons among different types of modulation codes introduced in this chapter
are listed in Table 4 and Table 5.
14 Data Storage

Block code Strip codes

One-to-one manner between information
Yes No
bits and code blocks
Constraint satisfaction across block borders Difficult Simple
Encoding/decoding complexity Low High
Table 4. Comparison between block codes and strip codes.

Fixed-length Variable-length
Freedom of choosing legal pattern Low High
Error propagation problem during decoding No Yes
BER performance with the same code rate Poor Good
Code rate with the same BER performance Low High
Table 5. Comparison between fixed-length and variable-length modulation codes

8. References
Ashley, J. J. & Marcus, B. H. (1998). Two-dimensional low-pass filtering codes. IEEE Trans.
on Comm., Vol. 46, No. 6, pp. 724-727, ISSN 0090-6778
Blaum, M.; Siegel, P. H., Sincerbox, G. T., & Vardy, A. (1996). Method and apparatus for
modulation of multi-dimensional data in holographic storage. US patent (Apr. 1996)
5,510,912
Blu-ray Disc Association (2006). White paper Blu-ray disc format: 1. A physical format
specifications for BD-RE, 2nd edition
Burr, G. W.; Ashley, J., Coufal, H., Grygier, R. K., Hoffnagle, J. A., Jefferson, C. M., &
Marcus,B. (1997). Modulation coding for pixel-matched holographic data storage.
Optics Letters, Vol. 22, No. 9, pp. 639-641, ISSN 0146-9592
Burr, G. W. & Marcus, B. (1999). Coding tradeoffs for high density holographic data storage.
Proceedings of the SPIE, Vol. 3802, pp. 18–29, ISSN 0277-786X
Coufal, H. J.; Psaltis, D. & Sincerbox, G. T. (Eds.). (2000). Holographic data storage, Springer-
Verlag, ISBN 3-540-66691-5, New York
Chen, C.-Y. & Chiueh, T.-D. A low-complexity high-performance modulation code for
holographic data storage, Proceedings of IEEE International Conference on Electronics,
Circuits and Systems, pp. 788-791, ISBN 978-1-4244-1377-5, Dec. 2007, Morocco
Daiber, A. J.; McLeod, R. R. & Snyder, R. (2003). Sparse modulation codes for holographic
data storage. US patent (Apr. 2003) 6,549,664 B1
Hwang, E.; Kim, K., Kim, J., Park, J. & Jung, H. (2002). A new efficient error correctible
modulation code for holographic data storage. Jpn. J. Appl. Phys., Vol. 41, No. 3B,
pp. 1763-1766, ISSN 0021-4922
Hwang, E.; Roh, J., Kim, J., Cho, J., Park, J. & Jung, H. (2003).A new two-dimensional
pseudo-random modulation code for holographic data storage. Jpn. J. Appl. Phys.,
Vol. 42, No. 2B, pp. 1010-1013, ISSN 0021-4922
Immink, K. A.; Nijboer, J. G., Ogawa, H. & Odaka, K. (1985). Method of coding binary data.
U.S. Patent (Feb. 1985) 4,501,000
Immink, K. A. (1997). Method of converting a series of m-bit information words to a
modulated signal, method of producing a record carrier, coding device, decoding
Modulation Codes for Optical Data Storage 15

device, recording device, reading device, signal, as well as record carrier. U.S.
Patent (Dec. 1997) 5,696,505
Kamabe, H. (2007). Representation of 2 dimensional RLL constraint, Proceedings of IEEE
International Symposium on Information Theory (ISIT), pp. 1171-1175, ISBN 978-1-
4244-1397-3, Jun. 2007, Nice
King, B. M. & Neifeld, M. A. (2000). Sparse modulation coding for increased capacity in
volume holographic storage. Applied Optics, Vol. 39, No. 35, pp. 6681-6688, ISSN
0003-6935
Kume, T.; Yagi, S., Imai, T. & Yamamoto, M. (2001). Digital holographic memory using two-
dimensional modulation code. Jpn. J. Appl. Phys., Vol. 40, No. 3B, pp. 1732-1736,
ISSN 0021-4922
Malki, O.; Knittel, J., Przygodda, F., Trautner, H. & Richter, H. (2008). Two-dimensional
modulation for holographic data storage systems. Jpn. J. Appl. Phys., Vol. 47, No. 7,
pp. 5993-5996, ISSN 0021-4922
McLaughlin, S. W. (1998). Shedding light on the future of SP for optical recording. IEEE
Signal Processing Magazine, Vol. 15, No. 4, pp. 83-94, ISSN 1053-5888
Pansatiankul, D. E. & Sawchuk, A. A. (2003). Variable-length two-dimensional modulation
coding for imaging page-oriented optical data storage systems. Applied Optics, Vol.
42, No. 26, pp. 5319-5333, ISSN 0003-6935
Proakis, J. G. (2001) Digital Communications, McGraw-Hill Science/Engineering/Math; 4th
edition, ISBN 0-07-232111-3, International
Roth, R. M.; Siegel, P. H. & Wolf, J. K. (2001). Efficient coding schemes for the hard-square
model. IEEE Trans. on Info. Theory, Vol. 47, No. 3, pp. 1166-1176, ISSN 0018-9448
Vadde, V. & Vijaya Kumar, B.V.K. (2000). Performance comparison of equalization and low-
pass coding for holographic storage, Conference Digest Optical Data Storage, pp. 113-
115, ISBN 0-7803-5950-X, May 2000
Vardy, A.; Blaum, M., Siegel, P. H., & Sincerbox, G. T. (1996). Conservative arrays:
multidimensional modulation codes for holographic recording. IEEE Trans. on Info.
Theory, Vol. 42, No. 1, pp. 227-230, ISSN 0018-9448
16 Data Storage
Signal Processing in Holographic Data Storage 17

X2

Signal Processing in Holographic Data Storage

Tzi-Dar Chiueh and Chi-Yun Chen
National Taiwan University
Taipei, Taiwan

1. Introduction
Holographic data storage (HDS) is regarded as a potential candidate for next-generation
optical data storage. It has features of extremely high capacity and ultra-fast data transfer
rate. Holographic data storage abandons the conventional method which records
information in one-dimensional bit streams along a spiral, but exploits two-dimensional (2-
D) data format instead. Page access provides holographic data storage with much higher
throughput by parallel processing on data streams. In addition, data are saved throughout
the volume of the storage medium by applying a specific physical principle and this leads
data capacity on the terabyte level.
Boosted data density, however, increases interferences between stored data pixels.
Moreover, the physical limits of mechanical/electrical/optical components also result in
misalignments in the retrieved images. Typical channel impairments in holographic data
storage systems include misalignment, inter-pixel interferences and noises, which will be
discussed in the following. One channel model that includes significant defects, such as
misalignment, crosstalk among pixels, finite pixel fill factors, limited contrast ratio and
noise, will also be introduced.
The overall signal processing for holographic data storage systems consists of three major
parts: modulation codes, misalignment compensation, equalization and detection. A block
diagram of such a system is illustrated in Fig. 1. Note that in addition to the three parts,
error-correcting codes (ECC) help to keep the error rate of the retrieved information under
an acceptable level. This topic is beyond the scope of the chapter and interested readers are
referred to textbooks on error-correcting codes.
To help maintain signal fidelity of data pixels, modulation codes are designed to comply
with some constraints. These constraints are designed based on the consideration of
avoiding vulnerable patterns, facilitation of timing recovery, and simple decoder
implementation. Modulation codes satisfying one or more constraints must also maintain a
high enough code rate by using as few redundant pixels as possible. The details are
discussed in the chapter on “Modulation Codes for Optical Data Storage.”
Even with modulation codes, several defects, such as inter-pixel interferences and
misalignments, can still cause degradation to the retrieved image and detection
performance. Misalignments severely distort the retrieved image and thus are better
handled before regular equalization and detection. Iterative cancellation by decision
18 Data Storage

feedback, oversampling with resampling, as well as interpolation with rate conversion are
possible solutions for compensating misalignment.
Equalization and detection are the signal processing operations for final data decision in the
holographic data storage reading procedure. Over the years, many approaches have been
proposed for this purpose. For one, linear minimum mean squared error (LMMSE)
equalization is a classical de-convolution method. Based on linear minimum mean squared
error (LMMSE) equalization, nonlinear minimum mean squared error equalization can
handle situations where there exists model mismatch. On the other hand, maximum
likelihood page detection method can attain the best error performance theoretically.
However, it suffers from huge computation complexity. Consequently, there have been
several detection algorithms which are modifications of the maximum likelihood page
detection method, such as parallel decision feedback equalization (PDFE) and two-
dimensional maximum a posteriori (2D-MAP) detection.

Fig. 1. Block diagram of signal processing in holographic data storage systems.

2. Recording and Retrieving Process

Holographic data storage systems store information carried by the interfered patterns
within the recording media. The interfered pattern is created by one object beam and one
reference beam. A spatial light modulator is adopted in holographic data storage systems to
produce object beams, which is similar to a three-dimensional (3-D) object from which light
is scattered and thus recorded in conventional holography. The spatial light modulator
displays information in a 2-D format and modulates the intensity, phase or both intensity
and phase of light beams. The 2-D data pages can consist of binary pixels (ON and OFF) or
gray-scale pixels (Burr et al., 1998; Das et al., 2009). Spatial light modulators can be
implemented by liquid crystal displays, digital micro-mirror devices, etc.
Signal Processing in Holographic Data Storage 19

Fig. 2. Holographic data storage system: (a) recording process and (b) retrieving process
(Chen et al., 2008).

Fig. 3. Illustration of recording and reading process (InPhase).

With the spatial light modulator well controlled by the computer, information-carrying
object beams can be created by passing a laser light through or being reflected by the spatial
light modulator. Next, the object beam is interfered with a reference beam, producing an
interferes pattern, namely, a hologram, which then leads to a chemical and/or physical
change in the storage medium. By altering one or more characteristics of the reference beam,
e.g., angle, wavelength, phase or more of them, multiple data pages can be superimposed at
the same location. The process is called multiplexing. There are many multiplexing methods;
interested readers can refer to (Coufal et al., 2000) for more details.
Reference beams of data pages at a certain location in the recording medium are like
addresses of items in a table. In other words, data pages can be distinguished by the
reference beam which interfered with the corresponding object beam. Therefore, by
illuminating the recording medium with the reference beam from a proper incident angle,
with a proper wavelength and phase, a certain object beam can be retrieved and then
captured by the detector array at the receiving end. The stored information is processed by
20 Data Storage

transforming captured optical signals to electrical signals. The detector array is usually a
charge coupled device (CCD) or a complementary metal oxide semiconductor (CMOS)
image sensor. Minimum crosstalk from other pages when retrieving individual pages is
attributed to a physical property called Bragg effect. Fig. 2 shows an example of the
holographic data storage systems and Fig. 3 illustrates more details in the recording and
reading processes.

3. Channel Defects
Ideally, a pixel of the spatial light modulator is directly imaged onto a detector pixel in a
pixel-matched holographic data storage system. Such perfect correspondence is practically
difficult to maintain due to non-ideal effects in holographic data storage channel that distort
light signal and deteriorate signal fidelity. These non-ideal effects are called channel defects.
Like most communication systems, optical storage systems have channel defects. Boosted
storage density makes holographic data storage signals especially sensitive to interferences,
noise or any tiny errors in storage media and optical/mechanical fabrication. From the
spatial light modulator to the detector, from optical to mechanical, quite a few factors have
significant influence on retrieved signals.
Laser light intensity determines the signal-to-noise ratio (SNR) but the intensity may not be
uniform across the whole page. Usually, corners of data pages are darkerdue to the
Gaussian wave front of the light source. This phenomenon and variations of light intensity
over a page both are categorized as non-uniformity effect of light intensity. Non-uniformity
of the retrieved signal level may cause burst errors at certain parts of data pages. Another
factor is the contrast ratio, which is defined as the intensity ratio between an ON pixel and
an OFF pixel. Ideally the contrast ratio is infinite. However, the OFF pixels actually have
non-zero intensity in the spatial light modulator plane. A finite contrast ratio between the
ON and OFF pixels makes the detection of pixel polarity more cumbersome. Another non-
ideal effect is related to fill factors of spatial light modulator pixels and detector pixels. A fill
factor is defined as the ratio of active area to the total pixel area, which is less than unity in
real implementation.
Crosstalk is actually the most commonly discussed channel defect. We discuss two kinds of
crosstalk here, including inter-pixel interference and inter-page interference. Inter-pixel
interference is the crosstalk among pixels on the same data page. It results from any or
combination of the following: band-limiting optical apertures, diffraction, defocus and other
optical aberrations. Inter-page interference is caused by energy leakage from other data
pages. This can be caused by inaccuracy of the reference beam when a certain data page is
being retrieved. As more pages are stored in the medium, inter-page interference becomes a
higher priority in holographic data storage retrieval. How good the interference can be
handled determines to a large extent the number of date pages that can be superimposed.
Mechanical inaccuracies bring about another type of channel defects. Errors in the optical
subsystem, mechanical vibration and media deformation can lead to severe misalignment,
namely magnification, translation and rotation, between the recorded images and the retrieved
images. Even without any inter-pixel interference or other optical distortions, misalignments
still destroy the retrieval results entirely, especially in pixel-matched systems. Fig. 4 shows
an example of misalignments (γx, γy, σx, σy, θ), where γx, γy are the magnification factors in
the X- and Y- directions, respectively; σx, σy are the translation in the range of ±0.5 pixel
Signal Processing in Holographic Data Storage 21

along the X- and Y- directions, respectively; and the rotation angle, θ, is positive in the
counter-clockwise direction.

Fig. 4. Misalignment effects: magnification, translation and rotation of a detector page with
respect to a spatial light modulator page.

Optical and electrical noises are inevitable in holographic data storage systems. They
include crosstalk noise, scattering noise, shot noise and thermal noise. Optical noise occurs
during the read-out process when the storage medium is illuminated by coherent reference
beams, resulting from optical scatter, laser speckle, etc. Electrical noise, such as shot noise
and thermal noise, occurs when optical signals are captured by detector arrays and
converted into electrical signals.

4. Channel Model
A holographic data storage channel model proposed in (Chen et al., 2008), shown in Fig. 5,
includes several key defects mentioned in Section 3. Multiple holographic data storage
channel impairments including misalignment, inter-pixel interference, fill factors of spatial
light modulator and CCD pixels, finite contrast ratio, oversampling ratio and noises are
modeled. The input binary data sequence, A(i, j), takes on values in the set {1, 1/ε}, where ε
is a finite value called the amplitude contrast ratio. The spatial light modulator has a pixel
shape function p(x, y) given by
 x   y ,
p( x, y )       (1)
 ff SLM    ff SLM  
where ffSLM represents the spatial light modulator’s linear fill factor, the symbol Δ represents
the pixel pitch and Π(．) is the unit rectangular function. Another factor that contributes to
inter-pixel interference is the point spread function, a low-pass spatial behavior with impulse
response hA(x, y) resulted from the limited aperture of the optics subsystem. This point
spread function is expressed as
    
hA x, y  hA x hA y , (2)
where
22 Data Storage

hA  x   D f L sinc  xD f L  . (3)
Note that D is the width of the square aperture, λ is the wavelength of the incident light, and
fL represents the focal length. The corresponding frequency-domain transfer function HA(fx,
fy) in this case is the ideal 2-D rectangular low-pass filter with a cut-off frequency equal to
D/2λfL.
A CCD/CMOS image sensor is inherently a square-law integration device that detects the
intensity of the incident light. The image sensor transforms the incoming signals from the
continuous spatial domain to the discrete spatial domain. Quantization in space causes
several errors due to offsets in sampling frequency, location, and orientation. Magnification,
translation and rotation are modeled as (γx, γy, σx, σy, θ), as explained previously. In addition,
the oversampling ratio M, the pixel ratio between spatial light modulator and CCD/CMOS
sensor, is another factor to be considered.
Taking all the aforementioned effects into account, we have the final image sensor output at
the (k, l)th pixel position given by
2
 
  A(i  a, j  b)h( x  a, y  b)  no ( x' , y ' ) dy'dx'  ne (k , l )
Cm (k , l )  (4)
 X Y a b 
where
 k  ff CCD 2   k  ff CCD 2  
 X     x ,    x  
 M  x   M x  
(5)
 l  ff 2   l  ff CCD 2  
 Y   CCD
  y  ,    y  
   M 
 M y   y  
and
x'  x cos   y sin 
. (6)
y '   x sin   y cos 
The subscript m in Cm(k, l) indicates misalignment; h(x, y) in Eq. (4) is also known as a pixel
spread function and h( x, y )  p ( x, y )  hA ( x, y ) ; ffCCD represents the CCD image sensor
linear fill factor; no(x‘, y‘) and ne(k, l) represent the optical noise and the electrical noise
associated with the (k, l)th pixel, respectively. Translation and magnification effects are
represented by varying the range of integration square as in Eq. (5), while the rotational
effect is represented by transformation from the x-y coordinates to the x’-y’ coordinates as
in Eq. (6).
The probability density function of optical noise no(x‘, y‘) can be described as a circular
Gaussian distribution and its intensity distribution has Rician statistics. On the other hand,
the electrical noise ne(k, l) is normally modeled as an additive white Gaussian noise with
zero mean (Gu et al., 1996).
Signal Processing in Holographic Data Storage 23

Fig. 5. Block diagram of a complete holographic data storage channel model.

5. Misalignment Compensation
Misalignments in retrieved images need be detected and compensated. To avoid
information loss in the case that the detector array receives only part of a data page, a larger
detector array or redundant (guard) pixels surrounding the information-carrying spatial
light modulator pages can be employed. Moreover, a misalignment estimation procedure
based on training pixels is needed before actual compensation. In fact, the estimation needs
to be performed locally due to non-uniformity of channel effects. Toward this end, a page is
divided into blocks of a proper size and training pixels are deployed in every block for local
misalignment estimation. One possible estimation method is correlation based. In this
method, the correlations of the received pixels and the known training pixels with different
values of magnification, translation and rotation are first computed. The parameter setting
with the maximum correlation is regarded as the estimation outcome. With the
misalignments estimated, the retrieved images need be compensated before the modulation
decoder decides the stored information. In the following several compensation schemes will
be introduced.

5.1 Decision Feedback

Two decision feedback algorithms capable of dealing with translation and rotational
misalignment compensation are proposed in (Menetrier & Burr, 2003) and (Srinivasa &
McLaughlin, 2005), respectively. The iterative detection method exploits decision feedback
to eliminate misalignment effects simultaneously. Hereafter, we explain the algorithm by
using a one-dimensional translation effect in the pixel-matched system with a simplified
channel model. A received pixel Cm(k, l) can be represented as a function of two adjacent
spatial light modulator pixels, A(k-1, l) and A(k, l), as shown in Fig. 6 and Eq. (7).
24 Data Storage

Fig. 6. Illustration of one-dimensional translation effect with parameter σ (Menetrier & Burr,
2003).

Cm k , l   
ffCCD 2

ff CCD 2

 ff CCD 2  ff CCD 2
 
Ak , l h  x   , y   Ak  1, l h  x    1, y  dxdy
2
(7)
Rewriting Eq. (7), we have
Cm k , l   Ak , l H 00    2 Ak , l  Ak  1, l H 01    Ak  1, l H11   . (8)
It is clear that if Cm(k, l) and A(k-1, l) are known, A(k, l) can be calculated according to Eq. (8).
The decision feedback detection scheme is based on this observation. If σ is positive in the
horizontal direction, then the scheme starts from the pixel at the top-left corner Cm(0, 0), A(0,
0) (the corner pixel) is first detected assuming that A(-1, 0) is zero. With A(0, 0) decided,
decision feedback detection moves to detect the next pixel A(1, 0) and repeats the same
process until all pixels are detected. When the translation is only in the horizontal
dimension, all rows can be processes simultaneously.
Extending the above case to 2-D, the retrieved pixel is a function of four spatial light
modulator pixels
Cm , s  As H ss  Ah H hh  Av H vv  Ad H dd
 2 As Ah H sh  2 As Av H sv  2 As Ad H sd , (9)
 2 Ah Av H hv  2 Ah Ad H hd  2 Av Ad H vd
where subscript s is for self, h for horizontal, v for vertical and d for diagonal. With the same
principle used in the one-dimensional decision feedback detection, one must detect three
pixels, horizontal, vertical and diagonal, before detecting the intended pixel. If both σx and σy
are positive, again we start from the top-left pixel calculating A(0, 0) assuming that pixel A(0,
-1), A(-1, -1), and A(-1, 0) are all zeros. The process is repeated row by row until all pixels are
detected.
A similar detection scheme for images with rotational misalignment is proposed in
(Srinivasa & McLaughlin, 2005). The process is somewhat more complicated because a
pixel’s relationship with associated SLM pixels depends on its location. For example, if the
detector array has rotational misalignment of an angle in the clockwise direction as shown
in Fig. 7, a pixel at the left top portion is a function of A(k, l), A(k, l-1), A(k+1, l-1) and A(k+1,
l), while pixel at the right-bottom corner is a function of A(k, l), A(k-1, l), A(k-1, l+1) and A(k,
l+1). Therefore, iterative decision feedback detection has to be performed differently
Signal Processing in Holographic Data Storage 25

depending on the location of the current pixel.

Fig. 7. Rotational misalignment entails different scan orders for the decision feedback
cancellation detection in different portions of a page (Srinivasa & McLaughlin, 2005).

5.2 Oversampling and Resampling

Oversampling is a good way to handle the magnification effect and avoid the situation that
a detector pixel is larger than a spatial light modulator pixel, leading to information loss.
Besides, even if there is no magnification effect, a sub-Nyquist oversampling is necessary
when the translation misalignment is around 0.5 pixel (Ayres et al., 2006b). With
oversampling, the misalignment compensation task is no more than resampling.
The resampling process is actually a combination of misalignment compensation and
equalization which will be discussed later. The resampling coefficients can be deduced by
using the minimum mean-square-error (MMSE) criterion. With different sets of
misalignment parameters, different sets of coefficients can be designed. For example, 20×20
sets of coefficients are required to handle 2-D translation misalignment with 5% precision.
The number of coefficients needed to handle all possible misalignment situations can lead to
a large memory size. Besides, oversampling uses more detector pixels to represent spatial
light modulator pixels, thus decreasing stored information and recording density.

5.3 Interpolation and Rate Conversion

To compensate misalignments and convert the retrieved pixel rate to that of the stored data
pixels, a 2-D filter capable of interpolation and rate conversion is proposed in (Chen et al.,
2008). Interpolation can tackle translation and rotational misalignments. Rate conversion
then converts the pixel rate to obtain pixel-matched pages free of misalignments.
Pixels in the captured images can be realigned by either 2-D interpolators or 2-D all-pass
fractional delay filters (Pharris, 2005). A simple example is the bilinear filter shown in Fig. 8.
26 Data Storage

Fig. 8. Realignment by using a bilinear interpolator with local fractional displacement μx and
μ y.

The reconstructed pixel corresponding to a spatial light modulator pixel is given by

Z m k , l   1   x   1   y  Cm ( k ' , l ' )   x  1   y  Cm ( k '1, l ' )
, (10)
 1   x    y  Cm ( k ' , l '1)   x   y  Cm ( k '1, l '1)
where Zm and Cm are the realigned and misaligned images, respectively.
In general, the 2-D interpolator can be formulated as
Z k , l  
m x  y 
f   p  f   q C ( k ' p, l ' q ) ,
 m (11)
( p ,q )S

where f(．) is the one-dimensional impulse response of the corresponding interpolator; S is

the input range of the 2-D interpolators; μx and μy indicate the local fractional displacement
from the nearest pixel (k’, l’) at Cm in horizontal and vertical directions, respectively. The
relationship between the local fractional displacement μx and μy and the corresponding pixel
position k’ and l’ on the misaligned image Cm can be expressed as
~
k  k cos   l sin    x   M   x ,
~ (12)
l  k sin   l cos    y  M   y
where M is the oversampling ratio and
~ ~

k  k , x  k  k
. (13)
~ ~

l  l ,  y  l  l
To further enhance the realigned image quality, higher-order filters can be adopted. In
(Chen et al., 2008), a 6×6-tap raised-cosine interpolator provides a better choice with
acceptable complexity and satisfactory performance than other filters.
The rate-conversion filter can properly restore oversampled images that have suffered
magnification effect to pixel-matched images for ensuing pixel detection. Without loss of
generality, the rate-conversion filter can be formulated as
Pmax Qmax
Z i, j      p  q Z k  p, l  q  ,
x y m (14)
p  Pmin q Qmin

where Zm indicates the realigned pixels and the weights νx(p) and νy(p) depend on the
magnification factors γx and γy as well as the oversampling ratio M.
Signal Processing in Holographic Data Storage 27

The realignment interpolator and the rate-conversion filter can be combined to reduce
complexity. First, both 2-D filters are separated into two respective one-dimensional
operations. Second, the realignment interpolators and the rate-conversion filters are
integrated to construct a misalignment-compensation block that consists of one-dimensional
compensation in the horizontal direction and one-dimensional compensation in the vertical
direction. With such rearrangement, 84% reduction in additions and 74% reduction in
multiplications are achieved (Chen et al., 2008).

6. Equalization and Detection

In general, equalization and detection is the final step of signal readout in holographic data
storage. At this stage, most of the distortion caused by misalignments should have been
removed. The remaining channel effects are mainly inter-pixel interference and noises.
Equalization tries to eliminate impairments on the page by making its pixel intensity
distribution as close to that of the spatial light modulator page as possible. A certain number
of known pixels are needed to train the equalizer coefficients. For instance, assume that
pixels on the distorted page are formulated as
Y = HX+N, (15)
where Y indicates the pixels on the distorted page; H is the pixel spread function of the
channel; X indicates the pixels on the SLM page and N indicates the noises. Equalization is
-1 -1
done by computing H Y, where H denotes the equalizer coefficients. There are several
well-known techniques for estimating these coefficients. However, as the holographic data
storage channel is nonlinear, the simplistic linear equalizer may have model mismatch,
making it ineffective.
Detection generates final decision for each pixel. The simplest detection is a slicer which
determines a binary pixel as “1” if the intensity is higher than a certain threshold and “0”
otherwise. The slicer must be preceded by equalization to achieve acceptable BER
performance. There exist other detection algorithms that operate without equalization.
However, they need additional channel state information for effective operation.

6.1 MMSE Equalization

LMMSE equalization is a popular equalization method in communication systems. In
holographic data storage systems, LMMSE equalization has to be extended to two-
dimensional (Singla & O’Sullivan, 2004; Chugg et al., 1999; Keskinoz & Vijaya Kumar, 1999;
Choi & Baek, 2003; Ayres et al., 2006a). Assume that the 2-D equalizer has (2K-1)×(2K-1) taps
and its output is given by
K K
Aˆ i, j     wm, n Z i  m, j  n  , (16)
mK nK
where w(m, n) is the (m, n)th tap of the equalizer, and Z(i, j) is the (i, j)th pixel of retrieved
data page. The LMMSE equalizer calculates w(m, n) that minimizes the average squared
error between equalized output and actual stored pixel, i.e.,

min  E  Aˆ i, j   Ai, j   ,

2
(17)
   
28 Data Storage

based on the MMSE criterion. According to the orthogonal principle (Haykin, 2002), we can
make use of auto-correlation of the received pixels and the cross-correlation between the
received and the desired pixels to solve for the optimal coefficients through the following
equations
K K
RAZ  p, q     wm, n R  p  m, q  n   w p, q   R  p, q  ,
ZZ ZZ (18)
mK nK
where  denotes convolution and
RAZ  p, q   E Ai, j Z i  p, j  q  .
(19)
RZZ  p, q   E Z i, j Z i  p, j  q 
(Keskinoz & Vijaya Kumar, 1999) provided a simple way to calculate equalizer coefficients
by applying Fourier transformation to Eq. (18). The equalizer coefficients can then be
obtained by
w p, q   IFFT FFT RAZ  p, q  FFT RZZ  p, q . (20)
Unfortunately, linear equalizers can suffer from model mismatch and render them
ineffective since the holographic data storage channel is inherently nonlinear. To this end,
nonlinear equalization was also proposed (Nabavi & Vijaya Kumar, 2006; He & Mathew,
2006).

6.2 Parallel Decision Feedback Equalization

Decision feedback equalization was proposed in (King & Neifeld, 1998) and (Keskinoz &
Vijaya Kumar, 2004). The algorithm iteratively detects pixels while considering the
interferences from neighboring pixels based on decisions that have been made in the
previous iterations. The process is executed on each pixel independently. The detection
results will converge after several iterations. Hereafter in this chapter, we called this method
parallel decision feedback equalization (PDFE).
PDFE is actually a simplified version of the optimal maximum likelihood page detection
algorithm which has a search space as large as an entire page and is thus impractical in
actual implementation. PDFE breaks the whole page into small blocks whose size equals the
extent of the channel pixel spread function and uses iterations to achieve near maximum
likelihood performance.
Signal Processing in Holographic Data Storage 29

Fig. 9. Illustration of PDFE.

Assume that we consider inter-pixel interference within a range of 3×3 pixels and there exist
no misalignment effects and the retrieved images have been made pixel-matched. Parallel
decision feedback equalization starts with computing hard decision for each pixel. Then two
hypotheses are tested to find the best decision for the current pixel. The process is shown in
Fig. 9. With eight surrounding pixels given by decisions from the previous iteration, the
central (current) pixel is decided as “1” or “0” according to
Aˆ i, j   1 if
2 2
Z (i, j )  H (1, nij )  Z (i, j )  H (0, nij ) ,
(21)
Aˆ i, j   0 if
2 2
Z (i, j )  H (1, nij )  Z (i, j )  H (0, nij )
where H(A(i, j), nij) is the inter-pixel-interference-inflicted channel output with A(i, j)=0 or 1
and a neighborhood pattern expressed as a binary vector, nij, consisting of the eight binary
pixels.
The performance of parallel decision feedback equalization depends on the correctness of
channel estimation and good initial condition. With inaccurate channel information or too
many errors in the initial condition, PDFE will have poor overall performance as initial
errors can propagate throughout the entire page.

6.3 2D-MAP
Two-dimensional maximum a posteriori (2D-MAP) detection, proposed in (Chen et al., 1998) as
the 2-D4 (Two-Dimensional Distributed Data Detection) algorithm, is actually the well-
known max-log-MAP algorithm. It is also a simplified sub-optimal maximum likelihood
page detection algorithm. Different from PDFE, the extrinsic information of each pixel is
now taken into consideration during searching for optimal decisions. Therefore, more than
two cases are tested in 2D-MAP. In this algorithm, a log-likelihood ratio (LLR) for each pixel
is used as extrinsic information and is maintained throughout the iterative process. An LLR
30 Data Storage

with higher absolute value indicates a greater probability of the pixel being “1” or “0.” As
the iteration goes on, the LLR value at each pixel will be re-calculated based on the
knowledge of the previous LLR values of its eight neighbors. In all likelihood, this process
makes each and every LLR move away from the origin and all pixel decisions become more
and more certain.
The procedure of the 2D-MAP detection comprises likelihood computation and update. In a
binary holographic data storage system, the likelihood computation formulas are given by
 1 2 
LL1U( k ) (i , j )  min  Z (i, j )  H (1, nij )  dij( k 1)  nij 
N ij  2 N 0 .
(22)
(k )  1 2 ( k 1) 
LL0U (i, j )  min  Z (i, j )  H (0, nij )  d ij  nij 
N ij  2 N 0 
Again H(A(i, j), nij) is the inter-pixel interference-inflicted channel output with A(i, j) = 0 or 1
and a neighborhood pattern expressed as a binary vector, nij, consisting of the eight binary
pixels; Nij is the set of all possible neighborhood patterns. In addition, d(k-1) is a vector
consisting of the corresponding LLR values of neighboring pixels in the (k-1)th iteration and
the symbol ‘ ’ represents inner product of two vectors. In the above, all misalignment effects
and oversampling have been properly handled and the only remaining channel effect is
inter-pixel interference and noises. LLRs are updated at the end of each iteration. To avoid
sudden changes in the LLR values, a forgetting factor β is applied and the updated LLR
takes the form of
 
L( k ) (i, j )  (1   ) L( k 1) (i, j )   LL1U( k ) (i, j )  LL0U( k ) (i, j ) , (23)
The value of β can affect the speed and accuracy of convergence. A larger β leads to faster
convergence but may lead to poor detection performance.
With the help of soft information, 2D-MAP indeed has better performance but with much
higher complexity than PDFE. In (Chen et al., 2008) several complexity reduction schemes,
including iteration, candidate, neighborhood and addition reduction, are proposed and thus
up to 95% complexity is saved without compromising the detection performance.

7. Conclusion
This chapter gives an overview to the processing of retrieved signals in holographic data
storage systems. The information to be stored is arranged in a 2-D format as binary or gray-
level pixels and recorded by interference patterns called hologram. The fact that multiple
holograms can be superimposed at the same location of the recording medium leads to
volume storage that provides very high storage capacity.
Two important channel defects, misalignments and inter-pixel interferences, are major
causes for degradation in detection performance and their model is formulated
mathematically. Several misalignments compensation algorithms are introduced. One
algorithm adopts decision feedback to handle misalignments and interference
simultaneously. Pixels are detected one by one after cancelling interference from
neighboring pixels. The scan orders should be carefully designed when misalignments may
involve pixels coming from other directions. Another algorithm makes use of oversampling,
and then resample at location. Another algorithm combines interpolation and rate
conversion to compensate various misalignments effects.
Signal Processing in Holographic Data Storage 31

Equalization and detection are crucial steps in restoring stored information from the
interference-inflicted signal. Albeit its popularity and low complexity, LMMSE equalization
algorithm suffers from the problem of model mismatch as the holographic data storage
channel is inherently nonlinear. In light of this fact, two nonlinear detection algorithms, both
simplified versions of the optimal maximum likelihood page detection, are introduced. They
achieve better performance than the LMMSE method at the cost of higher complexity.

8. References
Ayres, M.; Hoskins, A., & Curtis, K. (2006a). Processing data pixels in a holographic data
storage system. WIPO patent (Sep. 2006) WO/2006/093945 A2
Ayres, M.; Hoskins, A., & Curtis, K. (2006b). Image oversampling for page-oriented optical
data storage. Applied Optics, Vol. 45, No. 11, pp. 2459–2464, ISSN 0003-6935
Burr, G. W.; Coufal, H., Hoffnagle, J. A., Jefferson, C. M., & Neifeld, M. A. (1998). Gray-scale
data pages for digital holographic data storage. Optics Letters, Vol. 23, No. 15, pp.
1218–1220, ISSN 0146-9592
Chen, X.; Chugg, K. M. & Neifeld, M. A. (1998). Near-optimal parallel distributed data
detection for page-oriented optical memories. IEEE J. Sel. Top. Quantum Electron.,
Vol. 4, No. 5, pp. 866–879, ISSN 1077-260X
Chen, C.-Y.; Fu, C.-C. & Chiueh, T.-D. (2008). Low-complexity pixel detection for images
with misalignments and inter-pixel interference in holographic data storage.
Applied Optics, vol. 47, no. 36, pp. 6784–6795, ISSN 0003-6935
Choi, A.-S. & Baek, W.-S. (2003). Minimum mean-square error and blind equalization for
digital holographic data storage with intersymbol interference. Jpn. J. Appl. Phys.,
Vol. 42, No. 10, pp. 6424–6427, ISSN 0021-4922
Chugg, K. M.; Chen, X., Neifeld, M. A. (1999). Two-dimensional equalization in coherent
and incoherent page-oriented optical memory. J. Opt. Soc. Am. A, Vol. 16, No. 3, pp.
549–562, ISSN 1084-7529
Coufal, H. J.; Psaltis, D. & Sincerbox, G. T. (Eds.). (2000). Holographic data storage, Springer-
Verlag, ISBN 3-540-66691-5, New York
Das, B.; Joseph, J., Singh, K. (2009). Phase modulated gray-scale data pages for digital
holographic data storage. Optics Communications, Vol. 282, No. 11, pp. 2147–2154,
ISSN 0030-4018
Gu, C.; Dai, F., & Hong, J. (1996). Statistics of both optical and electrical noise in digital
volume holographic data storage. Electronic Letters, Vol. 32, No. 15, pp.1400–1402,
ISSN 0013-5194
Haykin, S. (2002). Adaptive filter theory, Prentice Hall, 4th edition, ISBN 0130901261
He, A. & Mathew, G. (2006). Nonlinear equalization for holographic data storage systems.
Applied Optics, Vol. 45, No. 12, pp. 2731–2741, ISSN 0003-6935
InPhase website: http://www.inphase-technologies.com
Keskinoz, M. & Vijaya Kumar, B. V. K. (1999). Application of linear minimum mean-
squared-error equalization for volume holographic data storage. Applied Optics,
Vol. 38, No. 20, pp. 4387–4393, ISSN 0003-6935
Keskinoz M. & Vijaya Kumar, B. V. K. (2004). Discrete magnitude-squared channel
modeling, equalization, and detection for volume holographic storage channels.
Applied Optics, Vol. 43, No. 6, pp. 1368–1378, ISSN 0003-6935
32 Data Storage

King, B. M. & Neifeld, M. A. (1998). Parallel detection algorithm for page-oriented optical
memories. Applied Optics, Vol. 37, No. 26, pp. 6275–6298, ISSN 0003-6935
Menetrier, L. & Burr, G. W. (2003). Density implications of shift compensation
postprocessing in holographic storage systems. Applied Optics, Vol. 42, No. 5, pp.
845–860, ISSN 0003-6935
Nabavi, S. & Vijaya Kumar, B. V. K. (2006). Application of linear and nonlinear equalization
methods for holographic data storage. Jpn. J. Appl. Phys., Vol. 45, No. 2B, pp. 1079–
1083, ISSN 0021-4922
Pharris, K. J. (2005). Methods and systems for holographic data recovery. U.S. Patent (Jan.
2005) 20050018263 A1
Singla, N. & O’Sullivan, J. A. (2004). Minimum mean squared error equalization using priors
for two-dimensional intersymbol interference. Proceedings of IEEE International
Symposium on Information Theory (ISIT), pp. 130, ISBN 0-7803-8280-3, Jun. 2004,
Chicago
Srinivasa, S. G. & McLaughlin, S. W. (2005). Signal recovery due to rotational pixel
misalignments. Proceedings of IEEE International Conference on Acoustics, Speech, and
Signal Processing (ICASSP), Vol. 4, iv/121- iv/124, ISBN 0-7803-8874-7, Mar. 2005,
Philadelphia
Optical data storage in photosensitive glasses and spin state transition compounds 33

X3

Optical data storage in photosensitive glasses

and spin state transition compounds
Matthieu Bellec, Lionel Canioni, Arnaud Royon, Bruno Bousquet,
Jérôme Degert and Eric Freysz
Centre de Physique Moléculaire Optique et Hertzienne (CPMOH), University of Bordeaux
France
Thierry Cardinal and Jean-François Létard
Institut de Chimie de la Matière Condensée de Bordeaux (ICMCB), University of Bordeaux
France

1. Introduction
Up to now, the common media for optical data storage are the discs (blue ray technology for
example). However, by definition, this technology is limited to two dimensions. The
necessity for increasing data storage capacity requires the use of three-dimensional (3D)
optically based systems. One of the methods for 3D optical data storage is based on volume
holography. The physical mechanism is photochromism, which is defined as a reversible
transformation of a single chemical species between two states that have different
absorption spectra and refractive indices. This allows for holographic multiplexing
recording and reading, such as wavelength (Rakuljic et al., 1992), angular (Mok, 1993), shift
(Psaltis et al., 1995) and phase encoding. Another promising 3D optical data storage system
is the bit-by-bit memory at the nanoscale (Li et al., 2007). It is based on the confinement of
multi-photon absorption to a very small volume because of its nonlinear dependence on
excitation intensity. This characteristic provides an income for activating chemical or
physical processes with high spatial resolution in three dimensions. As a result there is less
cross talk between neighbouring data layers. Another advantage of multi-photon excitation
is the use of infrared (IR) illumination, which results in the reduction of scattering and
permits the recording of layers at a deep depth in a thick material. Two-photon 3D bit
recording in photopolymerizable (Strickler & Webb, 1991), photobleaching (Pan et al., 1997;
Day & Gu, 1998) and void creation in transparent materials (Jiu et al., 2005; Squier & Muller,
1999) has been demonstrated with a femtosecond laser. Recording densities could reach
terabits per cubic centimeter. Nevertheless, these processes suffer from several drawbacks.
The index modulation associated with high bit density limits the real data storage volume
due to light scattering. The fluorescence can limit the data transfer rate and the lifetime of
the device.
Thanks to various available compositions, ease of implementation, stability and
transparency, both organic and inorganic materials are convenient for 3D data storage. To
be good candidates for 3D optical data storage, these materials must satisfy several norms
34 Data Storage

for the storing and the reading: Resistance to ageing due to temperature and recurrent
reading. Moreover, high-speed response for high rate data transfer, no optically scattering
for multilayer storage, erasure and eventually record with a grayscale to increase the data
density could be incontrovertible advantages.
Here, we present two particular media: A photosensitive glass (zinc phosphate glass
containing silver) in which the contrast mechanism is neither a change in refractive index
nor a change in absorption, but a change in the third-order susceptibility ((3)) induced by
femtosecond laser irradiation (Canioni et al., 2008); and a spin state transition material.

2. Photosensitive glass
2.1. Data storage medium: Photosensitive zinc phosphate glass containing silver
Glasses with composition 40P2O5-4Ag2O-55ZnO-1Ga2O3 (mol%) were prepared for 3D data
storage using a standard melt quench technique. (NH4)2HPO4, ZnO, AgNO3 and Ga2O3 in
powder form were used as raw materials and placed with the appropriate amount in a
platinum crucible. A heating rate of about 1°C.min-1 has been conducted up to 1000 °C. The
melt was then kept at this last temperature (1000 °C) from 24 to 48 hours. Following this
step, the liquid was poured into a brass mold after a short increase of the temperature at
1100°C in order to access the appropriate viscosity. The glass samples obtained were
annealed at 320 °C (55 °C below the glass transition temperature) for 3 hours, cut (0.5 to 1
mm-thick) and optically polished. The glass possesses an absorption cut-off wavelength at
280 nm (due to the silver ions associated absorption band around 260 nm) and emits
fluorescence mainly around 380 nm when excited at 260 nm. This intrinsic fluorescence is
due to Ag+ isolated in the glass (Belharouak et al., 1999).
The processed glass is highly photosensitive and was originally developed as a gamma
irradiation dosimeter (Schneckenburger et al., 1981; Dmitryuk et al., 1996). Following
exposure to gamma rays, the glass presents a broad UV absorption band, and when excited
by UV radiation, emits homogeneous fluorescence which intensity is proportional to the
irradiation dosage. This fluorescence is attributed to the presence of silver nanoclusters.

2.2. Silver nanoclusters formation following IR femtosecond irradiation

When an IR femtosecond laser is focused inside the photosensitive glass, a nonlinear
multiphoton interaction occurs (in this case 4-photon-absorption) at the vicinity of the focal
volume. Photoelectrons are ejected from the valence to the conduction band of the glass
(Jones & Reiss, 1977; Stuart et al., 1996; Keldysh, 1965). In a similar manner to the
mechanism utilized in a dosimeter (Schneckenburger et al., 1981; Dmitryuk et al., 1996), the
released electrons are trapped within a few picoseconds by silver Ag+ ions to form silver
Ag0 atoms. Then, due to the high laser repetition rate, the accumulation of the deposited
laser energy increases the local temperature leading to the Ag0 diffusion. Mobile Ag0 atoms
are trapped by the Ag+ ions to form silver clusters Agmx+ with the number of atoms m<10
and the ionization degree x. Since this is produced by a highly nonlinear interaction, the
silver nanoclusters can be distributed in 3D inside the glass in an area that is smaller than
the laser beam diameter (Sun et al., 2001). More details on the silver nanoclusters
distribution in this glass can be found in (Bellec et al., 2009). The photo-induced silver
nanoclusters have characteristic properties well adapted for the 3D optical data storage.
Optical data storage in photosensitive glasses and spin state transition compounds 35

2.3. Silver nanoclusters spectral and optical properties

The photosensitive glass is irradiated using a femtosecond laser source emitting 440 fs, 9
MHz repetition rate pulses at 1030 nm. The laser mode is TEM00, and the output polarization
is TM. The maximum output average power is close to 6 W, which results in a maximum
energy per pulse of 600 nJ. Acousto-optic ltering permits the tuning of the pulse energy
and the repetition rate for perfect control of the accumulated effect. The femtosecond laser is
focused using a reective 36 objective with a 0.52 numerical aperture (NA) (working
distance 15 mm) to a depth of 200 mm in the glass. The beam waist is estimated to be 1 mm.
Spectroscopic properties (i.e. absorption and emission spectra) of the silver nanoclusters are
shown in Fig 1. A transmission confocal setup, with a 36 - 0.52 NA reective objective, is
used for spectroscopy. A UV diode emitting at 405 nm with an appropriate emission lter,
and a white light source are used to obtain emission and absorption spectra. All spectra are
measured using a spectrometer equipped with a CCD camera. Figure 1(a) shows the
absorbance difference between the irradiated region (experimental conditions: Irradiance I =
6 TW.cm−2; Number of pulses N = 106) and the non-irradiated region. An absorption band
appears at 345 nm. The emission spectrum of the uorescent species when excited in the UV
is presented in Fig. 1(b). A wide band centered at 580 nm is observed. These spectral
properties can be attributed to the photoinduced silver nanoclusters (Dmitryuk et al., 1996;
Dai et al., 2007).

Fig. 1. (a) Differential absorbance spectrum between the irradiated and non-irradiated
regions; (b) Emission spectrum (excitation wavelength, 405 nm) of the irradiated region.
Experimental laser irradiation conditions: I = 6 TW.cm−2, N = 106. Differential absorbance
and emission spectra are assigned to laser induced Ag clusters.

The optical properties of the silver nanoclusters are studied for different irradiation
conditions by white light, epiuorescence and third-harmonic generation (THG)
microscopy. Thus, the glass is exposed to different irradiance levels (x axis) between 4
TW.cm−2 and 10 TW.cm−2 and a different number of pulses (y axis) from 102 to 106, as shown
on the experimental map sketch in Fig. 2(a). The sample is manipulated through patterning
of bits with a bit spacing of 20 mm using a precision xyz stage. Epiwhite light and
epiuorescence microscopy are performed with commercial microscopes. A transmission
confocal setup, with a 36 - 0.52 NA reective objective, is used for THG data collection. The
36 Data Storage

THG signal is ltered from the fundamental one by an emission bandpass lter @ (350 ± 50)
nm and is collected with a photomultiplier tube. In our case, THG is excited with the same
laser at low energy 10 nJ/pulse, but practically, a cheaper laser, such as a femtosecond ber
laser, could be used. Indeed, with a minimum energy of 0.1 nJ and an irradiance of 1010
W.cm−2 (corresponding to an average power of 10 mW with a 100 MHz repetition rate),
more than one third-harmonic photon by incoming pulse can be detected (Brocas et al.,
2004).

Fig. 2. Microscopy imaging of laser-induced species following the experimental map sketch.
(a) x axis, laser irradiance; y axis, number of laser pulses, 65 bits pattern; spacing, 20 mm.
(b) Epiwhite light microscopy image reveals linear refractive index modications; vertical
dashed line, damage threshold. (c) Epiuorescence microscopy image (excitation
wavelength, 365 nm; emission lter, (610 ± 40) nm). (d) THG image (excitation wavelength,
1030 nm; emission lter, (350 ± 50) nm); encircled area, data storage irradiation conditions (I
= 6TW.cm−2, N = 106).

Transmission, uorescence, and THG readout images of bits recorded with different
irradiance levels and laser shots are presented in figures 2(b)–2(d). Figure 2(b) shows
changes in refractive index. The damage threshold is achieved at an irradiance of 9 TW cm−2,
which is delimited by a vertical dashed line. At irradiances below this line, no apparent
modications are observed except for the high accumulated area in the upper part of Fig.
2(b). Nevertheless, in Fig. 2(c), we observe that uorescence is obtained in regions where the
refractive index is not modied. The same behavior is observed on the THG image in Fig.
2(d).
The THG image conrms that the induced Ag clusters absorb at 343 nm, corresponding to
Optical data storage in photosensitive glasses and spin state transition compounds 37

the resonant third-harmonic of the laser wavelength. Indeed, a coherent third-harmonic

signal could be generated by each center due to the resonant absorption. The irradiated
region is in the weak absorption regime and allows an enhancement of the third harmonic
signal (Shen, 1984; Barille et al., 2002). Moreover, the image contrast is proportional to |R(3)
– NR(3)|2, where R(3) and NR(3) are the third-order susceptibilities of the irradiated
(resonant) and non-irradiated (nonresonant) regions (Barille et al., 2002). The signal is due to
the electronic contribution to χ(3), since only the electronic polarization is able to quickly
respond to a high frequency all-optical eld excitation. Therefore, at the vicinity of the
nonlinear interface between the Ag clusters and the glass matrix, a coherent third-harmonic
signal can be generated under femtosecond laser excitation at 1030 nm.
Thus, as illustrated in Fig 3, the data can be stored inside the glass by femtosecond laser
irradiation below the refractive index modication threshold. Thanks to thermal cumulative
effects, stable Ag nanoclusters are created, giving rise to the formation of a nonlinear
resonant interface (Fig. 3(a)). THG readout becomes possible using the same laser with less
energy (Fig. 3(b)).

Fig. 3. Writing and reading processes. (a) Data are stored inside the photosensitive glass by
focusing an IR femtosecond laser. 1 bit corresponds to a silver nanoclusters aggregation
which presents no refractive index modifications. (b) Using the same laser with less energy,
the third-harmonic signal can be collected.

2.4. Reading process by third-harmonic generation

To demonstrate the 3D optical data storage performance according to the principle
38 Data Storage

explained before, a 3D bit pattern embedded in the photosensitive glass is written and read
by a THG imaging setup. To achieve a high bit density in the volume, the change in
refractive index must be kept as low as possible to minimize the effect of scattering of the
reading beam while the change in (3) must be as high as possible. We choose to work below
the damage threshold to minimize the refractive index modication and optimize creation
of aggregates by the accumulated effect (the corresponding area is encircled in Fig. 2(d)).
The sample is irradiated with a laser irradiance of 6 TW.cm−2 and with 106 pulses. Three
layers of data are embedded 200 mm inside the sample. Each layer contains a pattern of
1212 bits with a bit spacing of 3 mm. The letters U, B, and the numeral 1 (for University
Bordeaux 1) are recorded in the rst, second, and third layers, respectively, with a layer
spacing of 10 mm in the z direction. As shown previously, the same laser is used for the
reading procedure but with a lower irradiance. By scanning the sample in xyz through the
focus, the three layers (U, B, and 1) are reconstructed and presented in Fig. 3.

Fig. 4. THG readout of the three layers containing the bit patterns U, B, and 1, recorded in
the bulk of the glass (bit spacing, 3 mm; layer spacing, 10 mm). Laser writing parameters: I =
6 TW.cm−2, N = 106. The three THG images present high signal-to-noise ratio and no cross
talk.

As expected by the THG mechanism, an image with high contrast and no cross talk is
observed in Fig. 3. The main advantages of this technique compared to usual 3D data
storage are no photobleaching, no change in linear refractive index, and therefore no
scattering. Moreover, the THG signal is coherent and gives rise to a rather intense,
directional, and less divergent beam with a high signal-to-noise ratio. Due to the fast THG
response, the reading speed is limited only by the pulse duration.

2.5. Grayscale encoding

Fig. 2(d) shows that the THG signal depends on the number of pulses. This can be used to
encode the information onto a grayscale. Using an acousto-optic modulator, the number of
pulses is well controlled. As presented in the Fig. 4, the THG response for an irradiance I = 6
TW.cm-2 is linear with the number of pulses between 1104 and 5 104 pulses. Considering
the readout error, the encoding becomes possible with a 16 levels grayscale (4 bits).
Optical data storage in photosensitive glasses and spin state transition compounds 39

Fig. 5. THG signal versus number of the laser pulses with I = 6 TW.cm-2. The information
can be grayscale encoded.

2.6. 3D optical data storage stability

The photo-written glasses were submitted to different thermal treatments below the
transition temperature of the glass Tg = 375 °C at 100 °C, 200 °C, 300 °C, 350 °C during 3 h
and above Tg at 400 °C during 20 min. The THG signal is not modied for temperatures
below Tg but disappears for temperatures higher than Tg. The recording is erased only after
thermal treatment at 400 °C during 20 min corresponding to a total reorganization of the
glass. Then, bits could be rewritten inside the glass after polishing. Thus we can state that
recording is very stable in standard conditions (up to 85 °C). In our reading conditions, 10
nJ/pulse, the THG signal is not modied, even after several hours of unstopped exposure
(corresponding to 1010 readings). Indeed, the Ag clusters are created only if enough
photoelectrons are produced (threshold process). This is the reason why no modication
appears during the reading process.

2.7. 3D optical data storage capacity

In our experimental conditions (bit spacing, 3 mm; layer spacing, 10 mm), 1 Gbit.cm−3 could
be stored inside the glass. Better performances can be achieved with a high NA objective
(NA = 1.4) even if the objective working distance limits the number of layers. According to
Chen & Xie (Chen and Xie, 2002), the THG can reach radial and axial resolutions of 250 and
400 nm, respectively. Thus a storage capacity in the range of Tbits.cm−3 is possible in the
glass. Considering the grayscale encoding, this value can be increased.
40 Data Storage

3. Spin state transition materials

Spin-crossover systems are a paradigmatic example of bistable optical materials. In such
materials, the optical properties can be changed from one state to another through external
stimuli such as temperature (T), pressure (P) or light (h). As the system switches between
the two states, its optical properties drastically change in the visible and near IR spectral
ranges. These properties were demonstrated in many molecular samples at low
temperature. It is only very recently that we have shown that pulsed light irradiations make
possible to switch the optical properties of such bistable systems at room temperature. The
mechanism responsible for the switching in these particles has also been elucidated. It
indicates that less than a thousand photons are required to photo-switch a single
nanoparticle of 60 nm in diameter. The offered prospects by such materials in optical and
material sciences are therefore very rich and promising.

3.1. Data recording within the hysteresis loop at low temperature

In the last thirty years, the interest in molecular materials that makes possible the storing of
a digital information has considerably increased. In such a context, the molecular bistability
associated to the spin-crossover (SCO) phenomenon has triggered important research
activities. Indeed, the change of the spin state in a SCO material can be induced reversibly in
a solid state by adjusting or applying different constraints such as temperature, pressure,
light intensity or pulsed magnetic field (Gutlich et al., 1994; Bousseksou et al., 2000). The
SCO phenomenon induces for instance an important and reversible change of the optical
absorbance or of the magnetic susceptibility of the studied medium. The first photomagnetic
effect in an iron(II) SCO material was reported by Decurtins et al. in 1984. At low
temperature (T 50 K), these authors demonstrated the possibility to convert a compound
with a low spin (LS, S=0) state into a compound with a metastable high spin (HS, S=2) using
green light laser beam irradiation. This phenomenon, called Light-Induced Excited Spin-
State Trapping (LIESST), was extended later by Hauser, who showed that red light switches
the system back to the LS state (reverse-LIESST) (Hauser, 1996). This effect triggered a lot of
interest, but most of the published studies have failed to demonstrate the LIESST effect at
high temperature. In fact, above 50 K, the photo-induced HS state usually decays within a
few minutes. To overcome this limitation, an interesting approach would be to work within
the thermal hysteresis loop of a SCO material. Indeed, within the hysteresis loop, both HS
and LS states are thermodynamically stable. The first spin change reported in a thermal
hysteresis was, in fact, induced using intense magnetic pulses (Bousseksou et al., 2000).
However, for practical applications, the use of such external perturbation is not realistic. The
first influence of light on a thermal hysteresis loop was described by Renz et al. in 2000. This
effect, named Light Perturbed Thermal Hysteresis (LiPTH) was observed under continuous
laser irradiation. Depending on the laser wavelength, the hysteresis loop was shifted
towards lower temperature or higher temperature. However, the phenomenon relaxes as
soon as the laser light is switched-off and considerably limits its application for data storage.
Recently, light-induced valence tautomeric conversion was reported in some cobalt-iron
Prussian Blue analogs at room temperature (Shimamoto et al., 2002; Liu et al., 2003). It was
shown that a single laser pulse converts the electronic state from FeII(S=0)-CN-CoII(S=0) to
FeIII(S=1/2)-CN-CoII(S=3/2). The reverse phenomenon has also been observed (Liu et al.,
2003). This phenomenon was explained using the so-called domino effect (Koshino &
Ogawa, 1998): When the concentration of the local excited HS state is higher than a critical
Optical data storage in photosensitive glasses and spin state transition compounds 41

value, all the system commutes from an LS to an HS state. According to these experiments,
one may wonder if a single laser pulse may also induce the same phenomena in a
cooperative iron(II) SCO material. To demonstrate it, we have selected the [Fe(PM-
BiA)2(NCS)2] (PM-BiA = N-2’-pyridylmethylene-4-aminobiphenyl) complex which exhibits a
well defined abrupt hysteresis around 170 K (Létard et al., 1998; Létard et al., 2003). We have
shown that, under certain conditions, the use of a single pulse laser in the center of the
thermal hysteresis loop leads to a LS  HS photo-conversion (Freysz et al., 2004). Compared
to the experiments reported on valence tautomeric compounds, our results indicated that
the final state reached after a laser excitation is neither a pure HS nor a pure LS state, but it
is instead a “mixture” of HS/LS domains. Since the system is firstly photo-excited into the
HS state and then slowly relaxes to mixture of HS/LS state, our results cannot be accounted
by the so-called domino effect.
The set-up we used to perform these experiments is sketched on Figure 6a. The sample, a
powder composed of micro crystallites (a few microns in radius) is sandwiched in between
two optical windows and is placed into a cryostat. The specular light reflected by the sample
was collected and sent to a 150 mm spectrometer to select the wavelength centered at 600
nm. The resolution of the spectrometer was set to be ~ 2 nm. At the exit of the spectrometer,
the light was collected by a photomultiplier connected to a 1 Mega Ohm load. The voltage
drop across this load was recorded versus the temperature of the sample in the cryostat. As
shown in Figure 6b, this reflection set-up makes possible to record the temperature
hysteresis loop of the studied sample. To first record the LS to HS state transitions, the
sample is shined with a white light continuum and the light reflected by the sample and
transmitted at 600 nm through the spectrometer is recorded by the photo-multiplier. The
typical evolution of the reflectivity versus the temperature increase or decrease is presented
in Figure 6b. These data are in very good agreement with the measurements performed with
a SQUID.

1,0
(a) (b)
Trigger
0,8 -
P.D.
0,6
Cryostat
xHS

White light
continuum 0,4
F1
L1

+
0,2

F2 G1 0,0
P.M. Monochromator
162 164 166 168 170 172 174 176 178
=630 nm Nd:YAG
Temperature (K)
L2
2

Fig. 6. (a) Sketch of the experimental set-up used to induce and measure the laser induced
spin state transition. (b) Evolution of the reflected light versus the temperature. The
temperature is either decreased from 180 K to 160 K or increased from 160 K To 180 K. XHS is
the molar fraction of molecules in the HS state.

Using this set-up, we have studied the effect of pulsed light on the sample. In this later case,
a single laser pulse (Q-switched nanosecond frequency doubled Nd3+:YAG laser, =0.532
µm, pulse width 8 ns; energy between 0.5 and 10 mJ) was focused on a spot of about 3 mm
42 Data Storage

in diameter and only a small fraction of the light reflected by the sample at 600 nm within
the laser spot area was imaged and recorded by the spectrometer. As shown in Figure 7a,
when the sample is cooled at 140 K, a temperature below the hysteresis loop, we clearly
record a temporal evolution of the reflectivity of the sample. According to our calibration,
our data clearly show that all the sample probed by the reflected light is brought into the HS
state. Moreover, within 0.5 second, it relaxes back to the LS state. By studying the evolution
of the HS fraction within the laser spot versus the energy of the laser pulse, we note that the
number of HS particles steadily increases versus the energy of the laser pulse up to an
energy of ~ 1 mJ. Above 1 mJ, which corresponds to a fluence of 14 mJ.cm-2 per pulse, all the
probed molecules are photo-converted in HS state and the signal saturates. This situation
remains until 9 mJ, when a surface photo-degradation of the sample occurs. In conclusion,
below the thermal hysteresis and between 78 and 140 K, the photo-induced HS state is not
stable. We noted its lifetime decreases from hours (at 60 K) to minutes (at 78 K) and to
second (at 140 K) (Degert et al., 2005).
1,0

T=140 K (a) 1,0 T = 170 K (b)

0,8
0,8

0,6
0,6
x HS

xHS

0,4
0,4

0,2 0,2

0,0 0,0
0,0 0,2 0,4 0,6 0,8 1,0 1,2 1,4 1,6 1,8 0 5 10 15

Fig. 7. HS  LS relaxation curves (a) At 140 K below the hysteresis Time

Time (s)
loop.(s)
(b) At 170 K within
the hysteresis loop. The sample was firstly cooled down to a pure LS state and then carefully
warmed

Let us now consider the influence of a 1 mJ laser pulse when the sample is set at a
temperature within the thermal hysteresis loop. The sample was firstly set into the LS
configuration by slowly cooling it and then carefully warming it at T1= 166 K, i.e. the region
at the beginning of the thermal hysteresis loop. The sample was then excited with a single
laser pulse. As reported on Figure 7a, immediately after the single laser pulse, all the probed
particles reach the HS state, but they relax back to the LS state. This clearly indicates firstly,
that the photo-induced excited HS state is not stable, even within the hysteresis loop and
secondly, that the LS  HS transition cannot be induced by the photo-excitation of HS
molecules. This observation also considerably limits the contribution of a domino effect
associated with the excitation of the molecular HS state under our experimental conditions.
We then warmed the sample at 170K, i.e. the region in the centre of the thermal hysteresis
loop, and again excited it with a single laser pulse. Similarly to what reported on Figure 7b,
immediately after the single laser pulse, all the probed particles reach the HS state. But this
time, they neither relax toward the pure HS state nor the initial LS state. The final situation
after relaxation can be regarded as a “mixture” of HS/LS particles. At first, this result is
interesting in regard to the opened problem of the thermal effect accompanying the optical
excitation of the sample. Indeed, if an artificial and important temperature increase (T > 10
K) associated with the absorption of the particles brought the sample in the HS state, it
Optical data storage in photosensitive glasses and spin state transition compounds 43

should, at first glance, remain within this state. This is clearly not the reported experimental
result. Moreover, we have also shown that additional laser pulses do not affect the
measured HS/LS ratio (Freysz et al., 2004). We have also demonstrated the stability of the
photo-induced mixture state within the hysteresis loop. When one steadily increases the
temperature of the sample, it remains within this “mixture” state until it reaches the upper
and lower temperature branch of the hysteresis loop, respectively labelled T1/2↓ and T1/2.
For temperatures higher than T1/2, all the particles of the sample are converted in the HS
state. This latter result demonstrates, if necessary, that the sample is not damaged by the
laser pulse. This point was confirmed by the repetition of the experiment many times on the
same sample, which did not show any degradation.
Although these experiments were interesting in demonstrating optical data recording in
spin state compounds within the hysteresis loop, their practical use was quite limited.

3.2. Data recording within the hysteresis loop and at room temperature
One year after our work, it has been shown that different compounds can be photo-switched
at room temperature using a single Nd:YAG Q-switched laser pulse (Bonhommeau et al.,
2005). However, the details of the mechanisms giving rise to the switching of the spin state
compound within the thermal hysteresis was difficult to evidence (Fouché et al., 2009). In
order to investigate these mechanisms, we have carried out an optical study of the SCO
complex [Fe(NH2trz)3](NO3)2 3H2O coordination polymer (NH2trz = 4-amino-1,2,4-triazole)
known to display a well-defined thermal hysteresis loop at room temperature.

(a) (b) (c)

Fig. 8. [Fe(NH2trz)3](NO3)2 3H2O sample in the LS state (a), the HS state (c) and after photo-
excitation by a single laser pulse within the hysteresis loop of the compound (b).

As shown in Fig. 8, the colour of the powder, which is pink when the sample is in the LS
state (Fig 8a), becomes almost completely white when the sample is in the HS state (Fig. 8c).
Indeed, in the LS state, this material displays a broad absorption band at around 520 nm,
which is characteristic of the d-d transition of Fe2+ (1A1g → 1T1g), while in the HS state a d-d-
transition is recorded at lower energy in the near-infrared region (830 nm). Figure 8b shows
the state of the sample impacted by a sequence of pulses yielded by a frequency tripled
Nd:YAG laser that delivers pulses at 355nm which duration is ~ 6ns and fluence on the
sample is Ep ~ 52 mJ.cm-2. The same experiment has also been performed using pulses with
the same fluence but centred at 532 nm. In both cases, we noticed that the central part of the
sample impacted by the laser pulse becomes white. The sample remains as it is as long as its
temperature is kept within the thermal hysteresis loop. For instance, the whole sample
recovers its original pink colour when its temperature is below 10 °C. This clearly indicates
that within the hysteresis loop, the micro-crystallites of the compound impacted by the laser
pulses are brought in the HS state.
44 Data Storage

Fig. 9. (I) and (VII) colour of the sample in the LS state. Evolution of the sample set in the
thermal hysteresis loop and impacted by (II) one pulse, (III) two pulses, (IV) three pulses,
(V) four pulses, (VI) five pulses. (VIII) Colour of the sample impacted by five laser pulse and
cooled back in the LS state.

We have also recorded the evolution of the sample versus the laser excitation. Firstly, we
noticed that below a certain laser fluence, the sample remains unchanged. Above this
threshold fluence, the final state of the compounds depends on both the laser fluence and
the number of laser pulses used to excite the sample. As shown in Fig. 9, the central part of
the sample shined by the laser beam that is initially in the LS state (I) becomes whiter and
whiter as the number of excitation pulses increases ( Fig. 9 II to 9 VI). The pictures in Fig.
9VII and 9VIII underline that, for temperature above or below the hysteresis loop, the
sample recovers its HS or LS state. This indicates that the sample is not altered by the laser
excitation. This experiment also indicates that grayscale encoding is also possible in these
materials.

Fig. 10. In solid line: Thermal hysteresis loop deduced from the optical reflectivity and
irradiation studies performed at 328 K, 333 K and 338 K by one (), two (), three (), four
() and five () pulses at 355 nm (in blue) and 532 nm (in green) with a fluence of 52
mJ.cm-2. The inset is the absorption of the powder in the HS and LS states.
Optical data storage in photosensitive glasses and spin state transition compounds 45

To be more quantitative about the final state reached by the sample after each laser pulse,
we have replaced the monochromator by a spectrometer and we have recorded the
spectrum of the light reflected by the sample after each laser pulse. This enables to fix the
state of the sample within the hysteresis loop (Fig. 10). These last data clearly indicate that
one can easily perform data recording in this compound. However, they also stress that data
recording is easier when the absorption coefficient of the sample is higher. This also clearly
underlines that the laser induced heating of the sample is of central importance. To clearly
evidence the mechanism responsible for data recording in these materials, we have
performed a nanosecond time resolved reflectivity measurement that is described below.

Oven
G1 2
L1
Boxcar Nd:YAG

P.D.1 Synchronization

OPO

P.D.2 L2 F1

Fig. 11. Nanosecond time resolved pump-probe set-up.

3.2. Kinetic of the data recording

3.2.1. Pump-probe time resolved nanosecond set-up
The optical set-up depicted in Fig. 11 has been designed to record the kinetic of the LS to HS
state transition with a nanosecond time resolution. The sample placed in an oven is excited
by a "pump" pulse provided by a nanosecond Optical Parametric Oscillator (OPO Panther
EX from Continuum), this latter being pumped by a Q-switched-nanosecond-frequency-
tripled Nd:YAG laser emitting at = 355 nm (Surelite II from Continuum). The signal and
the idler pulses (pulse duration pump ~ 6 ns), produced by a parametric frequency
conversion in a type II BBO crystal, are spatially separated by dichroïc mirrors. The OPO is
tuneable from 400 to 680 nm and from 720 to 2500 nm. At = 400 nm, the maximum pulse
excitation energy is ~ 20 mJ. In practice, we adjusted the fluence on the sample by focusing
more or less tightly the laser beam. To measure the sample reflectivity change (R) induced
by the pump pulse, a "probe" pulse generated by a frequency doubled and Q-switched
Nd:YAG laser (probe ~ 6 ns) was used. The probe pulse light diffused by the sample at an
angle of 30° was collected by a fast photodiode which output signal is sent in a first boxcar
gated integrator. To compensate for pulse to pulse energy fluctuations, the probe pulse
reflected by an uncoated optical window placed in front of the sample is used as a reference
pulse. It is collected by a second fast photodiode which output signal is connected to a
second boxcar gated integrator. The relative sample reflectivity is determined by dividing
the signal delivered by the two photodiodes. To control the delay between the pump and
probe pulses, the two Nd:YAG lasers are synchronized using a home build-electronic
46 Data Storage

device. This device, that also delivers the signal that triggers the boxcar gates, makes it
possible to temporally delay, from few nanoseconds up to few seconds, the light pulses
delivered by these two systems. To limit the impact of OPO fluctuations, we measure the
energy of each OPO pulse and only kept the data corresponding to an OPO pulse energy
within  5% of the mean OPO energy value. The signal to noise ratio of our data is further
improved by averaging the data over ten laser shots. Finally, to measure the actual
reflectivity change of the sample, we record for each pump-probe time delay the relative
reflectivity of the sample with and without the pump pulse. In the actual set-up, the sample,
a powder composed of micro crystallites a few microns in radius, is sandwiched in between
two optical windows and placed into a thermally regulated oven. This set-up has been
mainly used when the temperature of the sample is set slightly below the hysteresis loop. In
such a case the sample recovers its initial LS state after each pump pulse. Therefore, under
these experimental conditions, one records both the formation and the relaxation of the HS
fraction induced by the pump pulse.

3.2.2. Experimental results

The reflectivity change (R ~ 8%) induced by a pump pulse fluence of Ep ~ 18 mJ.cm-2 is
presented in Fig. 12 for a sample at 283 K.

1.10

1.08 Pump pulse

Reflectivity

C D
1.06

1.04 B

1.02
A A
1.00
1 2 3 4 5 6 7
10 10 10 10 10 10 10
Log(t(ns))
Fig. 13. Time dependence of the reflectivity change at 283 K after photo-excitation with a
pump pulse of fluence Ep ~ 9 mJ.cm-2.

To grasp all the complexity of the dynamic, we have chosen to present our data on a
temporal logarithmic scale. Using this scale, one clearly evidences that the growing and
the relaxation of the reflectivity is governed by different characteristic times. Figure 12
also indicates that, to register the complete evolution of the system after the pulse
excitation, the reflectivity has to be recorded over at least five decades of time. This
stresses the advantage of our experimental set-up in which one can adjust continuously
the sampling steps along the experiment.
Optical data storage in photosensitive glasses and spin state transition compounds 47

3.2.3. Temporal evolution of the sample within the phase diagram

We have demonstrated that the dynamic of the reflectivity change presented in Fig. 12
can be assigned to respectively:
 The heating of a thin layer (~ 400 nm) of the sample for 0 < t  200 ns.
 The diffusion of the heat toward the bulk sample and the growing of the HS
fraction within the thin layer for 200 ns  t  100 µs.
 The relaxation of the HS fraction within the thin layer for 100 µs  t  100 ms.

C Tdown Tup

(1)
D C
B
A

T (K)

T(t) B (2)
C D Tup
Tdown
A
t

C(t)
(3)

A
t
1 ~40 ns 2 ~100 µs

Fig. 14. Mechanisms that account for the laser pulse induced SCO at the vicinity of the
thermal hysteresis loop.

As depicted in Fig. 14, this experiment makes also possible to follow the temporal evolution
of the sample within the HS concentration, temperature (i.e. (C,T)) phase diagram of the
compound outside the hysteresis loop (Fouché et al., to be published). Upon laser
excitation, absorption of the pump energy takes place in a very thin layer Lp (~ 400 nm).
As a result, the temperature of this layer increases drastically and is maximum about 40
ns after the laser excitation and sets the sample layer at the point B of the (C,T) diagram.
The heat subsequently diffuses toward the bulk sample. Hence, the temperature of the
layer steadily decreases so that the sample layer is moving within the phase diagram.
The evolution of the HS fraction is less abrupt. During the heating of the layer, it
48 Data Storage

remains almost constant and then it starts to increase. The growth lasts as long as the
temperature of the layer has not reached the ascending branch of the hysteresis loop
(point C of the phase diagram). Then, as long as the temperature of the layer remains
within the thermal hysteresis loop (i.e. between the point C and D of the phase
diagram), the HS fraction remains constant. It decreases as soon as the temperature of
the layer reaches the descending branch of the hysteresis loop (point D of the phase
diagram).

3.2.4. Mechanisms responsible for data recording within the hysteresis loop
The results recorded when the temperature of the sample was set below the hysteresis loop
and presented in the previous section makes possible to understand the evolution of the
sample when its temperature is set within the hysteresis loop as depicted in Fig.9. Figure 15
shows the evolution of the sample after each laser pulse.

Fig. 15. Path in the phase diagram followed by the sample after each laser shot and
evolution of the temperature (grey line) and the HS fraction (black line) of the sample after
each laser shot.

At first, just after the laser excitation, the absorption of the pulse by a thin layer of the
sample induces a local heating. Within a hundred of nanoseconds, the absorbed energy
heats the system; its temperature rises up very quickly at a temperature where the LS state is
unstable. The higher the absorption coefficient is, the higher the temperature increase is.
Thus the sample is thermally quenched (A → B), whereas only a very small HS fraction
appears. Then, the sample cools and an increase of the HS fraction takes place (B → C). As
shown in Fig. 14, this process lasts about 100 µs. Once the ascending branch of the hysteresis
loop is reached, the HS fraction no longer evolves and the created HS fraction coexists with
the LS fraction. Finally, the powder reaches the temperature of the oven (C → D). The whole
Optical data storage in photosensitive glasses and spin state transition compounds 49

process takes about 1 ms. The sample has then reached a state where its absorption
coefficient decreases. Indeed, as shown in inset of the Fig. 10, the absorption of the sample in
the HS state is reduced at both 532 nm and 355 nm. Thus, for the subsequent laser pulses,
the laser induced heating is less and less important, the growth lasts less and less time and
the increase of the HS fraction is less and less important. After a few pulses, the sample has
reached a state located within the thermal hysteresis loop (Fig. 14). In the last case, one can
further improve the LS to HS conversion efficiency by increasing the laser pulse energy.
Based on this scenario, the highest efficiency recorded at 355 nm is directly accounted by the
highest absorption of the sample in the blue spectral range (inset of Fig. 10), leading to a
greater heating of the sample.

3.3. Future and prospects

Our on going study deals with bistable nanoparticles with controlled sizes, shapes and
optical properties. Such particles offer many prospect for data recording. Our computations
indicate that less than a thousand of photons are required to photo-switch a single
nanoparticle of 60 nm in diameter. On the other hand, we are also performing experiments
that make possible to understand and to master the physical mechanisms responsible for the
photo-switching of these nanoparticles inserted in different controlled polymeric
environments. We are also currently testing their practical uses for holographic data
recording.

4. References
Barille, R.; Canioni, L.; Sarger, L. & Rivoire, G. (2002). Nonlinearity measurements of thin
films by third-harmonic-generation microscopy, Phys. Rev. E, 66, 067602
Belharouak, I.; Parent, C.; Tanguy, B.; Le Flem, G. & Couzy, M. (1999). Silver aggregates in
photoluminescent phosphate glasses of the ‘Ag2O-ZnO-P2O5’ system, J. Non-Cryst.
Solids., 244, 238-249
Bellec, M; Royon, A.; Bousquet, B.; Bourhis, K.; Treguer, M.; Cardinal, T.; Richardson, M. &
Canioni, L. (2009), Beat the diffraction limit in 3D direct laser writing in
photosensitive glass, Opt. Exp., 17, 10304-10318
Bousseksou, A; Negre, N.; Goiran, M.; Salmon, L.; Tuchagues, J.P.; Boillot, M.L.;
Boukheddaden, K. & Varret, J. F. (2000), Dynamic triggering of a spin-transition by
a pulsed magnetic field, Eur. Phys. J. B, 13, 451
S. Bonhommeau, G. Molnar, A. Galet, A. Zwick, J.A. Real, J.J. McGarvey & A. Bousseksou
(2005), One-Shot-Laser-Pulse-Induced Reversible Spin Transition in the Spin
Crossover Complex {Fe(C4H4N2)[Pt(CN)4]} at Room Temperature, Angew. Chem.
Int. Ed., 44 4069
Brocas, A.; Canioni, L. & Sarger, L. (2004). Efficient selection of focusing optics in non linear
microscopy design through THG analysis, Opt. Exp., 12, 2317-2322
Canioni, L.; Bellec, M; Royon, A.; Bousquet, B. & Cardinal, T. (2008). Three-dimensional
optical data storage using third-harmonic generation in silver zinc phosphate glass.
Opt. Lett., 33, 360-362
Chen, J. & Xie, S. (2002). Green's function formulation of third-harmonic generation
microscopy, J. Opt. Soc. Am. B, 19, 1604
50 Data Storage

Dai, Y.; Hu, X.; Wang, C.; Chen, D.; Jiang, X.; Zhu, C.; Yu, B. & Qiu, J. (2007). Fluorescent Ag
nanoclusters in glass induced by an infrared femtosecond laser, Chem. Phys. Lett.,
439, 81
Day, D. & Gu, M. (1998). Effects of Refractive-Index Mismatch on Three-Dimensional
Optical Data-Storage Density in a Two-Photon Bleaching Polymer, Appl. Opt., 37,
6299-6304
Decurtins, S.; Gütlich, P.; Köhler, C.P.; Spiering, H. & Hauser, A. (1984), Light-induced
excited spin state trapping in a transition-metal complex: The hexa-1-
propyltetrazole-iron (II) tetrafluoroborate spin-crossover system, Chem. Phys. Lett.,
105, 1-4.
Degert, J.; Lascoux, N.; Montant, S.; Létard, S.; Freysz, E.; Chastanet, G. & Létard, J.-F. (2005),
Complete temperature study of the relaxation from the high-spin state to low-spin
state in a strongly cooperative spin crossover compound, Chem. Phys. Lett. 415, 206-
210
Dmitryuk, A. V.; Paramzina, S. E.; Perminov, A. S.; Solov’eva, N. D. & Timofeev, N. T.
(1996). The influence of glass composition on the properties of silver-doped
radiophotoluminescent phosphate glasses, J. Non-Cryst. Solids, 202, 173-177
Fouché, O.; Degert, J.; Jonusauskas, G.; Baldé, C.; Desplanche, C.; Létard, J.-F. & Freysz, E.
(2009), Laser induced spin state transition: Spectral and temporal evolution, Chem.
Phys. Lett., 469, 274-278
Freysz E, Montant S, Letard S & Létard J.F. (2004), Single laser pulse induces spin state
transition within the hysteresis loop of an Iron compound, Chem. Phys. Lett., 394,
318-323
Jiu, H.; Tang, H.; Zhou, J.; Xu, J.; Zhang, Q.; Xing, H.; Huang, W. & Xia, A. (2005).
Sm(DBM)3Phen-doped poly(methylmethacrylate) for three-dimensional
multilayered optical memory, Opt. Lett., 30, 774-776
Gütlich, P.; Hauser, A. & Spiering, H. (1994), Thermal and Optical Switching of Iron(II)
Complexes, Angew. Chem. Int. Ed. Engl., 33, 2024-2054
Hauser, A. (1986), Reversibility of light-induced excited spin state trapping in the
Fe(ptz)6(BF4)2, and the Zn1−xFex(ptz)6(BF4)2 spin-crossover systems, Chem. Phys.
Lett. 124, 543-548
Jones, H. D. & Reiss, H. R. (1977). Intense-field effects in solids, Phys. Rev. B, 16, 2466-2473
Keldysh, L. V. (1965). Ionization in the field of a strong electromagnetic wave, Sov. Phys.
JEPT, 20, 1307-1314
Koshino, K. & Ogawa, T (1998), Domino effects in photoinduced structural change in one-
dimensional systems, J. Phys. Soc. Jap. 67, 2174
Létard, J.-F.; Guionneau, P.; Rabardel, L.; Howard, J. A. K.; Goeta, A. E.; Chasseau, D. &
Kahn, O. (1998), Structural, Magnetic, and Photomagnetic Studies of a Mononuclear
Iron(II) Derivative Exhibiting an Exceptionally Abrupt Spin Transition. Light-
Induced Thermal Hysteresis Phenomenon, Inorg. Chem, 37, 4432-4441
Létard, J.-F.; Chastanet, G.; Nguyen, O.; Marcen, S.; Marchivie, M.; Guionneau, P.; Chasseau,
D.; Gütlich, P. (2003), Spin Crossover Properties of the [Fe (PM-BiA) 2 (NCS) 2]
Complex-Phases I and II, Monatshefte für Chemie 134, 165
Li, X.; Bullen, C.; Chon, J. W. M.; Evans, R. A. & Gu, M. (2007). Two-photon-induced three-
dimensional optical data storage in CdS quantum-dot doped photopolymer, Appl.
Phys. Lett., 90, 161116
Optical data storage in photosensitive glasses and spin state transition compounds 51

Liu, H.W.; Matsuda, K.; Gu, Z.Z.; Takahashi, K.; Cui, A.L.; Nakajima, R.; Fujishima, A. &
Sato, O. (2003), Reversible Valence Tautomerism Induced by a Single-Shot Laser
Pulse in a Cobalt-Iron Prussian Blue Analog, Phys. Rev. Lett. 90, 167403
Mok, F. K. (1993). Angle-multiplexed storage of 5000 holograms in lithium niobate, Opt.
Lett., 18, 915-917
Pan, S.; Shih, A.; Liou, W.; Park, M.; Bhawalkar, J.; Swiatkiewicz, J.; Samarabandu, J.; Prasad,
P. N. & Cheng, P. C. (1997). Scanning 19, 156
Psaltis, D.; Levene, M.; Pu, A. & Barbastathis, G. (1995). Holographic storage using shift
multiplexing, Opt. Lett., 20, 782-784
Rakuljic, G. A.; Leyva, V. & Yariv, A. (1992). Optical data storage by using orthogonal
wavelength-multiplexed volume holograms, Optics Letters, 17, 1471-1473
Renz, F.; Spiering, H.; Goodwin, H.A. & Gütlich, P (2000), Light-perturbed hysteresis in an
iron (II) spin-crossover compound observed by the Mössbauer effect, Hyperfine
Interactions 126, 155
Schneckenburger, H.; Regulla, D. F. & Unsöld, E. (1981). Time-resolved investigations of
radiophotoluminescence in metaphosphate glass dosimeters, Appl. Phys. A, 26, 23-
26
Shen, Y. R. (1984). The Principles of Nonlinear Optics, Wiley
Squier, J. & Muller, M. (1999). Third-harmonic generation imaging of laser-induced
breakdown in glass, Appl. Opt., 38, 5789-5794
Strickler, J. H. & Webb, W. W. (1991). Three-dimensional optical data storage in refractive
media by two-photon point excitation, Opt. Lett., 16, 1780-1782
Stuart, B. C.; Feit, M. D.; Herman, S.; Rubenchik, A. M.; Shore, B. W. & Perry, M. D. (1996).
Nanosecond-to-femtosecond laser-induced breakdown in dielectrics, Phys. Rev. B,
53, 1749-1761
Sun, H. B.; Tanaka, T.; Takada, K. & Kawata, S. (2001). Finer features for functional
microdevices, Nature, 412, 697-698
Shimamoto, N.; Ohkoshi, S.-S.; Sato, O. & Hashimoto, K. (2002), One-Shot-Laser-Pulse-
Induced Cooperative Charge Transfer Accompanied by Spin Transition in a Co-Fe
Prussian Blue Analog at Room Temperature, Chem. Lett., 31, 486
52 Data Storage
Data Representation for Flash Memories 53

0
4

Data Representation for Flash Memories

Anxiao (Andrew) Jiang
Computer Science and Engineering Dept.
Texas A&M University
College Station, TX 77843, U.S.A.
[email protected]

Jehoshua Bruck
Electrical Engineering Department
California Institute of Technology
Pasadena, CA 91125, U.S.A.
[email protected]

In this chapter, we introduce theories on data representation for flash memories. Flash mem-
ories are a milestone in the development of the data storage technology. The applications of
flash memories have expanded widely in recent years, and flash memories have become the
dominating member in the family of non-volatile memories. Compared to magnetic record-
ing and optical recording, flash memories are more suitable for many mobile-, embedded-
and mass-storage applications. The reasons include their high speed, physical robustness,
and easy integration with circuits.
The representation of data plays a key role in storage systems. Like magnetic recording and
optical recording, flash memories have their own distinct properties, including block erasure,
iterative cell programming, etc. These distinct properties introduce very interesting coding
problems that address many aspects of a successful storage system, which include efficient
data modification, error correction, and more. In this chapter, we first introduce the flash
memory model, then study some newly developed codes, including codes for rewriting data
and the rank modulation scheme. A main theme is understanding how to store information
in a medium that has asymmetric properties when it transits between different states.

1. Modelling Flash Memories

The basic storage unit in a flash memory is a floating-gate transistor [3]. We also call it a cell.
Charge (e.g., electrons) can be injected into the cell using the hot-electron injection mechanism
or the Fowler-Nordheim tunnelling mechanism, and the injected charge is trapped in the cell.
(Specifically, the charge is stored in the floating-gate layer of the transistor.) The charge can
also be removed from the cell using the Fowler-Nordheim tunnelling mechanism. The amount
of charge in a cell determines its threshold voltage, which can be measured. The operation of
injecting charge into a cell is called writing (or programming), removing charge is called erasing,
and measuring the charge level is called reading. If we use two discrete charge levels to store
data, the cell is called single-level cell (SLC) and can store one bit. If we use q > 2 discrete
charge levels to store data, the cell is called multi-level cell (MLC) and can store log2 q bits.
54 Data Storage

A prominent property of flash memories is block erasure. In a flash memory, cells are organized
into blocks. A typical block contains about 105 cells. While it is relatively easy to inject charge
into a cell, to remove charge from any cell, the whole block containing it must be erased to the
ground level (and then reprogrammed). This is called block erasure. The block erasure opera-
tion not only significantly reduces speed, but also reduces the lifetime of the flash memory [3].
This is because a block can only endure about 104 ∼ 106 erasures, after which the block may
break down. Since the breaking down of a single block can make the whole memory stop
working, it is important to balance the erasures performed to different blocks. This is called
wear leveling. A commonly used wear-leveling technique is to balance erasures by moving
data among the blocks, especially when the data are revised [10].
There are two main types of flash memories: NOR flash and NAND flash. A NOR flash
memory allows random access to its cells. A NAND flash partitions every block into multiple
sections called pages, and a page is the unit of a read or write operation. Compared to NOR
flash, NAND flash may be much more restrictive on how its pages can be programmed, such
as allowing a page to be programmed only a few times before erasure [10]. However, NAND
flash enjoys the advantage of higher cell density.
The programming of cells is a noisy process. When charge is injected into a cell, the actual
amount of injection is randomly distributed around the aimed value. An important thing to
avoid during programming is overshooting, because to lower a cell’s level, erasure is needed.
A commonly used approach to avoid overshooting is to program a cell using multiple rounds
of charge injection. In each round, a conservative amount of charge is injected into the cell.
Then the cell level is measured before the next round begins. With this approach, the charge
level can gradually approach the target value and the programming precision is improved.
The corresponding cost is the slowing down in the writing speed.
After cells are programmed, the data are not necessarily error-proof, because the cell levels
can be changed by various errors over time. Some important error sources include write dis-
turb and read disturb (disturbs caused by writing or reading), as well as leakage of charge from
the cells (called the retention problem) [3]. The changes in the cell levels often have an asym-
metric distribution in the up and the down directions, and the errors in different cells can be
correlated.
In summary, flash memory is a storage medium with asymmetric properties. It is easy to
increase a cell’s charge level (which we shall call cell level), but very costly to decrease it due to
block erasure. The NAND flash may have more restrictions on reading and writing compared
to NOR flash. The cell programming uses multiple rounds of charge injection to shift the
cell level monotonically up toward the target value, to avoid overshooting and improve the
precision. The cell levels can change over time due to various disturb mechanisms and the
retention problem, and the errors can be asymmetric or correlated.

2. Codes for Rewriting Data

In this section, we discuss coding schemes for rewriting data in flash memories. The interest
in this problem comes from the fact that if data are stored in the conventional way, even to
change one bit in the data, we may need to lower some cell’s level, which would lead to the
costly block erasure operation. It is interesting to see if there exist codes that allow data to be
rewritten many times without block erasure.
The flash memory model we use in this section is the Write Asymmetric Memory (WAM)
model [16].
Data Representation for Flash Memories 55

Definition 1. W RITE A SYMMETRIC M EMORY (WAM) In a write asymmetric memory, there are
n cells. Every cell has q ≥ 2 levels: levels 0, 1, · · · , q − 1. The level of a cell can only increase, not
decrease.

The Write Asymmetric Memory models the monotonic change of flash memory cells before
the erasure operation. It is a special case of the generalized write-once memory (WOM) model,
which allows the state transitions of cells to be any acyclic directed graph [6, 8, 29].
Let us first look at an inspiring example. The code can write two bits twice in only three
single-level cells. It was proposed by Rivest and Shamir in their celebrated paper that started
the study of WOM codes [29].
We always assume that before data are written into the cells, the cells are at level 0.

Example 2. We store two bits in three single-level cells (i.e., n = 3 and q = 2). The code is shown in
Fig. 1. In the figure, the three numbers in a circle represent the three cell levels, and the two numbers
beside the circle represent the two bits. The arrows represent the transition of the cells. As we can see,
every cell level can only increase.
The code allows us to write the two bits at least twice. For example, if want to write “10”, and later
rewrite them as “01”, we first elevate the cell levels to “0,1,0”, then elevate them to “0,1,1”.

Fig. 1. Code for writing two bits twice in three single-level cells.

In the above example, a rewrite can completely change the data. In practice, often multiple
data variables are stored by an application, and every rewrite changes only one of them. The
joint coding of these data variables are useful for increasing the number of rewrites supported
by coding. The rewriting codes in this setting have been named Floating Codes [16].

Definition 3. F LOATING C ODE

We store k variables of alphabet size  in a write asymmetric memory (WAM) with n cells of q levels.
Every rewrite changes one of the k variables.
Let (c1 , · · · , cn ) ∈ {0, 1, · · · , q − 1}n denote the state of the memory (i.e., the levels of the n cells).
Let (v1 , · · · , vk ) ∈ {0, 1, · · · ,  − 1}k denote the data (i.e., the values of the k variables). For any
56 Data Storage

two memory states (c1 , · · · , cn ) and (c1 , · · · , cn ), we say (c1 , · · · , cn ) ≥ (c1 , · · · , cn ) if ci ≥ ci for
i = 1, · · · , n.
A floating code has a decoding function Fd and an update function Fu . The decoding function

Fd : {0, 1, · · · , q − 1}n → {0, 1, · · · ,  − 1}k

maps a memory state s ∈ {0, 1, · · · , q − 1}n to the stored data Fd (s) ∈ {0, 1, · · · ,  − 1}k . The update
function (which represents a rewrite operation),

Fu : {0, 1, · · · , q − 1}n × {1, 2, · · · , k } × {0, 1, · · · ,  − 1} → {0, 1, · · · , q − 1}n ,

is defined as follows: if the current memory state is s and the rewrite changes the i-th variable to value
j ∈ {0, 1, · · · ,  − 1}, then the rewrite operation will change the memory state to Fu (s, i, j) such that
Fd ( Fu (s, i, j)) is the data with the i-th variable changed to the value j. Naturally, since the memory is
a write asymmetric memory, we require that Fu (s, i, j) ≥ s.
Let t denote the number of rewrites (including the first write) guaranteed by the code. A floating code
that maximizes t is called optimal.

The code in Fig. 1 is in fact a special case of the floating code, where the number of variables
is only one. More specifically, the parameters for it is k = 1,  = 4, n = 3, q = 2 and t = 2.
Let us look at an example of floating codes for two binary variables.

Example 4. We store two binary variables in a Write Asymmetric Memory with n cells of q levels.
Every rewrite changes the value of one variable. The Floating codes for n = 1, 2 and 3 are shown
in Fig. 2. As before, the numbers in a circle represent the memory state, the numbers beside a circle
represent the data, and the arrows represent the transition of the memory state triggered by rewriting.
With every rewrite, the memory state moves up by one layer in the figure. For example, if n = 3, q ≥ 3
and a sequence of rewrites change the data as

(0, 0) → (1, 0) → (1, 1) → (0, 1) → (1, 1) → · · ·

the memory state changes as

(0, 0, 0) → (1, 0, 0) → (1, 0, 1) → (1, 0, 2) → (1, 1, 2) → · · ·

The three codes in the figure all have a periodic structure, where every period contains 2n − 1 layers
(as shown in the figure) and has the same topological structure. From one period to the next, the only
difference is that the data (0, 0) is switched with (1, 0), and the data (1, 1) is switched with (0, 1).
Given the finite value of q, we just need to truncate the graph up to the cell level q − 1.

The floating codes in the above example are generalized in [16] for any value of n and q (but
still with k =  = 2), and are shown to guarantee

q−1
t = (n − 1)(q − 1) +  
2
rewrites. We now prove that this is optimal. First, we show an upper bound to t, the number
of guaranteed rewrites, for floating codes [16].
Data Representation for Flash Memories 57

Fig. 2. Three examples of an optimal floating code for k = 2, l = 2 and arbitrary n, q. (a) n = 1.
(b) n = 2. (c) n = 3.

Theorem 5. For any floating code, if n ≥ k(l − 1) − 1, then

[ k ( l − 1) − 1] · ( q − 1)
t ≤ [ n − k ( l − 1) + 1] · ( q − 1) +  ;
2
if n < k(l − 1) − 1, then
n ( q − 1)
t≤ .
2
Proof. First, consider the case where n ≥ k( − 1) − 1. Let (c1 , c2 , · · · , cn ) denote the memory
k (−1)−1
state. Let WA = ∑i=1 ci , and let WB = ∑in=k(−1) ci . Let’s call a rewrite operation “adver-
sarial” if it either increases WA by at least two or increases WB by at least one. Since there are k
variables and each variable has the alphabet size , a rewrite can change the variable vector in
k( − 1) different ways. However, since WA is the summation of only k( − 1) − 1 cell levels,
there are at most k ( − 1) − 1 ways in which a rewrite can increase WA by one. So there must
be an “adversarial” choice for every rewrite.
Consider a sequence of adversarial rewrites operations supported by a generic floating code.
Suppose that x of those rewrite operations increase WA by at least two, and suppose that
y of them increase WB by at least one. Since the maximum cell level is q − 1, we get x ≤
58 Data Storage

[k(l −1)−1]·(q−1)
 2  and y ≤ [n − k(l − 1) + 1] · (q − 1). So the number of rewrites supported by
[k(l −1)−1]·(q−1)
a floating code is at most x + y ≤ [n − k(l − 1) + 1] · (q − 1) +  2 .
The case where n < k( − 1) − 1 can be analyzed similarly. Call a rewrite operation “adver-
sarial” if it increases ∑in=1 ci by at least two. It can be shown that there is always a adversarial
n ( q −1)
choice for every rewrite, and any floating code can support at most t ≤  2  adversarial
rewrites.
q −1
When k =  = 2, the above theorem gives the bound t ≤ (n − 1)(q − 1) +  2 . It matches the
number of rewrites guaranteed by the floating codes of Example 4 (and their generalization
in [16]). So these codes are optimal.
Let us pause a little to consider the two given examples. In Example 2, two bits can be written
twice into the three single-level cells. The total number of bits written into the memory is
four (considering the whole history of rewriting), which is more than the number of bits the
memory can store at any given time (which is three). In Example 4, every rewrite changes one
of two binary variables and therefore reflects one bit of information. Since the code guarantees
q −1
(n − 1)(q − 1) +  2  rewrites, the total amount of information “recorded” by the memory
q −1
(again, over the history of rewriting) is (n − 1)(q − 1) +  2  ≈ nq bits. In comparison, the
number of bits the memory can store at any give time is only n log2 q.
Why can the total amount of information written into the memory (over the multiple rewrites)
exceed n log2 q, the maximum number of bits the memory can store at any give time? It is be-
cause only the current value of the data needs to be remembered. Another way to understand
it is that we are not using the cells sequentially. As an example, if there are n single level cells
and we increase one of their levels by one, there are n choices, which in fact reflects log2 n
bits of information (instead of one bit).
We now extend floating codes to a more general definition of rewriting codes. First, we use
a directed graph to represent how rewrites may change the stored data. This definition was
proposed in [19].

Definition 6. G ENERALIZED R EWRITING

The stored data is represented by a directed graph D = (VD , ED ). The vertices VD represent all the
values that the data can take. There is a directed edge (u, v) from u ∈ VD to v ∈ VD , v = u, iff a
rewriting operation may change the stored data from value u to value v. The graph D is called the data
graph and the number of its vertices, corresponding to the input-alphabet size, is denoted by L = |VD |.
Without loss of generality (w.l.o.g.), we assume the data graph to be strongly connected.

It is simple to see that when the above notion is applied to floating codes, the alphabet size
L = k , and the data graph D has constant in-degree and out-degree k ( − 1). The out-degree
of D reveals how much change in the data a rewrite operation can cause. It is an important
parameter. In the following, we show a rewriting code for this generalized rewriting model.
The code, called Trajectory Code, was presented in [19].
Let (c1 , · · · , cn ) ∈ {0, 1, q − 1}n denote the memory state. Let VD = {0, 1, · · · , L − 1} denote
the alphabet of the stored data. Let’s present the trajectory code step by step, starting with its
basic building blocks.

Linear Code and Extended Linear Code

We first look at a Linear Code for the case n = L − 1 and q = 2. It was proposed by Rivest and
Shamir in [29].
Data Representation for Flash Memories 59

Definition 7. L INEAR C ODE FOR n = L − 1 AND q = 2

The memory state (c1 , · · · , cn ) represents the data
n
∑ ici mod (n + 1).
i =1

For every rewrite, change as few cells from level 0 to level 1 as possible to get the new data.

We give an example of the Linear Code.

Example 8. Let n = 7, q = 2 and L = 8. The data represented by the memory state (c1 , · · · , c7 ) is
7
∑ ici mod 8.
i =1

If the rewrites change the data as 0 → 3 → 5 → 2 → 4, the memory state can change as
(0, 0, 0, 0, 0, 0, 0) → (0, 0, 1, 0, 0, 0, 0) → (0, 1, 1, 0, 0, 0, 0) → (0, 1, 1, 0, 1, 0, 0) → (0, 1, 1, 1, 1, 1, 0).
The following theorem shows that the number of rewrites enabled by the Linear Code is
asymptotically optimal in n, the number of cells. It was proved in [29].
1
Theorem 9. The Linear Code guarantees at least n+
4 + 1 rewrites.
1
Proof. We show that as long as at least n+ 2 cells are still of level 0, a rewrite will turn at most
two cells from level 0 to level 1. Let x ∈ {0, 1, · · · , n} denote the current data value, and let
y ∈ {0, 1, · · · , n} denote the new data value to be written, where y = x. Let z denote

y−x mod (n + 1).

If the cell cz is of level 0, we can simply change it to level 1. Otherwise, let

S = {i ∈ {1, · · · , n}|ci = 0},

and let
T = {z − s mod (n + 1) | s ∈ S}.
n +1
Since |S| = | T | ≥ and |S ∪ T | < n (zero and z are in neither set), the set S ∩ T must be
2
nonempty. Their overlap indicates a solution to the equation

z = s1 + s2 mod (n + 1)

where s1 and s2 are elements of S.

When n ≥ L and q ≥ 2, we can generalize the Linear Code in the following way [19]. First,
suppose n = L and q ≥ 2. We first use level 0 and level 1 to encode (as the Linear Code
does), and let the memory state represent the data ∑in=1 ici mod n. (Note that rewrites here
will not change cn .) When the code can no longer support rewriting, we increase all cell
levels (including cn ) from 0 to 1, and start using cell levels 1 and 2 to store data in the same
way as above, except that now, the data represented by the memory state (c1 , · · · , cn ) uses
the formula ∑in=1 i (ci − 1) mod n. This process is repeated q − 1 times in total. The general
decoding function is therefore
n
∑ i ( ci − c n ) mod n.
i =1
60 Data Storage

Now we extend the above code to n ≥ L cells. We divide the n cells into b = n/L groups
of size L (some cells may remain unused), and sequentially apply the above code to the first
group of L cells, then to the second group, and so on. We call this code the Extended Linear
Code.
Theorem 10. Let 2 ≤ L ≤ n. The Extended Linear Code guarantees n(q − 1)/8 = Θ(nq) rewrites.
( q −1) n
Proof. The Extended Linear Code essentially consists of (q − 1) nL  ≥ 2L Linear Codes.

Code for Large Alphabet Size L

We now consider the case where L is larger than n. The rewriting code we present√here will
reduce it to the case n = L studied above [19]. We start by assuming that n < L ≤ 2 n .
√
Construction 11. R EWRITING C ODE FOR n < L ≤ 2 n
Let b be the smallest positive integer value that satisfies n/bb ≥ L.
For i = 1, 2, . . . , b, let vi be a symbol from an alphabet of size n/b ≥ L1/b . We may represent any
symbol v ∈ {0, 1, · · · , L − 1} as a vector of symbols (v1 , v2 , . . . , vb ). Partition the n flash cells into b
groups, each with n/b cells (some cells may remain unused). Encoding the symbol v into n cells is
equivalent to the encoding of each vi into the corresponding group of n/b cells. As the alphabet size
of each vi equals the number of cells it is to be encoded into, we can use the Extended Linear Code to
store vi .
√
Theorem 12. Let 16 ≤ n ≤ L ≤ 2 n. The code in Construction 11 guarantees
n(q − 1) log n nq log n
= Θ( )
16 log L log L
rewrites.
Proof. We first show that for b – the smallest positive integer value that satisfies n/bb ≥ L –
it holds that
2 log L
b≤ .
log n
√ √
n
Note that for all 1 ≤ x ≤ 2 , we have n/x  x ≥ n x/2 . Since 16 ≤ n ≤ L ≤ 2 n , it is easy to
verify that √
2 log L n
≤ .
log n 2
Therefore,
n log n 2log
log L log L
  n ≥ n log n = L,
2 log L
which implies the upper bound for b.
Using Construction 11, the number of rewrites possible is bounded by the number of rewrites
possible for each of the b cell groups. By Theorem 10 and the above upper bound for b, this is
at least  
n q−1 n log n q−1 nq log n
 · ≥( − 1) =Θ .
b 8 2 log L 8 log L
Data Representation for Flash Memories 61

For the codes we have shown so far, we have not mentioned any constraint on the data graph
D . Therefore, the above results hold for the case where the data graph is a complete graph.
That is, a rewrite can change the data from any value to any other value. We now show that the
above codes are asymptotically optimal. For the Linear Code and the Extended Linear Code,
this is easy to see, because every rewrite needs to increase some cell level, so the number of
rewrites cannot exceed n(q − 1) = O(nq). The following theorem shows that the rewriting
code in Construction 11 is asymptotically optimal, too [19].

Theorem 13. When n < L − 1 and the data graph D is a complete graph, a rewriting code can
nq log n
guarantee at most O( log L ) rewrites.

Proof. Let us consider some memory state s of the n flash cells, currently storing some value
v ∈ {0, 1, · · · , L − 1}. The next rewrite can change the data into any of the other L − 1 values.
If we allow ourselves r operations of increasing a single cell level of the n flash cells (perhaps,
operating on the same cell more than once), we may reach (n+rr−1) distinct new states. Let r
be the maximum integer sastisfying (n+rr−1) < L − 1. So for the next rewrite, we need at least
r + 1 such operations in the worst case. Since we have a total of n cells with q levels each, the
number of rewrite operations is upper bounded by

n ( q − 1) n ( q − 1) nq log n
≤ log( L−1) = O( ).
r+1  1+log n  + 1 log L

Trajectory Code Construction

Let’s now consider more restricted rewrite operations. In many applications, a rewrite often
changes only a (small) part of the data. So let’s consider the case where a rewrite can change
the data to at most ∆ new values. This is the same as saying that in the data graph D , the
maximum out-degree is ∆. We call such graph D a Bounded Out-degree Data Graph.
The code to be shown is given the name Trajectory Code in [19]. Its idea is to record the path
in D along which the data changes, up to a certain length. When ∆ is small, this approach is
particularly helpful, because recording which outgoing edge a rewrite takes (one of ∆ choices)
is more efficient than recording the new data value (one of L > ∆ choices).
We first outline the construction of the Trajectory Code. Its parameters will be specified soon.
Let n0 , n1 , n2 , . . . , nd be d + 1 positive integers such that ∑id=0 ni = n is the number of cells. We
partition the n cells into d + 1 groups, each with n0 , n1 , . . . , nd cells, respectively. We call them
registers S0 , S1 , . . . , Sd .
The encoding uses the following basic scheme: we start by using register S0 , called the anchor,
to record the value of the initial data v0 ∈ {0, 1, · · · , L − 1}. For the next d rewrite operations
we use a differential scheme: denote by v1 , . . . , vd ∈ {0, 1, · · · , L − 1} the next d values of the
rewritten data. In the i-th rewrite, 1 ≤ i ≤ d, we store in register Si the identity of the edge
(vi−1 , vi ) ∈ ED . (ED and VD are the edge set and vertex set of the data graph D , respectively.)
We do not require a unique label for all edges globally, but rather require that locally, for each
vertex in VD , its out-going edges have unique labels from {1, · · · , ∆}, where ∆ denotes the
maximal out-degree in the data graph D .
62 Data Storage

Intuitively, the first d rewrite operations are achieved by encoding the trajectory taken by the
input data sequence starting with the anchor data. After d such rewrites, we repeat the process
by rewriting the next input from {0, 1, · · · , L − 1} in the anchor S0 , and then continuing with
d edge labels in S1 , · · · , Sd .
Let us assume a sequence of s rewrites have been stored thus far. To decode the last stored
value all we need to know is s mod (d + 1). This is easily achieved by using t/q more cells
(not specified in the previous d + 1 registers), where t is the total number of rewrite operations
we would like to guarantee. For these t/q cells we employ a simple encoding scheme: in
every rewrite operation we arbitrarily choose one of those cells and raise its level by one.
Thus, the total level in these cells equals s.
The decoding process takes the value of the anchor S0 and then follows (s − 1) mod (d + 1)
edges which are read consecutively from S1 , S2 , · · · . Notice that this scheme is appealing in
cases where the maximum out-degree of D is significantly lower than the alphabet size L.
Note that each register Si , for i = 0, . . . , d, can be seen as a smaller rewriting code whose data
graph is a complete graph of either L vertices (for S0 ) or ∆ vertices (for S1 , . . . , Sd ). We use either
the Extended Linear Code or the code of Construction 11 for rewriting in the d + 1 registers.
The parameters √
of the Trajectory Code are shown by the following construction. We assume
that n ≤ L ≤ 2 n .
√
Construction 14. T RAJECTORY C ODE FOR n ≤ L ≤ 2 n
n log n
If ∆ ≤  2 log L , let
d = log L/ log n = Θ(log L/ log n).
n log n
If  2 log L  ≤ ∆ ≤ L, let
d = log L/ log ∆ = Θ(log L/ log ∆).
In both cases, set the size of the d + 1 registers to n0 = n/2 and ni = n/(2d) for i = 1, . . . d.
n log n
If ∆ ≤  2 log L , apply the code of Construction 11 to register S0 , and apply the Extended Linear Code
n log n
to registers S1 , · · · , Sd . If  2 log L  ≤ ∆ ≤ L, apply the code of Construction 11 to the d + 1 registers
S0 , · · · , S d .

The next three results show the asympotically optimality of the Trajectory Code (when ∆ is
small and large, respectively) [19].
n log n
Theorem 15. Let ∆ ≤  2 log L . The Trajectory Code of Construction 14 guarantees Θ(nq) rewrites.

Proof. By Theorems 10 and 12, the number of rewrites possible in S0 is equal (up to constant
factors) to that of Si (i ≥ 1):
     nq 
n0 q log n0 nq log n
Θ =Θ =Θ = Θ ( ni q )
log L log L d

Thus the total number of rewrites guaranted by the Trajectory Code is d + 1 times the bound
for each register Si , which is Θ(nq).
 
n log n nq log n
Theorem 16. Let  2 log L  ≤ ∆ ≤ L. The Trajectory Code of Construction 14 guarantees Θ log ∆
rewrites.
Data Representation for Flash Memories 63

Proof. By Theorem 12, the number of rewrites possible in S0 is:

   
n0 q log n0 nq log n
Θ =Θ
log L log L
Similarly the number of rewrites possible in Si (i ≥ 1) is:
     
ni q log ni nq log n nq log n
Θ =Θ =Θ .
log ∆ d log ∆ log L

Here we use the fact that as d ≤ log L it holds that d = o (n) and log ni = Θ(log n − log d) =
Θ(log n). Notice that the two expressions above are equal. Thus, as in Theorem 15, we con-
clude that the total number of rewrites
 guaranteed
 by the Trajectory Code is d + 1 times the
nq log n
bound for each register Si , which is Θ log ∆ .

The rewriting performance shown in the above theorem matches the bound shown in the
following theorem. We omit its proof, which interested readers can find in [19].
n log n
Theorem 17. Let ∆ >  2 log L . There exist data graphs D of maximum out-degree ∆ such that any
 
nq log n
rewriting code for D can guarantee at most O log ∆ rewrites.

We have so far focused on rewriting codes with asymptotically optimal performance. It is

interesting to study rewriting codes that are strictly optimal, like the floating code in Exam-
ple 4. Some codes of this kind have been studied in [16, 17]. In some practical applications,
more specific forms of rewriting can be defined, and better rewriting codes can be found. An
example is the Buffer Code defined for streaming data [2]. Besides studying the worst-case per-
formance of rewriting codes, the expected rewriting performance is equally interesting. Some
rewriting codes for expected performance are reported in [7, 19].

3. Rank Modulation
We focus our attention now on a new data representation scheme called Rank Modulation [21,
23]. It uses the relative order of the cell levels, instead the absolute values of cell levels, to
represent data. Let us first understand the motivations for the scheme.
Fast and accurate programming schemes for multi-level flash memories are a topic of signifi-
cant research and design efforts. As mentioned before, the flash memory technology does not
support charge removal from individual cells due to block erasure. As a result, an interative
cell programming method is used. To program a cell, a sequence of charge injection opera-
tions are used to shift the cell level cautiously and monotinically toward the target charge level
from below, in order to avoid undesired global erasures in case of overshoots. Consequently,
the attempt to program a cell requires quite a few programming cycles, and it works only up
to a moderate number of levels per cell.
In addition to the need for accurate programming, the move to more levels in cells also ag-
gravates the reliability problem. Compared to single-level cells, the higher storage capacity of
multi-level cells are obtained at the cost of a smaller gap between adjacent cell levels. Many of
the errors in flash memories are asymmetric, meaning that they shift the cell levels more likely
in one direction (up or down) than the other. Examples include the write disturbs, the read
disturbs, and the charge leakage. It will be interesting to design coding schemes that tolerate
asymmetric errors better.
64 Data Storage

The Rank Modulation scheme is therefore proposed in [21, 23], whose aim is to eliminate both
the problem of overshooting while programming cells, and to tolerate asymmetric errors bet-
ter. In this scheme, an ordered set of n cells stores the information in the permutation induced
by the charge levels of the cells. In this way, no discrete levels are needed (i.e., no need for
threshold levels) and only a basic charge-comparing operation (which is easy to implement) is
required to read the permutation. If we further assume that the only programming operation
allowed is raising the charge level of one of the cells above the current highest one (namely,
push-to-top), then the overshoot problem is no longer relevant. Additionally, the technology
may allow in the future the decrease of all the charge levels in a block of cells by a constant
amount smaller than the lowest charge level (block deflation), which would maintain their rela-
tive values, and thus leave the information unchanged. This can eliminate a designated erase
step, by deflating the entire block whenever the memory is not in use.
Let’s look at a simple example of rank modulation. Let [n] denote the set of integers
{1, 2, · · · , n}.

Example 18. We partition the cells into groups of three cells each. Denote the three cells in a group by
cell 1, cell 2, and cell 3. We use a permutation of [3] – [ a1 , a2 , a3 ] – to represent the relative order of the
three cell levels as follows: cell a1 has the highest level, and cell a3 has the lowest level. (The cell levels
considered in this section are real numbers. So no two cells can practically have the same level.)
The three cells in a group can introduce six possible permutations:
[1, 2, 3], [1, 3, 2], [2, 1, 3], [2, 3, 1], [3, 1, 2], [3, 2, 1]. So they can store up to log2 6 bits of infor-
mation. To write a permutation, we program the cells from the lowest level to the highest level. For
example, if the permutation to write is [2, 3, 1], we first program cell 3 to make its level higher than
that of cell 1, then program cell 2 to make its level higher than that of cell 3. This way, there is no risk
of overshooting.

In this section, we use n to denote the number of cells in a group. As in the example, we use
a permutation of [n] – [ a1 , a2 , · · · , an ] – to denote the relative order of the cell levels such that
cell a1 has the highest level and cell an has the lowest level.
Once a new data representation method is defined, tools of coding are required to make it
useful. In this section, we focus on two tools: codes for rewriting, and codes for correcting
errors.

3.1 Rewriting Codes for Rank Modulation

Assume that the only operation we allow for rewriting data is the “push-to-top” operation:
injecting charge into a cell to make its level higher than all the other cell levels in the same cell
group. Note that the push-to-top operation has no risk of overshooting. Then, how to design
good rewriting codes for rank modulation?
Let  denote the alphabet size of the data symbol stored in a group of n cells. Let the alphabet
of the data symbol be [] = {1, 2, · · · , }. Assume that a rewrite can change the data to any
value in []. In general,  might be smaller than n!, so we might end up having permutations
that are not used. On the other hand, we can map several distinct permutations to the same
symbol i ∈ [] in order to reduce the rewrite cost. We use Sn to denote the set of n! permuta-
tions, and let Wn ⊆ Sn denote the set of states (i.e., the set of permutations) that are used to
represent information symbols. As before, for a rewriting code we can define two functions, a
decoding function, Fd , and an update function, Fu .
Data Representation for Flash Memories 65

Definition 19. The decoding function Fd : Wn → [] maps every state s ∈ Wn to a value Fd (s)
in []. Given an “old state” s ∈ Wn and a “new information symbol” i ∈ [], the update function
Fu : Wn × [] → Wn produces a state Fu (s, i ) such that Fd ( Fu (s, i )) = i.

The more push-to-top operations are used for rewriting, the closer the highest cell level is to
the maximum level possible. Once it reaches the maximum level possible, block erasure will
be needed for the next rewrite. So we define the rewriting cost by measuring the number of
push-to-top operations.

Definition 20. Given two states s1 , s2 ∈ Wn , the cost of changing s1 into s2 , denoted α(s1 → s2 ), is
defined as the minimum number of “push-to-the-top” operations needed to change s1 into s2 .

For example, α([1, 2, 3] → [2, 1, 3]) = 1, α([1, 2, 3] → [3, 2, 1]) = 2.

Definition 21. The worst-case rewriting cost of a code is defined as maxs∈Wn ,i∈[] α(s → Fu (s, i )).
A code that minimized this cost is called optimal.

Before we present an optimal rewriting code, let’s first present a lower bound on the worst-
case rewriting cost. Define the transition graph G = (V, E) as a directed graph with V = Sn ,
that is, with n! vertices representing the permutations in Sn . For any u, v ∈ V, there is a
directed edge from u to v iff α(u → v) = 1. G is a regular digraph, because every vertex has
n − 1 incoming edges and n − 1 outgoing edges. The diameter of G is maxu,v∈V α(u → v) =
n − 1.
Given a vertex u ∈ V and an integer r ∈ {0, 1, . . . , n − 1}, define the ball centered at u with
radius r as Brn (u) = {v ∈ V | α(u → v) ≤ r }, and define the sphere centered at u with radius r
as Srn (u) = {v ∈ V | α(u → v) = r }. Clearly,

Brn (u) = Srn (u).
0≤ i ≤r

By a simple relabeling argument, both |Brn (u)| and |Srn (u)| are independent of u, and so will
be denoted by |Brn | and |Srn | respectively.

Lemma 22. For any 0 ≤ r ≤ n − 1,

n!
|Brn | =
(n − r )!

1 r=0
|Srn | = n! n!
( n −r ) !
− ( n −r +1) !
1 ≤ r ≤ n − 1.

Proof. Fix a permutation u ∈ V. Let Pu be the set of permutations having the following prop-
erty: for each permutation v ∈ Pu , the elements appearing in its last n − r positions appear in
the same relative order in u. For example, if n = 5, r = 2, u = [1, 2, 3, 4, 5] and v = [5, 2, 1, 3, 4],
the last 3 elements of v – namely, 1, 3, 4 – have the same relative order in u. It is easy to see
that given u, when the elements occupying the first r positions in v ∈ Pu are chosen, the last
n − r positions become fixed. There are n(n − 1) · · · (n − r + 1) choices for occupying the first
n!
r positions of v ∈ Pu , hence | Pu | = (n− r )!
. We will show that a vertex v is in Brn (u) if and only
if v ∈ Pu .
66 Data Storage

Suppose v ∈ Brn (u). It follows that v can be obtained from u with at most r “push-to-the-top”
operations. Those elements pushed to the top appear in the first r positions of v, so the last
n − r positions of v contain elements which have the same relative order in u, thus, v ∈ Pu .
Now suppose v ∈ Pu . For i ∈ [n], let vi denote the element in the i-th position of v. One can
transform u into v by sequentially pushing vr , vr−1 , . . . , v1 to the top. Hence, v ∈ Brn (u).
n! 
We conclude that |Brn (u)| = | P| = (n− r )!
. Since Brn (u) = 0≤i≤r Sin (u), the second claim
follows.

The following lemma shows a lower bound to the worst-case rewriting cost.

Lemma 23. Fix integers n and , and define ρ(n, ) to be the smallest integer such that |Bρn(n,) | ≥ .
For any code Wn and any state s ∈ Wn , there exists i ∈ [] such that α(s → Fu (s, i )) ≥ ρ(n, ), i.e.,
the worst-case rewriting cost of any code is at least ρ(n, ).

Proof. By the definition of ρ(n, ), |Bρn(n,)−1 | < . Hence, we can choose i ∈ [] \ { Fd (s ) | s ∈
Bρn(n,)−1 (s)}. Clearly, by our choice α(s → Fu (s, i )) ≥ ρ(n, ).

We now present a rewriting code construction. It will be shown that the code achieves the
minimum worst-case rewriting cost. First, let us define the following notation.

Definition 24. A prefix sequence θ = [ a(1) , a(2) , . . . , a(m) ] is a sequence of m ≤ n distinct symbols
from [n]. The prefix set Pn (θ ) ⊆ Sn is defined as all the permutations in Sn which start with the
sequence θ.

We are now in a position to construct the code.

Construction 25. R EWRITING C ODE FOR R ANK M ODULATION

Arbitrarily choose  distinct prefix sequences, θ1 , . . . , θ , each of length ρ(n, ). Let us define Wn =

i ∈[] Pn ( θi ) and map the states of Pn ( θi ) to i, i.e., for each i ∈ [] and s ∈ Pn ( θi ), set Fd ( s ) = i.
Finally, to construct the update function Fu , given s ∈ Wn and some i ∈ [], we do the following: let
(1) (2) (ρ(n,))
[ ai , ai , . . . , ai ] be the first ρ(n, ) elements which appear in all the permutations in Pn (θi ).
(ρ(n,)) (2) (1)
Apply push-to-top operations on the elements ai , . . . , ai , ai in s to get a permutation s ∈
Pn (θi ) for which, clearly, Fd (s ) = i. Set Fu (s, i ) = s .

Theorem 26. The code in Construction 25 is optimal in terms of minimizing the worst-case rewriting
cost.

Proof. It is obvious from the description of Fu that the worst-case rewriting cost of the con-
struction is at most ρ(n, ). By Lemma 23 this is also the best we can hope for.

Example 27. Let n = 3,  = 3. Since |B13 | = 3, it follows that ρ(n, ) = 1. We partition the n! = 6
states into (n−ρn!
(n,))!
= 3 sets, which induce the mapping:

P3 ([1]) = {[1, 2, 3], [1, 3, 2]} → 1,

P3 ([2]) = {[2, 1, 3], [2, 3, 1]} → 2,
P3 ([3]) = {[3, 1, 2], [3, 2, 1]} → 3.

The cost of any rewrite operation is 1.

Data Representation for Flash Memories 67

If a probability distribution is known for the rewritten data, we can also define the perfor-
mance of codes based on the average rewriting cost. This is studied in [21], where an variable-
length prefix-free code is optimized. It is shown that the average rewriting cost of this prefix-
free code is within a small constant approximation ratio of the minimum possible cost of all
codes (prefix-free codes or not) under very mild conditions.

3.2 Error-correcting Codes for Rank Modulation

Error-correcting codes are often essential for the reliability of data. An error-correcting rank
modulation code is a subset of permutations that are far from each other based on some dis-
tance measure. The distance between permutations needs to be defined appropriately accord-
ing to the error model. In this section, we define distance as the number of adjacent transpo-
sitions. We emphasize, however, that this is not the only way to define distance. The results
we present here were shown in [23].
Given a permutation, an adjacent transposition is the local exchange of two adja-
cent elements in the permutation: [ a1 , . . . , ai−1 , ai , ai+1 , ai+2 , . . . , an ] is changed to
[ a1 , . . . , a i −1 , a i +1 , a i , a i +2 , . . . , a n ].
In the rank modulation model, the minimal change to a permutation caused by charge-level
drift is a single adjacent transposition. Let us measure the number of errors by the minimum
number of adjacent transpositions needed to change the permutation from its original value
to its erroneous value. For example, if the errors change the permutation from [2, 1, 3, 4] to
[2, 3, 4, 1], the number of errors is two, because at least two adjacent transpositions are needed
to change one into the other: [2, 1, 3, 4] → [2, 3, 1, 4] → [2, 3, 4, 1].
For two permutations A and B, define their distance, d( A, B), as the minimal number of adja-
cent transpositions needed to change A into B. This distance measure is called the Kendall Tau
Distance in the statistics and machine-learning community [24], and it induces a metric over
Sn . If d( A, B) = 1, A and B are called adjacent. Any two permutations of Sn are at distance at
n ( n −1)
most 2 from each other. Two permutations of maximum distance are a reverse of each
other.

Basic Properties

Theorem 28. Let A = [ a1 , a2 , . . . , an ] and B = [b1 , b2 , . . . , bn ] be two permutations of length

n. Suppose that b p = an for some 1 ≤ p ≤ n. Let A = [ a1 , a2 , . . . , an−1 ] and B =
[b1 , . . . , b p−1 , b p+1 , . . . , bn ]. Then,

d( A, B) = d( A , B ) + n − p.

The above theorem can be proved by induction. It shows a recursive algorithm for computing
the distance between two permutations. Let A = [ a1 , a2 , . . . , an ] and B = [b1 , b2 , . . . , bn ] be two
permutations. For 1 ≤ i ≤ n, let Ai denote [ a1 , a2 , . . . , ai ], let Bi denote the subsequence of B
that contains only those numbers in Ai , and let pi denote the position of ai in Bi . Then, since
d( A1 , B1 ) = 0 and d( Ai , Bi ) = d( Ai−1 , Bi−1 ) + i − pi , for i = 2, 3, . . . , n, we get
n
(n − 1)(n + 2)
d( A, B) = d( An , Bn ) = − ∑ pi .
2 i =2
68 Data Storage

Example 29. Let A = [1, 2, 3, 4] and B = [4, 2, 3, 1]. Then A1 = [1], A2 = [1, 2], A3 = [1, 2, 3],
A4 = [1, 2, 3, 4], B1 = [1], B2 = [2, 1], B3 = [2, 3, 1], B4 = [4, 2, 3, 1]. We get
d( A1 , B1 ) = 0,
d( A2 , B2 ) = d( A1 , B1 ) + 2 − p2 = 0 + 2 − 1 = 1,
d( A3 , B3 ) = d( A2 , B2 ) + 3 − p3 = 1 + 3 − 2 = 2,
d( A4 , B4 ) = d( A3 , B3 ) + 4 − p4 = 2 + 4 − 1 = 5.
(n−1)(n+2) (4−1)(4+2)
Indeed, d( A, B) = 2 − ∑in=2 pi = 2 − (1 + 2 + 1) = 5.
We now define a coordinate system for permutations. We fix A = [1, 2, . . . , n]. For every
permutation B = [b1 , b2 , . . . , bn ], we define its coordinates as XB = (2 − p2 , 3 − p3 , . . . , n − pn ).
Here pi is defined as above for 2 ≤ i ≤ n. Clearly, if XB = ( x1 , x2 , . . . , xn−1 ), then 0 ≤ xi ≤ i
for 1 ≤ i ≤ n − 1.
Example 30. Let A = [1, 2, 3, 4, 5]. Then X A = (0, 0, 0, 0). If B = [3, 4, 2, 1, 5], then XB =
(1, 2, 2, 0). If B = [5, 4, 3, 2, 1], then XB = (1, 2, 3, 4). The full set of coordinates for n = 3 and n = 4
are shown in Fig. 3 (a) and (c), respectively. 
The coordinate system is equivalent to a form of Lehmer code (or Lucas-Lehmer code, inver-
sion table) [26]. It is easy to see that two permutations are identical if and only if they have the
same coordinates, and any vector (y1 , y2 , . . . , yn−1 ), 0 ≤ yi ≤ i for 1 ≤ i ≤ n − 1, is the coor-
dinates of some permutation in Sn . So there is a one-to-one mapping between the coordinates
and the permutations.
n ( n −1)
Let A ∈ Sn be a permutation. For any 0 ≤ r ≤ 2 , the set Br ( A) = { B ∈ Sn | d( A, B) ≤ r }
is a ball of radius r centered at A. A simple relabeling argument suffices to show that the
size of a ball does not depend on the choice of center. We use |Br | to denote |Br ( A)| for any
A ∈ Sn . We are interested in finding the value of |Br |. The following theorem presents a way
to compute the size of a ball using polynomial multiplication.
n ( n −1) x i +1 − 1
Theorem 31. For 0 ≤ r ≤ 2 , let er denote the coefficient of xr in the polynomial ∏in=−11 x −1 .
Then |Br | = ∑ri=0 er .
Proof. Let A = [1, 2, . . . , n]. Let B = [b1 , b2 , . . . , bn ] be a generic permutation. Let XB =
(y1 , y2 , . . . , yn−1 ) be the coordinates of B. By the definition of coordinates, we get d( A, B) =
∑in=−11 yi . The number of permutations at distance r from A equals the number of integer
solutions to ∑in=−11 yi = r such that 0 ≤ yi ≤ i. That is equal to the coefficient of xr in the
x i +1 − 1
polynomial ∏in=−11 ( xi + xi−1 + · · · + 1) = ∏in=−11 x −1 . Thus, there are exactly er permutations
at distance r from A, and |Br | = ∑ri=0 ei .

Theorem 31 induces an upper bound for the sizes of error-correcting rank-modulation codes.
By the sphere-packing principle, for such a code that can correct r errors, its size cannot exceed
n!/|Br |.

Embedding of Permutation Adjacency Graph

Define the adjacency graph of permutations, G = (V, E), as follows. The graph G has |V | = n!
vertices, which represent the n! permutations. Two vertices u, v ∈ V are adjacent if and only if
n ( n −1)
d(u, v) = 1. G is a regular undirected graph with degree n − 1 and diameter 2 . To study
the topology of G, we begin with the follow theorem.
Data Representation for Flash Memories 69

Lemma 32. For two permutations A = [ a1 , a2 , . . . , an ] and B = [b1 , b2 , . . . , bn ], let their coordinates
be X A = ( x1 , x2 , . . . , xn−1 ) and XB = (y1 , y2 , . . . , yn−1 ). If A and B are adjacent, then ∑in=−11 | xi −
yi | = 1.

The above lemma can be proved by induction. Interested readers can see [23] for details.
Let Ln = (VL , EL ) denote a 2 × 3 × · · · × n linear array graph. Ln has n! vertices VL . Each
vertex is assigned integer coordinates ( x1 , x2 , . . . , xn−1 ), where 0 ≤ xi ≤ i for 1 ≤ i ≤ n − 1.
The distance between vertices of Ln is the L1 distance, and two vertices are adjacent (i.e., have
an edge between them) if and only if their distance is one.
We now build a bijective map P : V → VL . Here V is the vertex set of the adjacency graph
of permutations G = (V, E). For any u ∈ V and v ∈ VL , P(u) = v if and only if u, v have
the same coordinates. By Lemma 32, if two permutations are adjacent, their coordinates are
adjacent in Ln . So we get:

Theorem 33. The adjacency graph of permutations is a subgraph of the 2 × 3 × · · · × n linear array.

We show some examples of the embedding in Fig. 3. It can be seen that while each permu-
tation has n − 1 adjacent permutations, a vertex in the array can have a varied degree from
n − 1 to 2n − 3. Some edges of the array do not exist in the adjacency graph of permutations.

Fig. 3. Coordinates of permutations, and embedding the adjacency graph of permutations, G,

in the 2 × 3 × · · · × n array, Ln . In the two arrays, the solid lines are the edges in both G and
Ln , and the dotted lines are the edges only in Ln . (a) Coordinates of permutations for n = 3.
(b) Embedding G in Ln for n = 3. (c) Coordinates of permutations for n = 4. (d) Embedding
G in Ln for n = 4.

The observation that the permutations’ adjacency graph is a subgraph of a linear array shows
an approach to design error-correcting rank-modulation codes based on Lee-metric codes. We
skip the proof of the following theorem due to its simplicity.

Theorem 34. Let C be a Lee-metric error-correcting code of length n − 1, alphabet size no less than n,
and minimum distance d. Let C  be the subset of codewords of C that are contained in the array Ln .
Then C  is an error-correcting rank-modulation code with minimum distance at least d.
70 Data Storage

Single-error-correcting Rank-Modulation Code

We now present a family of rank-modulation codes that can correct one error. The code is
based on the perfect sphere packing in the Lee-metric space [11]. The code construction is as
follows.
Construction 35. (S INGLE - ERROR - CORRECTING RANK - MODULATION CODE )
Let C1 , C2 denote two rank-modulation codes constructed as follows. Let A be a general permutation
whose coordinates are ( x1 , x2 , . . . , xn−1 ). Then A is a codeword in C1 if and only if the following
equation is satisfied:
n −1
∑ ixi ≡ 0 (mod 2n − 1).
i =1
A is a codeword in C2 if and only if the following equation is satisfied:
n −2
∑ ixi + (n − 1) · (− xn−1 ) ≡ 0 (mod 2n − 1).
i =1
Between C1 and C2 , choose the code with more codewords as the final output. 
We analyze the code size of Construction 35.
( n −1) !
Lemma 36. The rank-modulation code built in Construction 35 has a minimum cardinality of 2 .
Proof. Let H = (VH , EH ) be a 2 × 3 × · · · × (n − 1) × (2n − 1) linear array. Every vertex
in H has integer coordinates ( x1 , x2 , . . . , xn−1 ), where 0 ≤ xi ≤ i for 1 ≤ i ≤ n − 2, and
−n + 1 ≤ xn−1 ≤ n − 1.
Given any choice of ( x1 , x2 , . . . , xn−2 ) of the coordinates, we would like to see if there is a
solution to xn−1 (note that −n + 1 ≤ xn−1 ≤ n − 1) that satisfies the following equation:
n −1
∑ ixi ≡ 0 (mod 2n − 1).
i =1
Since ∑in=−11 ixi = (n − 1) xn−1 + ∑in=−12 ixi , and n − 1 and 2n − 1 are co-prime integers, there
is exactly one solution to xn−1 that satisfies the above equation. If xn−1 ≥ 0, clearly
( x1 , x2 , . . . , xn−1 ) are the coordinates of a codeword in the code C1 . If xn−1 ≤ 0, then
∑in=−12 ixi + (n − 1) · [−(− xn−1 )] ≡ 0 (mod 2n − 1), so ( x1 , x2 , . . . , xn−2 , − xn−1 ) are the coor-
dinates of a codeword in the code C2 .
Since 0 ≤ xi ≤ i for 1 ≤ i ≤ n − 2, there are (n − 1)! ways to choose x1 , x2 , . . . , xn−2 . Each
choice generates a codeword that belongs either to C1 or C2 . Therefore, at least one of C1 and
( n −1) !
C2 has cardinality no less than 2 .

Lemma 37. The rank-modulation code built in Construction 35 can correct one error.
Proof. It has been shown in [11] that for an infinite k-dimensional array, vertices whose co-
ordinates ( x1 , x2 , . . . , xk ) satisfy the condition ∑ik=1 ixi ≡ 0 (mod 2k + 1) have a minimum L1
distance of 3. Let k = n − 1. Note that in Construction 35, the codewords of C1 are a subset
of the above vertices, while the codewords in C2 are a subset of the mirrored image of the
above vertices, where the last coordinate xn−1 is mapped to − xn−1 . Since the permutations’
adjacency graph is a subgraph of the array, the minimum distance of C1 and C2 is at least 3.
Hence, the code built in Construction 35 can correct one error.
Data Representation for Flash Memories 71

Theorem 38. The code built in Construction 35 is a single-error-correcting rank-modulation code

whose cardinality is at least half of optimal.

Proof. Every permutation has n − 1 adjacent permutations, so the size of a radius-1 ball, |B1 |,
is n. By the sphere packing bound, a single-error-correcting rank-modulation code can have at
most n!n = ( n − 1) ! codewords. The code in Construction 35 has at least ( n − 1) !/2 codewords.

4. Extended Information Theoretic Results

Flash memory is a type of constrained memory. There has been a history of distinguished
theoretical study on constrained memories. It includes the original work by Kuznetsov and
Tsybakov on coding for defective memories [25]. Further developments on defective mem-
ories include [12, 14]. The write once memory (WOM) [29], write unidirectional memory
(WUM) [28, 30, 32], and write efficient memory [1, 9] are also special instances of constrained
memories. Among them, WOM is the most related to the Write-Asymmetric Memory model
studied in this chapter.
Write once memory (WOM) was defined by Rivest and Shamir in their original work [29]. In
a WOM, a cell’s state can change from 0 to 1 but not from 1 to 0. This model was later general-
ized with more cell states in [6, 8]. The objective of WOM codes is to maximize the number of
times that the stored data can be rewritten. A number of very interesting WOM code construc-
tions have been presented over the years, including the tabular codes, linear codes, etc. in [29],
the linear codes in [6], the codes constructed using projective geometries [27], and the coset
coding in [5]. Fine results on the capacity of WOM have been presented in [8, 13, 29, 33]. Fur-
thermore, error-correcting WOM codes have been studied in [35]. In all the above works, the
rewriting model assumes no constraints on the data, namely, the data graph D is a complete
graph.
With the increasing importance of flash memories, new topics on coding for flash memories
have been studied in recent years. For the efficient rewriting of data, floating codes and buffer
codes were defined and studied in [16] and [2], respectively. A floating code jointly encodes
multiple variables, and every rewrite changes one variable. A buffer code, on the other hand,
records the most recent data of a data stream. More results on floating codes that optimize the
worst-case rewriting performance were presented in [17, 34]. Floating codes that also correct
errors were studied in [15]. In [19], the rewriting problem was generalized based on the data
graph model, and trajectory codes for complete data graphs and bounded-degree data graphs
were presented. The paper [19] also contains a summary and comparison of previous results
on rewriting codes.
Optimizing rewriting codes for expected performance is also an interesting topic. In [7], float-
ing codes of this type were designed based on Gray code constructions. In [19], randomized
WOM codes of robust performance were proposed.
The rank modulation scheme was proposed and studied in [21, 23]. In addition to rewrit-
ing [21] and error correction [23], a family of Gray codes for rank modulation were also pre-
sented [21]. One application of the Gray codes is to map rank modulation to the conventional
multi-level cells. A type of convolutional rank modulation codes, called Bounded Rank Modu-
lation, was studied in [31].
To study the storage capacity of flash memories, it is necessary to understand how accruately
flash cells can be programmed using the iterative and monotonic programming method. This
was studied in [18] based on an abstract programming-error model.
72 Data Storage

The errors in flash cell levels often have an asymmetric property. In [4], error-correcting codes
that correct asymmetric errors of limited magnitude were designed for flash memories.
In a storage system, to avoid the accumulation of errors, a common practice is to write the cor-
rect data back into the storage system once the errors accumulated in the data reach a certain
threshold. This is called memory scrubbing. In flash memories, however, memory scrubbing is
more difficult because to write one correct codeword back into the system, the whole block
needs to be erased. A new type of error-correcting codes, called Error-Scurbbing Codes, were
defined in [20] for multi-level cells. It is shown that even if the only allowed operation is to
increase cell levels, a higher rate of ECC can be still achieved by actively scrubbing errors.
The block erasure property of flash memories affects not only rewriting and cell programming,
but also data movement. In [22], it is shown that by appropriately using coding, the number
of erasures needed for moving data among n NAND flash blocks can be reduced by a factor
of O(log n).
The rewriting codes and the rank modulation scheme are useful not only for flash memories,
but also for other constrained memories. A promonent example is the phase-change memory
(PCM). In a phase-change memory, a memory cell can be switched between a crystalline state
and an amorphous state. Intermediate states are also possible. Since changing the cell to the
crystalline state takes a substantially longer time than changing it to the amorphous state,
the memory is closely related to the Write-Asymmetric Memory model. How to program
PCM cells with more intermediate states fast and reliably is an active ongoing topic. The rank
modulation scheme may provide an effective tool in this area.

Acknowledgment
We would like to thank all our co-authors for their collaborative work in this area. In particu-
lar, we would like to thank Mike Langberg and Moshe Schwartz for many of the main results
discussed in this chapter.

5. References
[1] R. Ahlswede and Z. Zhang, “On multiuser write-efficient memories,” IEEE Trans. on In-
form. Theory, vol. 40, no. 3, pp. 674–686, 1994.
[2] V. Bohossian, A. Jiang and J. Bruck, “Buffer codes for asymmetric multi-level memory,” in
Proc. IEEE International Symposium on Information Theory (ISIT), 2007, pp. 1186-1190.
[3] P. Cappelletti, C. Golla, P. Olivo and E. Zanoni (Ed.), Flash memories, Kluwer Academic
Publishers, 1st Edition, 1999.
[4] Y. Cassuto, M. Schwartz, V. Bohossian and J. Bruck, “Codes for multilevel flash memories:
Correcting asymmetric limited-magnitude errors,” in Proc. IEEE International Symposium
on Information Theory (ISIT), Nice, France, June 2007, pp. 1176-1180.
[5] G. D. Cohen, P. Godlewski, and F. Merkx, “Linear binary code for write-once memories,”
IEEE Trans. on Inform. Theory, vol. IT-32, no. 5, pp. 697–700, Sep. 1986.
[6] A. Fiat and A. Shamir, “Generalized “write-once” memories,” IEEE Trans. on Inform. The-
ory, vol. IT-30, no. 3, pp. 470–480, May 1984.
[7] H. Finucane, Z. Liu and M. Mitzenmacher, “Designing floating codes for expected perfor-
mance,” in Proc. 46th Annual Allerton Conference, 2008.
[8] F. Fu and A. J. Han Vinck, “On the capacity of generalized write-once memory with state
transitions described by an arbitrary directed acyclic graph,” IEEE Trans. on Inform. Theory,
vol. 45, no. 1, pp. 308–313, Jan. 1999.
Data Representation for Flash Memories 73

[9] F. Fu and R. W. Yeung, “On the capacity and error-correcting codes of write-efficient mem-
ories,” IEEE Trans. on Inform. Theory, vol. 46, no. 7, pp. 2299–2314, Nov. 2000.
[10] E. Gal and S. Toledo, “Algorithms and data structures for flash memories,” in ACM Com-
puting Surveys, vol. 37, no. 2, pp. 138-163, June 2005.
[11] S. W. Golomb and L. R. Welch, “Perfect codes in the Lee metric and the packing of poly-
ominoes,” SIAM J. Appl. Math., vol. 18, no. 2, pp. 302–317, Jan. 1970.
[12] A. J. Han Vinck and A. V. Kuznetsov, “On the general defective channel with informed
encoder and capacities of some constrained memories,” IEEE Trans. on Inform. Theory,
vol. 40, no. 6, pp. 1866–1871, 1994.
[13] C. D. Heegard, “On the capacity of permanent memory,” IEEE Trans. on Inform. Theory,
vol. IT-31, no. 1, pp. 34–42, Jan. 1985.
[14] C. D. Heegard and A. A. E. Gamal, “On the capacity of computer memory with defects,”
IEEE Trans. on Inform. Theory, vol. IT-29, no. 5, pp. 731–739, Sep. 1983.
[15] A. Jiang, “On the generalization of error-correcting WOM codes,” in Proc. IEEE Interna-
tional Symposium on Information Theory (ISIT’07), 2007, pp. 1391-1395.
[16] A. Jiang, V. Bohossian and J. Bruck, “Floating codes for joint information storage in write
asymmetric memories,” in Proc. IEEE International Symposium on Information Theory (ISIT),
2007, pp. 1166-1170.
[17] A. Jiang and J. Bruck, “Joint coding for flash memory storage,” in Proc. IEEE International
Symposium on Information Theory (ISIT), 2008, pp. 1741-1745.
[18] A. Jiang and J. Bruck, “On the capacity of flash memories,” in Proc. International Sympo-
sium on Information Theory and Its Applications (ISITA), 2008, pp. 94-99.
[19] A. Jiang, M. Langberg, M. Schwartz and J. Bruck, “Universal rewriting in constrained
memories,” in Proc. IEEE International Symposium on Information Theory (ISIT), Seoul, Ko-
rea, June-July 2009. Technical report online at http://www.paradise.caltech.edu/
papers/etr096.pdf.
[20] A. Jiang, H. Li and Y. Wang, “Error scrubbing codes for flash memories,” in Proc. Canadian
Workshop on Information Theory (CWIT), May 2009, pp. 32-35.
[21] A. Jiang, R. Mateescu, M. Schwartz and J. Bruck, “Rank modulation for flash memories,”
in Proc. IEEE International Symposium on Information Theory (ISIT), 2008, pp. 1731-1735.
[22] A. Jiang, R. Mateescu, E. Yaakobi, J. Bruck, P. Siegel, A. Vardy and J. Wolf, “Storage coding
for wear leveling in flash memories,” in Proc. IEEE International Symposium on Information
Theory (ISIT’09), Seoul, Korea, June-July 2009.
[23] A. Jiang, M. Schwartz and J. Bruck, “Error-correcting codes for rank modulation,” in Proc.
IEEE International Symposium on Information Theory (ISIT), 2008, pp. 1736-1740.
[24] M. Kendall and J. D. Gibbons, Rank correlation methods. Oxford University Press, NY,
1990.
[25] A. V. Kuznetsov and B. S. Tsybakov, “Coding for memories with defective cells,” Problemy
Peredachi Informatsii, vol. 10, no. 2, pp. 52–60, 1974.
[26] D. H. Lehmer, “Teaching combinatorial tricks to a computer,” in Proc. Sympos. Appl. Math.
Combinatorial Analysis, vol. 10, Amer. Math. Soc., Providence, R.I., pp. 179-193, 1960.
[27] F. Merkx, “WOM codes constructed with projective geometries,” Traitment du Signal,
vol. 1, no. 2-2, pp. 227–231, 1984.
[28] W. M. C. J. van Overveld, “The four cases of write unidirectional memory codes over
arbitrary alphabets,” IEEE Trans. on Inform. Theory, vol. 37, no. 3, pp. 872–878, 1991.
[29] R. L. Rivest and A. Shamir, “How to reuse a ‘write-once’ memory,” Information and Con-
trol, vol. 55, pp. 1-19, 1982.
74 Data Storage

[30] G. Simonyi, “On write-unidirectional memory codes,” IEEE Trans. on Inform. Theory,
vol. 35, no. 3, pp. 663–667, May 1989.
[31] Z. Wang, A. Jiang and J. Bruck, “On the capacity of bounded rank modulation for flash
memories,” in Proc. IEEE International Symposium on Information Theory (ISIT), Seoul, Ko-
rea, June-July 2009.
[32] F. M. J. Willems and A. J. Vinck, “Repeated recording for an optical disk,” in Proc. 7th
Symp. Inform. Theory in the Benelux, May 1986, Delft Univ. Press, pp. 49-53.
[33] J. K. Wolf, A. D. Wyner, J. Ziv, and J. Korner, “Coding for a write-once memory,” AT&T
Bell Labs. Tech. J., vol. 63, no. 6, pp. 1089–1112, 1984.
[34] E. Yaakobi, A. Vardy, P. H. Siegel and J. K. Wolf, “Multidimensional flash codes,” in Proc.
46th Annual Allerton Conference, 2008.
[35] G. Zémor and G. Cohen, “Error-correcting WOM-codes,” IEEE Trans. on Inform. Theory,
vol. 37, no. 3, pp. 730–734, May 1991.
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 75

X5

Design and Implementation of FPGA- based

Systolic Array for LZ Data Compression
Mohamed A. Abd El Ghany, Magdy A. El-Moursy* and Aly E. Salama**
Electronics Engineering Dept., German University in Cairo, Cairo, Egypt
Electronics Research Institute, Cairo, Egypt, Mentor Graphics Corporation, Cairo, Egypt*
Electronics and communication Dept., Cairo University, Cairo, Egypt**

1. Introduction
Data compression is becoming an essential component of high speed data communications
and storage. Lossless data compression is the process of encoding ("compressing") a body of
data into a smaller body of data which can, at a later time, be uniquely decoded
("decompressed") back to the original data. In lossy compression, the decompressed data
contains some approximation of the original data.
Hardware implementation of data compression algorithms is receiving increasing attention
due to exponential expansion in network traffic and digital data storage usage. Many lossless
data compression techniques have been proposed in the past and widely used, e.g., Huffman
code (Huffman, 1952) ; (Gallager, 1978);( Park & Prasanna, 1993), arithmetic code (Bodden et
al., 2004);(Said, 2004);(Said, 2003); (Howard & Vetter, 1992), run-length code (Golomb, 1966),
and Lempel–Ziv (LZ) algorithms (Ziv & Lempel, 1977);( Ziv & Lempel, 1978);(Welch,
1984);(Salomon, 2004). Among those, LZ algorithms are the most popular when no prior
knowledge or statistical characteristics of the data being compressed are available. The
principle of the LZ algorithms is to find the longest match between the recently received
string which is stored in the input buffer and the incoming string. Once this match is located,
the incoming string is represented with a position tag and a length variable linking the new
string to the old existing one. Since the repeated data is linked to an older one, more concise
representation is achieved and compression is performed. The latency of the compression
process is defined by the number of clock cycles needed to produce a codeword (matching
results).
To fulfill real-time requirements, several hardware realizations of LZ and its variants have
been presented in the literature. Different hardware architectures, including content
addressable memory (CAM) (Lin & Wu, 2000);(Jones, 1992);(Lee & Yang, 1995), Systolic array
(Ranganathan & Henriques, 1993);(Jung & Burleson, 1998);(Hwang & Wu, 2001), and
embedded processor (Chang et al., 1994), have been proposed in the past. The microprocessor
approach is not attractive for real- time applications, since it does not fully explore hardware
parallelism (Hwang & Wu, 2001). CAM has been considered one of the fastest architectures
to search for a given string in a long world, which is necessary process in LZ. A CAM- based
LZ data compressor can process one input symbol per clock cycle, regardless of the buffer
length and string length. A CAM- based LZ can achieve optimum speed for compression.
76 Data Storage

However, CAMs require highly complex hardware and dissipate high power. The CAM
approach performs string matching through full parallel search, while the systolic-array
approach exploits pipelining.
As compared to CAM- based designs, systolic –array-based designs are slower, but better in
hardware cost and testability (Hwang & Wu, 2001); (Hwang & Wu, 1995); (Hwang & Wu,
1997). Preliminary design for systolic – array contains thousands of processing elements (PEs)
(Ranganathan & Henriques, 1993). High speed designs were then reported later, requiring
only tens of PEs (Jung & Burleson, 1998);(hwang et al., 2001). A technique to enhance the
efficiency of systolic- array approach which is used to implement Lempel- Ziv algorithm is
described in this chapter. A parallel technique for LZ-based data compression is presented.
The technique employs transforming a data–dependent algorithm to a data – independent
algorithm. A control variable is introduced to indicate early completion which improves the
latency. The proposed implementation is area and speed efficient. The effect of the input
buffer length on the compression ratio is analyzed. An FPGA implementation for the
proposed technique is carried out. The implemented design verifies that data can be
compressed and decompressed on-the- fly which opens new areas of research in data
compression.
The organization of this chapter is as follows: In Section 2, the LZ compression algorithm is
explained. The results and comments about some software simulations are discussed. The
dependency graph (DG) to investigate the data dependency of every computation step in the
algorithm is shown. The most recent systolic array architecture is described and an area and
speed efficient architecture is proposed in Section 3. In Section 4, the proposed systolic array
structure is compared with the most recent structures (Hwang et al., 2001) in terms of area,
and latency. An FPGA implementation for the proposed architecture showing the real time
operations is demonstrated in Section 5. Finally, conclusions are provided in Section 6.

2. Lempel-ziv coding algorithm

The LZ algorithm was proposed by Ziv and Lempel in (Ranganathan & Henriques, 1993).
The relationship between n and Ls for optimal compression performance is briefly examined.
The data dependency of every computation step in the LZ compression algorithm is
investigated.

2.1 The compression algorithm:

The LZ algorithm and its variants use a sliding window that moves a long with the cursor.
The window can be divided into two parts, the part before the cursor, called the dictionary,
and the part starting at the cursor, called the look-ahead buffer. The lengths of these two
parts are input parameters to compression algorithm. The basic algorithm is very simple, and
loops executing the following steps:
1. Find the longest match of a string starting at the cursor and completely contained in
the look-ahead buffer to a string starting in the dictionary.
2. Output a triple parameter (Ip, Lmax, S) containing the position Ip of the occurrence in
the window, the length Lmax of the match and the next symbol S past the match.
3. Move the cursor Lmax + 1 symbols forward.
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 77

Let us consider an example with window length of (n=9) and look-ahead buffer length (Ls=3)
shown in Fig. 1.
Let the content of the window be denoted as Xi, i = 0,1,…..,n-1 and that of the look-ahead
buffer be Yj, j = 0,1,……Ls-1 (i.e., Yj = Xi+ n-Ls). According to LZ algorithm, the content of look-
ahead buffer is compared with the dictionary content starting from X0 to Xn-Ls-1 to find the
longest match length. If the best match in the window is found to start from position Ip and
the match length is Lmax. Then Lmax symbols will be represented by a codeword (Ip, Lmax). The
codeword length is Lc:
Lc = 1 + [log2 (n-Ls)] + [Log2 Ls] bits (1)
Lc is fixed. Assume w bits are required to represent a symbol in the window, l = [log2 Ls] bits
are required to represent Lmax, and p = [log2 (n-Ls)] bits are required to represent Ip. Then the
compression ratio is (l + p) / (Lmax * w), where 0 ≤ Lmax ≤ Ls. Hence the compression ratio
depends on the match situation.

Fig. 1. window of the LZ compressor example

The codeword design and the choice of widow length are crucial in achieving maximum
compression. The LZ technique involves the conversion of variable length substrings into
fixed length codewords that represent the pointer and the length of the match. Hence,
selection of values of n and Ls can greatly influence the compression efficiency of the LZ
algorithm.

2.2 Compression Algorithm Paramters Selection

Simulation for the performance of the LZ algorithm for different buffers lengths is performed
using the Calgary corpus and the Silesia Corpus (Deorowicz, 2003) [16], as shown in Fig. 2
and Fig. 3, respectively. In these experiments, the codeword is up to 2 bytes long.
From Fig. 2 and Fig. 3, the compression ratio decreases when n exceeds 1024. The above
improvement in the compression ratio can be obtained only when Ls = 8, 16, or 32. Based on
the results, the best Ls for a good compression ratio is 24. Increasing Ls beyond that will
require a much faster growing n (as well as hardware cost and computation time), with
saturating or even decreasing compression ratio. The reason is that repeating patterns tend to
be short, and that increasing Ls and n also increases the codeword length (l + p). To achieve a
good performance for different data formats, Ls may range from 8 to 32, while n may be from
1k to 8 k. The simulation results verify the results proposed in (Arias et al., 2004).
78 Data Storage

1,2

1 n=256
compression ratio

n=512
0,8
n=102
0,6 4

0,4
4 8 16 32 64 128
Ls values
Fig. 2. The relationship between the compression ratio of Calgary corpus and Ls for different
values of n

1,2

1 n=25
compression ratio

6
n=51
0,8
2
n=10
0,6 24

0,4
4 8 16 32 64 128
Ls values
Fig. 3. The relationship between the compression ratio of Silesia corpus and Ls for different
values of n.

2.3 Dependency graph:

A dependency graph (DG) is a graph that shows the dependence of the computations that
occur in an algorithm. The DG of the LZ algorithm can be obtained as shown in Fig. 4. In the
DG, L (match length) and E (match signal) are propagated from cell to cell. X (content of the
window) and Y (content of the look-ahead buffer) are broadcast horizontally and diagonally
to all cells, respectively. The DG shown in Fig. 4 is called a global DG, because it contains
global signals. The global DG can be transformed into a localized DG, by propagating the
input data Y and X from cell to cell instead of broadcasting them. Processor assignment can
be done by projection of the DG onto the surface normal to the projection vector selected.
After processor assignment, the events are scheduled using a schedule vector.
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 79

Fig. 4. The dependence graph of the LZ compression algorithm.

3. Systolic array architecture

The hardware architectures of LZ data compression demonstrate that systolic array
compressors are better in hardware cost and testability. The main difference between
systolic arrays and other architectures is that systolic arrays can operate at a higher clock
rate (due to nearest-neighbor communication) and can easily be implemented and tested
(due to regularity and homogeneity). In the following subsections, the most recent systolic
array architectures are described. The high performance architecture is proposed.

3.1 Design-1
This architecture was first proposed in (Ziv & Lempel, 1977). The space-time diagram and
its final array architecture are given in Fig. 5, where D represents a unit delay on the signal
line between two processing elements. In Table 1, the six –sets of comparisons have to be
done in sequence in order to find the maximum matching substring.

(1) (2) (3) (4) (5) (6)

X0- Y0 X1- Y0 X2- Y0 X3- Y0 X4- Y0 X5- Y0
X1- Y1 X2- Y1 X3- Y1 X4- Y1 X5- Y1 X6- Y1
X2- Y2 X3- Y2 X4- Y2 X5- Y2 X6- Y2 X7- Y2
Table 1. The six- sets of required comparisons

Let us consider six processing elements (PE’s) in parallel, each performing one vertical set of
comparisons. Each processing element would require 3 time units (Ls =3) to complete its set
of comparisons. As shown in Fig. 5, the delay blocks in each PE delay the Y by two time
steps and the X by one time step. A space- time diagram is used to illustrate the sequence of
comparisons as performed by each PE. The data brought into PE0 are routed systolically
through each processor from left to right. In the first time unit, X0 and Y0 are compared at
PE0. In the second time unit, X1 and Y1 are compared, X0 flows to PE1, and Y0 is delayed by
one cycle (time unit). In the third time unit, X2 andY2 are compared at PE0. At this time, Y0
gets to PE1along with X1, and PE1 performs its first comparison at the third cycle, PE0
80 Data Storage

completes all its required comparisons and stores an integer specifying the number of
successful comparisons in a register called Li. Another register called Lmax, holds the
maximum matching length obtained from the previous PE's. In the fourth time unit, PE0
compares the values of Lmax (which for PE0 is 0) and Li, and the greater of the two is sent to
the Lmax register of the next PE. The result of the Li - Lmax comparison is sent to the next PE
after a delay of one time unit for proper synchronization. Finally, the Lmax value emerging
out from the last PE (PE5 in this case) is the length of the longest matching substring. There
is another register called PE's id, whose contents are passed along with the Lmax value to the
next PE. Its contents indicate the id of the processor element where the Lmax value occurred
which becomes the pointer to the match.

time PE0 PE1 PE2 PE3 PE4 PE5

1 X0-Y0
2 X1-Y1
3 X2-Y2 X1-Y0
4 Lmax X2-Y1
5 X3-Y2 X2-Y0
6 Lmax X3-Y1
7 X4-Y2 X3-Y0
8 Lmax X4-Y1
9 X5-Y2 X4-Y0
10 Lmax X5-Y1
11 X6-Y2 X5-Y0
12 Lmax X6-Y1
13 X7-Y2
14 Lmax
Fig. 5. Design-1 and its space-time diagram indicating the sequence of events in the 6 PE’s.

The functional block of the PE is shown in Fig. 6, in which the control circuit is not included.
Two comparators are needed in the PE: one is for equality check of Yj and Xi and the other
together with two multiplexers are for determining Lmax and Ip. If Yj and Xi are equal, a
counter is incremented each time until an unsuccessful comparison occurs. Sequences Xi and
Yj can be generated by the buffer shown in Fig. 7, which is organized in two levels the upper
level of the buffer holds the incoming symbols to be compressed. The contents of the upper
level are copied into the lower level whenever the "load" line goes high. The lower level is
used to provide data to the PE's in the correct sequence.
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 81

Yj
w
Xi
w
a b
Comparator
a=b

Incremente
r
D Li
Lmax M
u D
x
l a b
Comparator
a>b

Ip
M
p u D
PE’s id x

Fig. 6. The functional block of design-1 PE.

The operation of the buffer is as follows. When the longest match length is found, the same
number of symbols are shifted into the in upper buffer from the source and then the
symbols in the upper buffer are copied to the lower buffer in parallel to generate the next
sequence to the processor array. In the Design- 1 array, The number of clock cycles needed
to produce a codeword is 2 (n-Ls), so the utilization rate of each PE is Ls/ [(2 (n-Ls)], which is
low since the PE is idle from the moment when Li is determined until the time the codeword
is produced. The reason is that it seems impossible to compress subsequent input symbols
before the present compression process is completed, because the number of input symbols
needed to be shifted into the buffer is equal to the longest match length which is not
available before the completion of the present compression process. Therefore, the design
with more than Ls pipeline stages must have some idle PEs before the present codeword is
produced.

Fig. 7. the buffer.

3.2 Design-2:
The Design-2 was first proposed in (Hwang & Wu, 2001). The space-time diagram and its
array architecture are given in Fig. 8. It consists of Ls process elements. The match element
Yjs stay in the PEs, and Xi and Li both flow leftwards with delays of 1 and 2 clock cycles,
82 Data Storage

respectively. The first Li from the leftmost PE will be obtained after 2* Ls clock cycles. After
that, one Li will be obtained every clock cycle. The block diagram of the Design-2 PE is
shown in Fig. 9.

PE2 D D PE1 D D PE0

Li X0 X1 X2 X3 X4 X5 X6 X7
D D
Y2 Y1 Y0

Li PE2 PE1 PE0 Time

L0 1
X0-Y0
L1 2
X1-Y0
L0 L2 3
X1-Y1 X2-Y0
L1 L3 4
X2-Y1 X3-Y0
L0 L2 L4 5
X2-Y2 X3-Y1 X4-Y0
L0 L1 L3 L5 6
X3-Y2 X4-Y1 X5-Y0
L1 L2 L4 7
X4-Y2 X5-Y1
L2 L3 L5 8
X5-Y2 X6-Y1
L3 L4 9
X6-Y2
L4 L5 10
X7-Y2
L5 11
Fig. 8. Design- 2 space- time diagram and array.
Yj
w
Reg
w Xi
D
a b
Comparator
a=b Ei

Accumulator Li
D D
D D

Fig. 9. The structure of Design- 2 PE.

Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 83

3.3 Design-3:
Design-3 was proposed in (Ranganathan & Henriques, 1993). The space-time diagram and
the resulting array are given in Fig. 10. The value of Yi stays in the PE. The buffer element Xi
moves systolically from right to left with delay of 2 clocks. Li propagates from right to left
with 1 clock cycle. The first Li from the leftmost PE will be obtained after Ls clock cycles.
After that, the subsequent ones will be obtained every clock cycle. The structure of Design-3
PE is shown in Fig. 11. MRB is needed to determine Lmax and Ip.

space

Li PE2 PE1 PE0 time

X5-Y0 1 time
L5
X6-Y1 X4-Y0 2
L5 L4
X7-Y2 X5-Y1 X3-Y0 3
L5 L4 L3
L5 X6-Y2 X4-Y1 X2-Y0 4
L4 L3 L2
L4 X5-Y2 X3-Y1 X1-Y0 5
L3 L2 L1
L3 X4-Y2 X2-Y1 X0-Y0 6
L2 L1 L0
L2 X3-Y1 X1-Y1 7
L1 L0
L1 X2-Y2 8
L0
L0 9

Fig. 10. The space-time diagram and the resulting array of Design-3.
Yj
w

Reg
w Xi
D D

a b
comparator
a=b
Ei

Li
D Accumulator
D
Fig. 11. The structure of Design-3 PE.
84 Data Storage

3.4 The Interleaved Design:

From the dependency graph shown in Fig. 4, the interleaved design is obtained by projecting
all the nodes in a particular raw to a single processor element. This design was first proposed
in (Hwang & Wu, 2001). The space–time diagram and the resulting array of the interleaved
design are given in Fig. 12. The Yjs which need to be accessed in parallel do not change
during the encoding step. The special buffer is needed to generate the interleaving symbols of
Xi and Xi + [n- Ls/2] as shown in Fig. 13. The first Li will be obtained after Ls clock cycles from the
leftmost PE and subsequent ones will be obtained every one clock cycle.
Li D D
PE2 PE1 PE0
X7 X4 X6 X3 X5 X2 X4 X1 X3 X0 D D
Y2 Y1 Y0

Li PE2 PE1 PE0 time

X0-Y0 1
L0
X1-Y1 X3-Y0 2
L0 L3
X2-Y2 X4-Y1 X1-Y0 3
L0 L3 L1
L0 X5-Y2 X2-Y1 X4-Y0 4
L3 L1 L4
L3 X3-Y2 X5-Y1 X2-Y0 5
L1 L4 L2
L1 X6-Y2 X3-Y1 X5-Y0 6
L4 L2 L5
L4 X4-Y1 X6-Y1 7
L2 L5
L2 X7-Y2 8
L5
L5 9
Fig. 12. The space-time diagram and the resulting array of interleaved design.

X0 X1 X2 X3 X4 X5 X6 X7 X8
Input sequence
Mux

Fig. 13. buffer design for generating the interleaving symbols.

Before the encoding process, the Yis are preloaded which take Ls extra cycles. During the
encoding process, the time to preload new source symbols depend on how many source
symbols were compressed in the previous compression step, Lmax. The block diagram of the
processor element is shown in Fig. 14.
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 85

Fig. 14. The block diagram of interleaved design PE.

Match results block MRB is needed to determine Lmax among the serially produced Lis. MRB
is shown in Fig. 15. The PEs do not need to store their ids to record the position of the Lis. A
special counter is needed to generate the sequence which interleaves the position of the first
half of Lis and the position for the second half as shown in Fig. 16. The compression time of
the interleaved design is n clock cycles.
Lmax
Reg
Lmax
Li M l
u
x
a b
comparator
a>b

Ip
p Ip
M
p u
counter n - Ls
x

a b Code word
comparator
a=b d

Fig. 15. the match results block

Fig. 16. Special position counter in MRB for interleaved Design.

3.5 Proposed Design (Design- P)

From the dependence graph shown in Fig. 4, all the nodes in a particular row are projected to
a single processor element (Abd El Ghany, 2007). This produces an array of length Ls. The
space- time diagram and the resulting array of Design-P are given in Fig. 17, where D
represents a unit delay on the signal line between two processing elements.
As shown in Fig. 17, the architecture consists of Ls processing elements which are used for the
comparison and L-encoder which is used to produce the matching length. Consequently, the
look-ahead buffer symbols Yjs which do not change during the encoding step, stay in PEs.
86 Data Storage

The dictionary element Xi moves systolically from left to right with a delay of 1 clock cycle.
The match signal Ei of the processing element moves to the L-encoder. The output Li of the
encoder is the matching length resulting from the comparisons at step i-1. The first Li will be
obtained after one clock cycle and the subsequent ones will be obtained every clock cycle.
Before the encoding process, the Yis are preloaded to be processed and this takes Ls extra
cycles. During the encoding process, the time to preload new source symbols depends upon
how many source symbols were compressed in the previous compression step, Lmax.
The functional block of the PE is shown in Fig. 18. Only one equality comparator is needed
for comparing Yj and incoming Xi. The comparator result Ei (match signal) propagates to L-
encoder. The block diagram of L- encoder is shown in Fig. 19. According to Eis (match
signals), L-encoder computes the match length Li corresponding to position i.

Space
Time PE2 PE1 PE0 Li

1 X2-Y0 X1-Y1 X0-Y0

2 X3-Y0 X2-Y1 X1-Y0 L0

3 X4-Y0 X3-Y1 X2-Y0 L1

4 X5-Y0 X4-Y1 X3-Y0 L2

5 X6-Y0 X5-Y1 X4-Y0 L3

6 X7-Y0 X6-Y1 X5-Y0 L4

7 L5

Fig. 17. Architecture of Design-P.

Fig. 18. The functional block of Design-P PE.

Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 87

Fig. 19. L-encode

As shown in Fig. 17, it is clear that the maximum matching length is not produced by the L-
encoder. So, a match results block (MRB) is needed, as shown in Fig. 20, to determine Lmax
among the serially produced Lis. Also, the PEs need not store their ids to record the position
of the Lis (Ip). Since p = [log2(n-Ls)] bits are required to represent Ip, only a p- bit counter is
required to provide the position i associated with each Li, since the time when Li is produced
corresponds to its position. MRB uses a comparator to compare the current input Li and the
present longest match length Lmax stored in the register. If the current input Li is larger than
Lmax, then Li is loaded into the register and the content of the position counter is also loaded
into another register which is used to store the present Ip. Another comparator is used to
determine whether the whole window has been searched. It compares the content of the
position counter with n- Ls, whose output is used as the codeword – ready signal. During the
searching process, Li might be equal to Ls when I< n- Ls, i.e., the content in the look-ahead
buffer can be fully matched to a subset of the dictionary, and hence searching the whole
window is not always necessary. An extra comparator is used to determine whether Lmax is
equal to Ls, and hence the string matching process is completed. Therefore, encoding a new
set of data could start immediately. This will reduce the average compression time. The
number of clock cycles needed to produce a codeword is [(n-Ls) + 1] clock cycles, so the
utilization rate of each PE is (n-Ls) / [(n-Ls) + 1], which almost equals to one. This result is
consistent since the PE is busy once Li is determined until the time at which the codeword is
produced.
88 Data Storage

Lmax
Reg

M Lmax
u
l
x
Li
a b Ls
comparator a b
a>b comparator
a=b
Ip
Reg
done
Ip
M p
p u
counter x
n - Ls
Code word
a b d
comparator
a=b

Fig. 20. The match results block (MRB) for design-P.

Parallel compression can be achieved by using an appropriate number of Design- P modules.

For example, two modules of Design– P can be used, as shown in Fig. 21. The input sequence
of the first module (Xi) is obtained from the first position of Buffer. The input sequence of the
second module (Xi +[(n – Ls)/2 ] ) is obtained starting at ((n –Ls)/2 ) position in the Dictionary.
Note that the MRB now needs to determine Lmax among LI, LII that are produced at the same
time. So, the MRB could be modified. The speed is two times the Design –P array.

Fig. 21. Parallel compression.

Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 89

4. The design comparison

The comparison of four designs is given in Table 2. The Design-1 contains n- Ls processing
elements, which are always active, resulting in a large area and high- power consumption. It
is also slower than the others. Design-P has the maximum utilization rate, minimum latency
time (high compression speed) and minimum area for LZ.
Interleaved Design-P Design-1 Design-2 Design-3
Design
Number of PEs Ls Ls n – Ls Ls Ls
Utilization (n-Ls)/ (n- Ls /2(n-Ls) (n-Ls)/ (n-Ls)/ n
(n-Ls)/ n
Ls+ 1) (n+Ls-1)
Latency n n- Ls+ 1 2 (n –Ls)+1 n+Ls-1 n
DFFs per PE 2w+l+1 2w + 1 3w+l+2p+1 2w+2l+2 3w+l+1
accumulator Per PE 1 1 1
1 ---

Counter per MRB 2 1 --- 2 2

Multiplexer per --- 3 3
3 2
MRB
Multiplexer per --- --- ---
1 ---
buffer
Equality comparator --- 1 1
1 2
per MRB
Table 2. The comparison between Design-P and Design-i.

5. FPGA Implementation
In this section, the proposed architecture for LZ is implemented on FPGA. As shown in Fig.
22, the architecture consists of 3 major components: systolic-array LZ component (SALZC),
block RAM, and host controller. The length of window (n) is assumed to be 1K, and the
length of look-ahead buffer (Ls) is assumed to be 16. SALZ component contains 16 PEs and
implements the most cost effective array architecture (Design –P). Full – custom layout is
straight forward since the array is very regular. Only a single cell (PE) was hand – laid out.
The other 15 PEs are copies of it. Since the array is systolic, routing also is simplified. Block
RAM that is used as the data buffer (Dictionary) is not included in SALZ component.
Thereby, the dictionary length can be increased by directly replacing the block RAM with a
larger one. The dictionary length is a parameter to cover a broad range of applications, from
text compression to lossless image compression.

Fig. 22. LZ compression chip

90 Data Storage

The implementation of the proposed design (Design-P) and interleaved design (Design-i) are
carried out using XILINX (Spartan II XC200) FPGA, for n= 1 k, Ls=16, w =8. The
implementation results are shown in Table 3. The number of components and their
percentage as compared with the available components on the chip are calculated.
The compression rate of a compressor is defined as the number of input bits which can be
compressed in one second. The compression rate (Rc) can be estimated as follows:
Rc = clk × [(LsW) /(n-Ls+1)] (2)

Where clk is the operating frequency. Note that only estimated Rc can be obtained, since it
depends on the input data. It is not possible to predict exactly how many words will be
compressed (Ls at most) and how many clock cycles will be required ((n – Ls +1) at most) for
every compression step. In the proposed implementation, if the window length (n) is 1k, Ls
=16, w =8, and clk = 105 MHz, The Rc is 13 M bit per second.
Number Number Number
Number of Slice of 4 input of Maximum
of Slices Flip Flops LUTs BRAMs Frequency
Design-P 310 13% 408 8% 419 8% 2 14% 105 MHz
Design-i 471 20% 511 10% 650 13% 2 14% 79 MHz
Table 3. The implementation results of Design-P and Design-i

In order to use the parallel scheme to increase the compression rate, the host controller could
be modified and an appropriate number of LZ compressor components could be connected
in parallel, as shown in Fig. 23. Note that, the MRB now needs to determine Lmax among LI, LII
and LIII. e.g., ten components could be implemented in one chip of large size to achieve a
compression rate about 130 M bits per second. Moreover, by modifying the host controller
and including, e.g., dictionaries, the proposed design can be used for other string-matching
based LZ algorithms, such as LZ78 and LZW. The Design-P is flexible.

Fig. 23. Multiple SALZC system.

6. Conclusions
In this chapter, a parallel algorithm for LZ based data compression is described by
transforming a data–dependent algorithm to a data – independent regular algorithm. To
further improve the latency, a control variable to indicate early completion is used. The
Design and Implementation of FPGA- based Systolic Array for LZ Data Compression 91

proposed implementation is area and speed efficient. The compression rate is increased by
more than 40% and the design area is decreased by more than 30%. The design can be
integrated into real – time systems so that data can be compressed and decompressed on –
the - fly.

7. References
Abd El Ghany, M. A.; Salama, A. E. & Khalil, A. H. “Design and Implementation of FPGA-
based Systolic Array for LZ Data Compression” in Proc. IEEE International
Symposium on Circuits and Systems ISCAS07, May 2007
Arias, M.; Cumplido, R. & Feregrino C., (2004) "On the Design and Implementation of an
FPGA-based Lossless Data Compressor.", ReConFig’04, Colima, Mexico, Sep., 2004
Bodden, E.; Clasen, M.; Kneis, J. & Schluter, R., (2004) "Arithmetic coding revealed" pp. 8-57,
May 2004
Chang, J.; Jih, H. J & Liu, J. W., (1994) “A lossless data compression processor,” in Proc. 4th
VLSI Design/CAD Workshop, Aug. 1994, pp.134-137
Deorowicz, S., (2003) “Universal lossless data compression algorithms” Dissertation, Silesian
University, pp 92-119, 2003
Gallager, R. (1978)“Variations on a theme by Huffman,” IEEE Trans. Inform. Theory, vol. IT-
24, pp. 668-674, 1978
Golomb, S., (1966) “Run-length encoding,” IEEE Trans. Inform. Theory, Vol. IT-12, pp. 399-
401, July 1966
Howard, P. G. & Vetter, J. S., (1992) "Practical Implementations of Arithmetic Coding"
Brown University, Department of Computer Science, Technical Report No. (CS-91-45),
April 1992
Huffman, D. A. (1952) “A method for the construction of minimum-redundancy codes" Proc.
IRE, Vol. 40, pp. 1098-1101, September, 1952.
Hwang, S. –A & Wu, C. -W., (2001) “Unified VLSI Systolic Array Design for LZ Data
compression,” IEEE Trans. VLSI Systems, vol. 9, pp. 489- 499, August 2001
Jung, B. & Burleson, W. P., (1998) “Efficient VLSI for Lempel-Ziv compression in wireless
data communication networks,” IEEE Trans. VLSI Syst. Vol. 6, pp. 475-483, Sept.
1998
Lin, K. -J. & Wu, C.-W., (2000) "A low-power CAM design for LZ data compressor," IEEE
Transactions on Computers, Vol.49, No.10, pp. 1139-1145, Oct. 2000
Ranganathan, N. & Henriques, S., (1993) ”High-Speed VLSI Designs for Lempel-Ziv-Based
Data Compression,” IEEE Trans. Circuits and Systems. II: Analog and Digital Signal
Processing, vol. 40, pp.96-106, Feb. 1993
Park, H. & Prasanna, V.K. (1993) “Area efficient VLSI architectures for Huffman coding”,
IEEE Trans. Analog and Digital Signal Processing, Vol.40, pp.568-575,Sep.1993
Said, A., (2003), "Fast Arithmetic Coding," in Lossless Compression Handbook, 2003
Said, A., (2004) " Introduction to Arithmetic coding – Theory and Practice" Imaging Systems
Laboratory, HP Laboratories Palo Alto, April 2004
Salomon, D., (2004) "Data Compression" The complete reference, Computer Science
Department, California State University, Northridge, pp. 22-205, 2004
Sandoval, M. M. & Uribe, C. F. (2005) “A hardware Architecture for Elliptic Curve
Cryptography and Lossless Data Compression,” IEEE Computer Society, May 2005
92 Data Storage

Welch, T., (1984) "A technique for high-performance data compression," IEEE computer, Vol.
17, pp. 8-19, 1984
Ziv, J. & Lempel, A., (1977) “A universal algorithm for sequential data compression,” IEEE
Trans. Inform. Theory, Vol. IT-23, pp. 337-343, 1977
Ziv, J. & Lempel, A., (1978) “Compression of individual sequences via variable rate coding,”
IEEE Trans. Inform. Theory, Vol. IT-24, pp. 530-536, 1978
Design of Simple and High Speed Scheme to Protect Mass Storages 93

X6

Design of Simple and High Speed Scheme to

Protect Mass Storages
Ming-Haw Jing, Yan-Haw Chen, Zih-Heng Chen,
Jian-Hong Chen and Yaotsu Chang
I-Shou University
Taiwan, R.O.C.

1. Introduction
The computer industry has entered a stage of unprecedented improvement in CPU
performance. However, the speed of file system management of huge information is
commonly considered as the main factor that affects the computer performance; for
example, the I/O bandwidth is limited by magnetic disks. The capacity and cost of magnetic
disks per megabyte have been continually improved, but the rotation speed and seek time
are improved very slowly. Recently, many computers have become I/O bound in the
applications of video, audio, commercial database, etc. If such an I/O crisis can be resolved,
the computer system performance will be improved. In 1988, Patterson et al. proposed the
redundant array of independent disks (RAID) system which allows the data to be separated
into several disks (Patterson et al., 1988). We can access the data in parallel so that the
throughput of I/O systems will be improved. On the other hand, more disks in RAID
system have a higher risk of losing data because of high component failure rates. As a result,
the safety and reliability have become the major issues in the RAID system.
When designing a highly available and reliable RAID system, the method of bit wise parity
checking is mostly used to correct errors and to enhance reliability of the RAID system.
However, the parity checking method is limited so that only single disk failure can be
tolerated. In 1995, Blaum et al. proposed a method called even-odd code, which tolerates up
to two disk failures in the RAID system (Blaum et al., 1995). Even-odd code is the first
known scheme for tolerating single or double disk failures, providing an optimal solution
with regard to both storage and performance. However, the major problem concerning the
even-odd code is a variety of modes of operations when solving erasures or up to 2 disk
failures. In practical, it is not easy to be integrated into a VSLI. On the other hand, a small
write problem is difficult to be solved with the even-odd code (Liao & Jing, 2002).
In 1997, Plank proposed a tutorial by using the Reed-Solomon (RS) code to provide error
correction in the RAID system (Plank, 1997). In 2000, Jing et al. also proposed a simple
algorithm, called RS-RAID system, to combine the RS codes with the RAID system (Jing et al.
2000). In this chapter, we aim to improve RS codes codec to design a fast error and erasure
correction for RS-RAID system, and to solve the small write problem in RS codes. In a RS
decoder, there are various algorithms to solve the error locator polynomial, which affect the
94 Data Storage

complexity and performance in hardware design such as the Berlekam-Massey algorithm,

Euclidean algorithm, and Peterson-Gorenstein-Zierler (PGZ) algorithm (Wicker, 1995). The
PGZ algorithm is mostly applied to correct less than six or seven errors because it is very
simple with enhanced error tolerant capability. This chapter contributes to simplify the PGZ
algorithm for single error or double erasure disks correction in the RS-RAID system by
using the support of a set of combinational circuits. Its error and erasure correction become
faster without the use of Chien search to find the root of the error locator polynomial, as
well as the Forney method to find the error magnitudes from evaluator polynomial. Hence,
this straightforward algorithm is then suitable for VLSI design.
This chapter is organized as follows: The history of RAID system is presented in Section 2.
The simple RS encoder and decoder are described in Section 3. The design of small write
modules of RS-RAID system is shown in Section 4. The RS-RAID system is proposed in
Section 5. Finally, the result and conclusion are given in Section 6.

2. The Redundant Array of Independent Disks

A landmark paper, titled “A case for redundant arrays of inexpensive disks (RAID)”, was
presented by Patterson, Katz and Gibson in 1988 (Patterson et al., 1988). RAID systems can
be configured into various ways to get a compromised result on data access speed, system
reliability and size of storage. The general trade-off is to increase data access speed by
writing the data into more places, which increases the amount of storage available by a
factor N. On the other hand, more disks cause lower reliability on the disk system, leading
to data redundancy and the need for additional algorithms to enhance the reliability of
valuable data. There are several levels in the RAID system, such as RAID-0, RAID-1, RAID-
10, RAID-2, RAID-3, RAID-4, RAID-5, RAID-6, and RAID-7. The mostly used versions for
the trade-off are RAID-0, RAID-10, RAID-5, and RAID-6.

2.1 The RAID-0

RAID-0, as shown in Fig. 1, strips all data across multiple drives in a disk array. This is a
high I/O performance solution, since it can simultaneously support many small/individual
and large means of access with all disks transferring in parallel. Thus, very high data
transfer rates (both reads and writes) may be achieved. RAID-0 is very suitable for I/O time
critical or real time applications, such as video on demand (VoD) systems. However, RAID-
0 maximizes the access speed and space available while being low in reliability. Because
RAID-0 provides no data protection, the probability of disk failure increases with increasing
number of disk drives. Any failing single drive will break the entire disk array.

Fig. 1. The hierarchy of RAID-0

Design of Simple and High Speed Scheme to Protect Mass Storages 95

2.2 The RAID-1

RAID-1 writes data to two drives simultaneously in a replicated way, as shown in Fig. 2. If
one drive fails, the data can still be retrieved from the counterpart of the RAID set. This
process is also called “disk mirroring”. Mirroring refers to the 100% duplication of data from
one disk to another so that it is the most expensive RAID (double hardware storage
required). It offers the advantage in reliability only. RAID-1 increases the reliability of
protection of single disk failure, but doubles the cost for the available storage. RAID-1 is
very suitable for both reliability and performance applications, such as OS disk and
accountant data.

Fig. 2. The hierarchy of RAID-1

2.3 The RAID-10

RAID-10 is a combination of RAID-0 and RAID-1, where the data are striped and mirrored
as shown in Fig. 3. This provides a higher speed and reliability, but possesses the same cost
problem as RAID-1. Again, it can only tolerate single disk failure in each disk pair.
RAID 1+0 Host
Command read
write

RAID Controller

Fig. 3. The hierarchy of RAID-10

2.4 The RAID-5

RAID-5 uses a parity encoder to produce parity information and to provide data recovery at
a low cost. The architecture of RAID-5 is shown in Fig. 4. Although one disk is added, data
are striped over all disks so that large files can be fetched with high bandwidth. To balance
the disk access load, a rotating parity is used. Many small random blocks can be accessed in
parallel without hot spots/bottlenecks in any single disk. When a single disk fails, the data
can still be reconstructed from the parity information.
96 Data Storage

Fig. 4. The hierarchy of RAID-5

2.5 The RAID-6

Fig. 5 shows the model of RAID-6. RAID-6 is different from RAID-5 as it has two additional
disks to recover the loss of two disks. This capability provides a higher fault tolerant
capacity for disk array. The popular techniques for RAID-6 are even-odd and RS codes,
which will be discussed later.

Fig. 5. The hierarchy of RAID-6

3. The Design of Reed-Solomon Codes for Error and Erasure Correction

The RS codes have been widely used in error control coding, especially in the applications of
communication, satellite, and storage (Wicker, 1994). The RS codes have maximum distance
separable (MDS) and hence can extend the largest possible minimum distance for codes of
their size and dimension. In addition, RS codes are good at correcting burst errors. In the
following, we discuss the basic definition of the encoder and illustrate those algorithms to
correct one error or two erasure conditions using the PGZ algorithm (Wicker, 1995).

3.1 The Reed-Solomon Codes Encoder

The (n, k) RS codes over GF(2m) have the capability to correct 2t  n  k erasures or t errors,
where n is the total length of a codeword, k is the number of information symbols in GF(2m),
and t is the number of errors that the RS codes can correct. Consider the construction of a t-
error-correcting RS code with a length of 2 m  1 . 2t consecutive powers of  are required as
Design of Simple and High Speed Scheme to Protect Mass Storages 97

zeros of the generator polynomial g(x) for the t-error-correcting RS code. The generator
polynomial is the production of the associated minimal polynomials:
2t
(1)
g( x )  
( x   j ), for   GF( 2 m ) .
j 1

We define each codeword as a polynomial C ( x )  c 0  c 1 x    c n 1 x n1 . Let a message

polynomial can be expressed in terms of I ( x )  i0  i1 x    ik 1 x k 1 . The I(x) can be divided by
a generator polynomial g(x) to obtain the remainder polynomial d(x), such that an encoded
codeword is
C ( x )  I ( x )  x 2 t  d( x ) . (2)

3.2 The Reed-Solomon Codes Decoder

Since the discovery of RS codes, the efficient decoding algorithm has been highly used for
the high-speed data process. Peterson provided the first decoding algorithm for binary BCH
codes (Peterson, 1960). Then Peterson’s algorithm was improved and extended to non-
binary codes by Gorenstein and Zierler (Gorenstein & Zierler, 1961), Chien (Chien, 1964),
and Forney (Forney, 1965). In the following, we discuss how to solve two erasures or one
error using the PGZ algorithm, and propose a fast error and erasure correction algorithm
based on the PGZ algorithm. From the example, further design for tolerating more errors or
erasures may be developed when needed.

3.2.1 The Syndrome Evaluation Algorithm

The syndrome evaluation is the first procedure in the error detection. Assume that a
received codeword polynomial R( x )  r0  r1 x    rn  1 x n 1 can be expressed in terms of the
sum of the codeword polynomial C(x) and the error polynomial e( x )  e0  e1 x    en x n ,
denoted as
R( x )  C ( x )  e( x ) , (3)
and the syndromes are
Sk ( x )  R( k )  C ( k )  e( k ) for 1  k  2t . (4)
The  is the root of C(x), so C ( )  0 ; consequently, the syndrome is actually a form of
k k

evaluation for the error pattern e(x). If there is an error e(x), the syndrome will be nonzero.

3.2.2 The Design of Single Random Error Correction

In the applications of high speed storage, the requirement for single random error or double
erasure correction is commonly applied. The PGZ algorithm is applied to find the error
locator polynomial  ( x )  1   1 x     t x t . We represent the syndromes with t errors in the
received word as follows:
t
(5)
Sk   Y X , for 1  k  2t ,
i 1
i
k
i

where Xi is the error locator for the ith error and Yi is the magnitude of the ith error. If a
random error ei has been introduced into the received word as ri  c i  ei , the syndromes
from equation (4) can be represented as
98 Data Storage

S1  R( )  ei   i (6)
and
S2  R( 2 )  ei   2 i . (7)
From equations (6) and (7), the ei can directly affect the syndromes S1 and S2. With single
error, the error magnitude ei is substituted by Yi and the error location  i is substituted by
Xi. Rearranging the equations (6) and (7), we have
S1  Yi  X i (8)
and
S2  Yi  X i2 . (9)
Finally, the direct solution of the error location Xi and magnitude Yi are obtained from
equations (8) and (9) as
S (10)
X i  2  S2  S11
S1
and
S12 (11)
Yi   S12  S21 .
S2

3.2.3 The Design of Single or Double Erasure Correction

In the definition, the erasure means that the error location has been known. Starting from
the case of double erasures, we assume those magnitudes as Yi and Yj, on the ith and jth
locations in the codeword, respectively, and the effected syndromes are
S1  Yi   i  Yj   j (12)
and
S2  Yi   2 i  Yj   2 j . (13)
By solving equations (12) and (13), the magnitudes can be obtained from the syndromes as
S  j  S2 (14)
Yi  21i
  ( i  j )
and
S1 i  S2 (15)
Yj  .
 2 j  ( i  j )
In the event of one erasure error, Yi  0 and Yj  0 ; the syndromes become S1  Yi   i and
S2  Yi   2 i . We obtain the location or Yi as
S1  S2 (16)
Yi  .
 2i   i
For this event, we can also use the same equation of solving double erasure errors by setting
Yj  0 and j  0 , such that  j  1 ; therefore, equation (16) can be substituted by equation
(14). Thus, only one set of erasure module is needed.
Design of Simple and High Speed Scheme to Protect Mass Storages 99

4. The Design of Reed-Solomon Codes with Small Write Capability

In the design of highly reliable systems for banks, stock markets and hospitals, RAID-6
systems are normally applied. A problem that will affect the system is the frequent
transactions or updating of data/information. Those frequent writing is a small amount of
bytes compared with a block of record in kilo-bytes. This is called the small write problem
which is an important factor that influences the performance of a RAID system. The small
write problem is caused by the frequent modifications of partial codewords, and the parity
bytes of each codeword also need to be updated. In other words, to update the parities,
block data reading is needed for the task of partial data updating in RAID systems.
As shown in Fig. 6, in the RAID-5, assume the D0 is modified, so that the parity should be
updated, too. For example, an inefficient method fetches the unused/whole data and
encodes them again, causing a great delay and a serious problem on writing a small amount
of data. Another smart algorithm is as follows: first, read the old D0 and old parity P and
then perform exclusive-OR operation with the new data D0’ to obtain the new party P’.
Second, write the new data D0’ and new parity P’ back. In this proposed algorithm, we need
2 reads and 2 writes operation to update D0.

Fig. 6. The data updating for single parity

On the other hand, the RAID-6 systems have 2 parities, and the modification of parities
might need to fetch a block of data to calculate the new parity. This is another inefficient
method that fetches more data and performs encoding again. The proposed algorithm has
limited times of access so that only the changed data can be read and the calculation of the
parities is performed, such as the c0 and c1 , which may be different from the original c0 and
c1 in the encoder. Since the advance of VLSI, the proposed IP has absolutely become a
combinational circuit to perform the calculation which will provide a high speed
performance in the case of low delay and a simple interface to the current RAID systems, as
presented in the following sections.

4.1 The Algorithm of Small Write Encoder

Regarding RS-RAID, according to the design of RS codes in Section 3.1, if a symbol/data in
a set of codeword C(x) has been modified, the original parity symbols c0 and c1 have to be re-
encoded. Assuming the new symbol ci  C (x ) , where 2  i  k , is an updated parity in the
codeword C(x). Take ci as the coefficient of the new input, the encoder should generate the
100 Data Storage

correspondence parity symbols c0 and c1 to construct a new set of codeword from the
original parity symbols c0 and c1. The fast algorithm is explained as follows.
First, the codeword C(x) can be expressed in terms of
n1
(17)
C ( )  c 
i 0
i
i
0

and
n1
(18)
C ( 2 )  c 
i 0
i
2i
0.

We also know the c0 and c1 are parity check symbols, thus the equations (17) and (18) can be
expressed in terms of
n1
(19)
c 0 0  c 1 1  c j j  c i i 
i  2 ,i  j

and
n1
(20)
c 0 0  c 1 2  c j 2 j  c 
i  2 ,i  j
i
2i
.

where j is the index or the location, cj is the original coefficient and cj is the new coefficient
of the updated codeword.
Secondly, the new coefficients c0 and c1 can be expressed in terms of the equations
c0  c 0   0 and c1  c 1   1 . In a similar manner, cj  c j   j , where  j stands for the
differences between the original and new coefficients of the jth symbol in the codeword.
Since  j is known, we need to solve  0 and  1 so that we can substitute  0 ,  1 and  j
into equations (19) and (20); therefore, we have
n1
(21)
(c 0   0 ) 0  (c 1   1 ) 1  (c j   j ) j  c 
i  2. i  j
i
i

and
n 1
(22)
(c 0   0 ) 0  (c 1   1 ) 2  (c j   j ) 2 j  c 
i  2. i  j
i
2i
.

To solve the  0 and  1 , subtract equation (19) from equation (21) and use the same way on
equations (20) and (22); we obtain
 0 0   1 1   j j (23)
and
 0 0   1 2   j 2 j . (24)
From the equations (23) and (24), it is found that we do not need the whole codeword to
generate the new set of parity symbols. This is the key to calculate the new parity symbols
on line. To solve  0 and  1 , the set of equations in equations (23) and (24) can be solved
simultaneously, and we have
( j   2 j ) (25)
 1  (cj  c j ) .
(   2 )
Finally, the new parity check coefficient c1 can be expressed in terms of
Design of Simple and High Speed Scheme to Protect Mass Storages 101

( j   2 j ) (26)
c1  (cj  c j )  c1 .
(   2 )
Extending this representation to the new parity check coefficient c0 , we obtain
( j   2 j ) (27)
c0  (cj  c j ) j  (cj  c j )  c0 .
(   2 )
This proves the decoder without a sequential stream of data. If all the elements are over
GF(28), equations (26) and (27) can be rearranged to obtain
c1  (cj  c j )( j   2 j ) 229  c 1 (28)
and
c0  (cj  c j )( j  ( j   2 j ) 230 )  c 0 . (29)
By observing equations (28) and (29), we have a combinational circuit to construct a VLSI
module to finally realize this function.

5. The RS-RAID System Design

Based on the explanations in Sections 3 and 4, the basic modules of RS codes are included to
develop a reliable disk system, or RS-RAID system. The RS-RAID system design not just
tolerates up to two or more disks failure but also corrects error(s) and erasures
transparently. Transparency features high speed and real-time processes without
complicated software control. This section will discuss the design of this RS-RAID system
from its operation to system architecture.

5.1 System Design

In regard to modern mass storage systems, there are usually ten or more disks in the RAID
system with less reliability or a higher risk of data loss. A reliable storage system must
satisfy the following requirements:
1. High-performance disk failure recovery: It not just features high-speed access but also
tolerates up to two or more disks failure.
2. Low recovery time: When one or more disks are not returning data within a limited
period of time, the system control assumes that the disk(s) is/are slow disk(s) and solves its
original information from the existing data/codeword. This strategy must be realized with
low access or recovery time and speed up the system performance.
3. High confidence on individual disk: The disk can identify whether its data are reliable or
not.
Aiming to meet the above requirements, we first use CRC 32 as part of the major checking
data to judge the health of data disks. The rate of miss checking using CRC 32 is
comparatively low. Secondly, a Reed-Solomon Product Code (RS-PC) is proposed with the
support of CRC 32 checking bits to construct a highly reliable RS-RAID. The RS-PC is a
combination of two (n, k) RS codes, denoted by inner-codes Crow and outer-codes Ccol. The
Crow and Ccol codes are presented as the parity symbols of line blocks in row and column
directions, as shown in Fig. 7. The codes Crow and Ccol are combined for double protection,
which is a “check-on-check” of the RS-PC to enhance the error-correcting capability. Since
the two parity symbols are utilized as the line blocks in rows and columns, the architecture
102 Data Storage

of RS-RAID is then partitioned into dual levels, namely the system level and disk level, as
shown in Fig. 8.

Fig. 7. The format of Reed-Solomon product code

Internal Bus
PCI

Fig. 8. The architecture of RS-RAID

In this design, all disks are considered as large logical/unified storage. The host can access
the data in RS-RAID through the IDE or SCSI interface. At the system level, there are n disks,
and all data are encoded and decoded by L1 (level one) RS-code codec through the PC
interface. This design guarantees the reliability of data reading from the large logical disk.
System Cache memory, which temporarily stores the data encoded from L1 RS-code codec,
is used to buffer the currently used data. Cache buffer will improve the RS-RAID
performance in frequent access to/from the system.
At the disk level, each disk has n stripes space. When the data are read from a disk, they are
decoded by L2 (level two) RS-code codec and then checked by CRC 32 codec. If the amount
Design of Simple and High Speed Scheme to Protect Mass Storages 103

of errors is too many for the correcting capability of L2 RS-code codec, this situation will be
detected by CRC 32 codec. This guarantees the data reading from individual disk in a
reliable condition. This system has advantages such as high capacity, throughput and
reliability, because all encoding/decoding processes are operating in real time.

5.2 System Operation

For easier explanations, the operation of the system is based on the concept of the error
correction design. The operations in dual levels can enhance the system reliability and
increase the number of errors tolerated in the system.

5.2.1 The System Level

At the system level, the L1 RS-code codec can be partitioned into the encoder and the
decoder parts, as shown in Fig. 9. When bulk write from the host is performed, the
information of stripe u can be expressed in terms of I u ( x )  {ik 1 , ik 2 , , i0 } and encoded into a
inner-code C cow ( x )  {c u ,n1 , c u ,n  2 , , c u , 0 } for each tripe of data, where 0  u  n , n is the
u

number of total disks in system, and k is the number of data disks. When the frequently
rewritten data are sent to the system cache sequentially from the host, the c0 and c1 become
new parity symbols and must be updated in real time, as shown in Fig. 9.

Iu(x), c'j c'0, c'1

Encoder
c0, c1 Disk0

Decoder Idisk(x)
System Cache Disk1
Syndromes Xi and Write Back
S1, S2 Yj, Yi Microprocessor

Erasure and RStripe(x) Rdisk(x)

Error
Corrector
Iu(x) Diskn-1

System Level Disk Level

Fig. 9. The RS codes codec block diagram of system level

Before being written back to disks, all data are temporarily stored into system cache. During
write back, the data are firstly encoded into the outer-code C col ( x )  {c n1 , v , c n2 , v , , c 0 , v } for
v

each disk, where 0  v  n , and n is the stripes of a disk.

When read from an unreliable disk or communication channel, the received data from disk
Rdisk ( x )  { Rn 1 , Rn 2 , , R0 } are decoded into I disk ( x )  { I n 1 , I n 2 , , I 0 } and stored as inner-code
Crow in system cache. When read from cache, the inner-code Crow can be represented in terms
104 Data Storage

of stripe Rstripe ( x )  {ru ,n1 , ru ,n2 , , ru , 0 } , where 0  u  n , and n is the total number of disks.
u

Decoding Rstripe ( x ) , from the previous research in (Jing et al. 2001), the procedure is as
follows:
A. Rstripe ( x ) is firstly sent to the error corrector and generates its syndromes S1 and S2 for
u

error checking and correction purposes.

B. When a random error occurs, the corrector will use the syndromes to calculate its
magnitude Yi on position Xi.
C. With erasure(s), the corrector firstly sets one or both of Xi and Xj as the already known
position(s) of erasure(s) to solve their correlated error magnitudes Yi and Yj.
When an error or erasure(s) is found, the corrector will correct the Rstripe ( x ) using S1 and S2
u

immediately after completion of reading. The timing of this procedure is shown in Fig. 10.
With such high-speed correction, we consider this operation as a real-time correction
process.

Fig. 10. The RS codes codec block diagram of system level

5.2.2 The Disk Level

Internal Bus

Fig. 11. The RS codes codec block diagram of system level

At the disk level as shown in Fig. 11, when the bulks are written from cache into disks, the
encoding procedure at the disk level is as follows:
1. The data I disk ( x )  { I n 1 , I n 2 , , I 0 } in cache is written into disks.
Design of Simple and High Speed Scheme to Protect Mass Storages 105

2. Each disk receives its own data I which is then encoded by CRC 32 codec.
3. The encoded data stream from CRC 32 codec are encoded again by L2 RS-codes codec and
stored in the disks.
Thus, the RS-PC encoding is finished.
When a system reads data from disks into cache, all data will be checked by CRC 32 codec
and decoded by L2 RS-codes codec. The decoding procedure at the disk level is as follows:
1. The data in each disk are decoded and corrected by its own L2 RS-codes codec.
2. The decoded data stream from each L2 RS-codes codec are decoded again by its own CRC
32 codec.
Finally, if the quantity of errors is greater than the error-correcting capacity of L2 RS-codes
codec, the CRC 32 codec can detect the errors and report to the system level, so this disk is
marked as an erasure at the system level. This strategy will enhance the system reliability
and increase the data access speed with less possibility to retry failure disk(s). This design
provides support of double protection for the RAID system in real time.

6. Conclusion
This paper provides an example of coding to implement a RS code in redundant array of
independent disks system in correcting single random error and double erasures. There are
new directions such as the small write module and higher correction capabilities for the
design of a RS-RAID system. As a result, the proposed RS-RAID system has the following
advantages:
1. Expandable design: As the design of RAID-6, this paper does not only propose a solution
for dual disks failure, but also adopt the PGZ algorithm to correct less than six or seven
errors.
2. Integrated concept: This system presents a unified RSPC concept to partite the system into
dual levels of abstract/structure. Thus, the modules at the disk level mainly deal with burst
or random errors in disks, and the control of the system level does the correction for
multiple failures of the system. On the other hand, each disk may be a surface of disk so that
a fault tolerant hard disk is produced.
3. Real-time updating capability: In regard to the applications for banks, stock markets,
hospitals or military purposes, the system requires frequent transactions or updating of
data/information. The small write module may support the system cache with a real time
requirement and solve the frequent update operations in the RAID system with very low
overhead.
4. Suitability for co-design: The proposed algorithm is suitable for both hardware and
software designs of the modules in the applications of RS codes by using the finite field. The
math of finite field belongs to modern algebra which has been largely applied to the
applications of error correction code and cryptography. This suggests that the hardware
modules will be integrated into the math processor in CPU of the future versions. The
reliable control may be easily integrated into microcontrollers and general processors.
5. More applications: With the advantages from the expandability to the co-design of the
system, this concept may extend its applications to most memory systems such as the flash
memory, DRAM, and so on.
106 Data Storage

7. Acknowledgments
This work is supported in part by the Nation Science Council, Taiwan, under grant NSC 91-
2213-E-214-010.

8. References
Blaum, M.; Brady, J. ; Bruck, J. & Menon, J. (1995). EVENODD: An Efficient Scheme for
Tolerating Double Disk Failures in RAID Architectures. IEEE Transaction on
Computers, Vol. 44, No. 2, pp. 192-202, ISSN: 0018-9340
Chien, R.T. (1964). Cyclic Decoding Procedure for the Bose-Chaudhuri- Hocquenghem
Codes. IEEE Transaction on Information Theory, Vol. IT-10, No. 10, pp. 357-363, ISSN:
0018-9448
Forney, G.D. (1965). On Decoding BCH Codes. IEEE Transaction on Information Theory, Vol.
IT-11, No. 4, pp. 549-557, ISSN: 0018-9448
Gorenstein, D. & Zierler, N. (1961). A Class of Error Correcting Codes in pm Symbols. Journal
of the Society of Industrial and Applied Mathematics, Vol. 9, No. 2, pp. 207-214, ISSN:
0368-4245
Jing, M.H.; Chen, Y.H. & Yuan, K.Y. (2000). The Comparison of Evenodd Code and RS Code
for RAID Applications, Asia Pacific Conference on Multimedia Technology and
Application, pp. 261-268, December, 2000, Kaohsiung, Taiwan
Jing, M.H.; Chen, Y.H. & Liao, J.E. (2001). A Fast Error and Erasure Correction Algorithm for
a Simple RS-RAID, Proceedings of the 2001 International Conferences on Info-tech and
Info-net, pp. 333-338, ISBN: 0-7803-7010-4, October, 2001, Beijing, China
Liao, J.E & Jing, M.H. (2002). The Research and Implementation of High-Speed RAID Using Reed-
Solomon Codes, Master's thesis, Department of Information Engineer, I-Shou
University, Kaohsiung, Taiwan
Patterson, D.A.; Gibson, G. & Katz, R.H. (1988). A Case for Redundant Arrays of
Inexpensive Disks (RAID), Proceedings of the 1988 ACM SIGMOD international
conference on Management of data, pp. 109-116, ISBN: 0-89791-268-3, June, 1988,
ACM, Chicago, Illinois
Peterson, W.W. (1960). Encoding and Error-Correction Procedures for the Bose-Chaudhuri
Codes. IRE Transaction on Information Theory, Vol. IT-6, No. 4, pp. 459-470, ISSN:
0096-1000
Plank, J.S. (1997). A Tutorial on Reed-Solomon Coding for Fault-Tolerance in RAID-like
Systems. Software - Practice and Experience, Vol. 27, No. 9, pp. 995-1012, ISSN: 0038-
0644
Wicker, S.B. & Bhargava, V.K. (1994). Reed-Solomon Codes and Their Applications, IEEE Press,
ISBN: 0-7803-5391-9, NJ
Wicker, S.B. (1995). Error Control Systems for Digital Communication and Storage, Prentice Hall,
ISBN: 0-13-308941-X, US
Administration and Monitoring Service for Storage Virtualization in Grid Environments 107

X7

Administration and Monitoring Service for

Storage Virtualization in Grid Environments
Salam Traboulsi
Institut de recherche en informatique de Toulouse IRIT
Université Paul Sabatier Toulouse III
31062 Toulouse Cedex 3, France
[email protected]

1. Introduction
A grid is collection of computers and storage resources maintained to serve the needs of
some community (Foster et al., 2002). It addresses collaboration, data sharing, cycle sharing
and other patterns of interaction that involve distributed resources.
Actually, the de facto building block is for high performance storage systems. In grid
context, the scale and the reliability have key issues as many independently failing and
unreliable components need to be continuously accounted for and managed over time
(Porter & Katz, 2006). Manageability also becomes of paramount importance, since
nowadays the grid commonly consists of hundred or even thousands of storage and
computing nodes (Foster et al., 2002). One of the key challenges faced by high performance
storage system is scalable administration and monitoring of system state.
A monitoring system captures subset of interactions amongst the myriad of computational
nodes, links, and storage devices. These interactions are interpreted in order to improve
performance in grid environment.
On a grid scale basis, ViSaGe aims at providing to the grid users a transparent, reliable and
powerful storage system underpinned by a storage virtualization layer. ViSaGe is based on
three services, namely as: administration and monitoring service, storage virtualization
service and distributed file system. The virtualization service incorporates storage resources
distributed on the grid in virtual spaces. These virtual spaces will be attributed by the
distributed file system various qualities of service and data placement policies.
In this paper, we present our scalable distributed system: Admon. Admon consists of an
administration module that manages virtual storage resources according to their workloads
based on the information collected by a monitoring module. It is known that the
performance of a system depends deeply on the characteristics of its workload. Usually, the
node's workload is associated to the service response time of a storage application. As the
workload increases, the service response time becomes longer. However, the utilization
percentage of system resources (CPU load, the Disk load and Network load) must be taken
into more consideration. Therefore, Admon traces ViSaGe's applications and collects system
resources (CPU, Disk, Network…) percentage of utilization. It is characterized by an
108 Data Storage

automatic instrumentation. It is an auto-manager monitoring system. In this paper, we

demonstrate Admon efficiency due to nodes workload and user constraint like CPU usage
workload.
The remainder of this paper is composed as follows: Section 2 presents the existing
monitoring systems and motivates the need of Admon. Section 3 illustrates ViSaGe and grid
environment, followed by Admon architecture, in Section 4, where we delve into details of
Admon functionalities and API. In Section 5, we show a performance experiment tests for an
Admon evaluation in a storage virtualization system. Finally, we conclude the paper in
Section 6 and some future work is also presented.

2. Related work
In a distributed system such as ViSaGe, the storage management is principally funded on
automatic decisions to improve performance. These decisions identify nodes which can be
accessed efficiently. Therefore, analyzing node’s workload is primary to make an adequate
decision. Moreover, this workload is mainly supported by the monitoring systems. Several
existing monitoring systems are available for monitoring grid resources (computing
resources, storage resources, networks) and grid applications. Examples of existing
monitoring systems are: Network Weather Service (NWS) (Wolski et al., 1999), Relational
grid Monitoring Architecture (R-GMA) (Cooke et al., 2004), NetLogger (Gunter et al., 2000),
etc.
Network Weather Service (NWS) is a popular distributed system for producing short term
performance forecasts based on historical performance measurements. NWS is specific to
monitor grid computing and network resources. It provides CPU percentage availability
rather than CPU usage of a particular application. It supposes that nodes are available for
the entire application. This conclusion is very limited to be used to achieve high throughput
in a storage virtualization system like ViSaGe.
Relational grid Monitoring Architecture (R-GMA) is based on a relational model. It is an
implementation of GGF Grid Monitoring Architecture (GMA). It is being used both for
information about the grid and for application monitoring. R-GMA collected information
was used only to find out about what services are available at any time.
Netlogger is a monitoring and a performance analysis tool for grids. It provides tracing
services for distributed applications to detect bottlenecks. However, performance analysis is
carried out in post mortem. In ViSaGe, we don't need a system for studying the application's
state, but a system that analyzes the availability of grid resources.
The aforementioned monitoring systems are designed to produce and deliver measurements
on grid resources and application efficiently. Nevertheless, none of these systems presents
the whole of the pertinent characteristics for a virtualization system like ViSaGe. ViSaGe
needs a monitoring system providing a full view of node's workload in order to choose
better nodes according to its target (replicating data, distributing workload …) and during
nodes workload variations. Therefore, our monitoring system, Admon, traces applications,
and collects information about storage resources, grid resources and networks. It is
considered as an intersection point between all the aforementioned monitoring systems. It
should use its monitoring knowledge for choosing the accurate node. The choice will be
according to a prediction model in order to place data, efficiently, during runtime execution.
Administration and Monitoring Service for Storage Virtualization in Grid Environments 109

3. ViSaGe Architecture
A grid consists of three levels: node level represents storage node and computing node, site
level represents site’s administrator of the grid, and grid’s level represents grid’s
administrator. The components of our storage virtualization system are distributed on this
architecture (Fig.1) which brought an additional service for improving storage quality, and
the access to the data.
 At the node level, we have :
1. The virtualization service (Vrt).
2. The distributed file system service (Visagefs).
3. The administration and monitoring service (Admon).
 At the upper levels (the site level and the grid level), we found the tools of the
administration and monitoring service (Admon).
ViSaGe aggregates distributed storage resources. It proposes for providing a self adaptive
storage management system. The ViSaGe storage virtualization layer allows simplifying the
management of sharing storage resources: flexible and transparent access to the data, high
performance (data replication, links configuration, etc).

Fig. 1. ViSaGe’s Architecture

The traditional application (like creation file, reading/writing data) uses the virtual storage
resource through our file system, and thus reached data by means of virtualization layer.
Furthermore, to handle the dynamic and the inherent nature of the grid, and to control the
virtualization (storage resources attribution, virtual storage resources creation/destruction
or modification), our storage virtualization system proposes the administration and the
monitoring layer, Admon. Therefore, Admon’s design must be strongly related to grid
environment. Next, we present the hierarchical architecture and the fundamental concept of
this system.
110 Data Storage

4. Administration and Monitoring Service

The Admon's components (Fig.2) are distributed throughout the grid. These components
are:
1. An intelligent administration and monitoring agent, distributed among the storage
and computing resources.
2. The administration and monitoring tools (server types), installed on site and grid
controllers.
These components rely on a send/request protocol to monitor and manage the state of grid
nodes (Foster et al., 2002).

Fig. 2. Admon's Architecture

4.1 The administration components

The administration service manages the grid resources. It consists of (Fig.2): administration
agent (AA_event), site’s administration tool (SA_Event) and grid’s administration tool
(GA_event).
a) The administration agent:
In order to facilitate the deployment, the configuration and the update of the resources, an
administration agent runs at the storage and computing elements. It constitutes a means of
communication with the site's administration tool, and the other layers’ interfaces of our
storage virtualization system. It receives or sends events for instance when a node asks to be
integrated into the site, its administration agent contacts the administration tool of its site to
send its identity.
b) The site’s administration tool:
It is a centralized tool for the site management, able to send to the grid’s administration tool
the information concerning the site that it manages. For example:
1. The storage elements configuration within the site, to allow the resource sharing.
2. The management of the available physical resources to be shared.
3. Sending and receiving messages from the grid’s administration tool.
Administration and Monitoring Service for Storage Virtualization in Grid Environments 111

c) The grid’s administration tool:

At the grid’s level, the Admon’s administration part is represented by the administration
tool which runs in front of the grid owner. It manages grid user requests, and the
information concerning the grid components in order to make decisions according to the
nodes state.

4.2 The monitoring components

The Monitoring service collects information and sends notifications to the administrator. It
consists of (Fig.2): monitoring agent (MA_sensor and MA_event), site’s monitoring tool
(SM_event) and grid’s monitoring tool (GM_event).
a) The monitoring agent:
The intelligent monitoring agent is launched on all single machines where the processes of
the application are executed. Distributed among the computing and storage elements, it
allows collecting relevant information by the mean of MA_sensor and observing the
behaviour of its node by the mean of MA_event. This information is represented by: CPU
usage percent, Memory usage percent, Input/Output throughput to storage devices,
Network throughput and event log.
The monitoring information is collected by the system command as “top”. We usually poll
at 100 ms intervals. Since the kernel events are not time-stamped, the data obtained this way
represents all events in this interval. We use a simple statistical method to send the collected
information when a significant change occurred. In order to no overload the grid, our agent
monitoring is autonomous. It changes its interval of polling.
Event log analysis: for studying the node stability it is very important to collect the times-
tamping of interest events to the logical volume of this node. This information will be sent to
Admon grid's tool, by the mean of site's tool, in order to analyze the storage virtual state.
The adopted method for collecting relevant information is no disrupted of the running
application, which means that our system is a non intrusive system.
b) The site's monitoring tool:
It is introduced as an intermediate monitoring process between the grid's monitoring tool
and the monitoring agent. These monitoring tools are used to forward data from the
monitoring agent to the grid's monitoring tool. The introspection mechanism, adopted by
the monitoring agent to the node level, leverages the collected and optimized information of
each observed node. Thus, it will have a global vision concerning the state of the
components of the corresponding site.
c) The grid's monitoring tool:
The grid's monitoring tool incorporates the synthesized information by the monitoring tools
of each site into the grid information system. This information describes the state of each site
in the grid and represents the significant behaviours of the various components.
The synthesized information consists of:
 Site's state: overloaded or failing sites.
 Network's state: overloaded or broken links.
This collected information makes it possible to notify the grid's administration tool in order
to make appropriate management decisions, and helps to have a global vision concerning
the state of virtual storage resources created.
112 Data Storage

The collection of traces and logs creates a complete view of what the system is being ordered
to be achieved. The next step should allow combining the system’s knowledge with the
statistically gathered data, for evaluating the stability state of the virtual storage resources.

5. Experimental results
In a data storage distributed system, the major concern is the data access performance.
Many works focus on the data access performance has been concentrated on a static value
and is not related to nodes load. However, whole system’s performance is affected not only
by the behavior of each single application, but also by the resulting execution of several
different applications combined together. We have implemented Admon to study the
service response time of ViSaGe’s application influencing the system resources utilization. It
traces applications and collects system resources' percentages (CPU, Disk, Network…) of
utilization. The contribution of our work is to uncover which nodes are dedicated or not to
ViSaGe. The Admon's performance metrics are represented by the CPU, Disk and Network.
Admon sets the maximum value of performance metrics; constraint for achieving an
experiment and making an adequate decision. This solution is used by Admon to identify
grid dedicated node.
Here we present a use case study in order to show the Admon automatic instrumentation
due to its decision. This decision improves data storage management and performance in
our data storage system. Before on going on our experiment, we choose the CPU load (CPU
< 70) like a constraint. This constraint is like a user demand to improve experiment
performance.

5.1 Experimental setup

In our experimental, we choose the meshing AMIBE 3D. The European consortium EADS-
CCR makes use of AMIBE drive electromagnetic simulations. We used grid platform, which
gathered 4 sites distributed in France (IRIT, SEANODES, EADS-CCR, and CS-SI). For our
preliminary experiments, we used 4 nodes (node1, node2, node3 and node4) distributed in
these sites, where nodes are connected through gigabit Ethernet network.
The meshing full process encompasses three phases, each one being data dependent from its
elder. First step is the 1D meshing which is started with a structural description. This job
will run on a single node and generates one directory filled with one file per face to mesh.
After completion, ViSaGe administration tools replicate data from current logical volume to
all of the involved sites with at least one replica per site. Having 1D stage data replicated,
we can launch the 2D stage on the reserved ViSaGe nodes for our application. We use a job
scheduler (torque) to dispatch AMIBE 2D jobs on a face per face basis. On each run, a new
directory is filled with an intermediate files. The final AMIBE 3D job is to collect and
compound all those intermediate 2D files in one final set of 3D files. Thereby, although 2D
data sets now span over a wide range of local storage resources. In AMIBE, before
launching, an edge length must be determined. We used an AMIBE executable version
where this value is set to 50. Modifying this value will modify performance result. However,
in our experiment, we fix this value to 3. The reason from that is to show the administration
and monitoring efficiency and utility by using AMIBE.
Administration and Monitoring Service for Storage Virtualization in Grid Environments 113

Fig. 3. nodes dedicated to ViSaGe

5.2 Results and discussion

We installed on node1 the Admon grid tools and site tools. On the three other nodes, we
installed the Admon agents. The site monitoring tools know the nodes of its site where
ViSaGe turns. For AMIBE, we used 3 nodes: one node for 1D or 3D and two other nodes for
2D phase. The CPU load worse case is fixed to 70 %.
We started by mounting the ViSaGe file system (Visagefs) on node1.
Before starting 2D stage, the Admon administration contacts the monitoring part to ask
about computing nodes state reserved for the task. The grid scheduler has not all control
over nodes tasks will be delegated. In our case study we have three nodes. The Admon
administration chooses nodes according to their workload.
The first test (Fig.3) showed the CPU’s nodes state in our experiment. This figure (Fig.3)
shows the normal case of our test: all the nodes are dedicated to ViSaGe. We conclude that
CPU usage percentage, in the 2D phase reaches highest percentage. Here, I must mention
that the 2D phase is computing of 2D faces followed by storage of information in order to
initialize the next step: we had CPU usage and I/O usage. That's mean, if the node2 is
loaded node and it was choose in the 2D phase. We will have an experiment failure or
increasing of service time.
114 Data Storage

Fig. 4. node2 is dedicated to the grid

In the other hand, the second test, represented by the figure 4, shows from the workload of
node2 (in terms of CPU load), that node2 is dedicated to the grid- that mean that this node is
used by another applications (local or on the grid).
Before the launching the 2D phase, we launch on the node2 applications simulating the use
of this node by a virtual user. The application of this user uses the nodes CPU, stores and
sends requests to the disk (read or write data). The goal of this operation is to increase the
load of the CPU in a progressive way in order to stabilize it on a maximum value. The
monitoring agent collects the percentages of CPU node and sends this significant change to
the monitoring tools. The monitoring tools, starting from the information collected by each
monitoring agent, test the node’s state. For ongoing on the 2D phase, Admon administration
doesn’t choose this node – “node2” for the 2D stage.
The other two nodes, (node3 and node4), aren't detected in term of dedicated node to the
grid, Admon launches commands to mount the Visagefs. For adding these nodes in the
same virtual space, Admon contacts the Vrt (the visage virtualization layer). The Vrt
replicates data in order to launch the 2D phase.
We finalize our experiment by the 3D phase that synthesizes information on the node1.

6. Conclusion
In this paper, we have described a non-intrusive, scalable administration and monitoring
system for high performance grid environment. It is a fundamental building block for
achieving and analyzing the performance of the storage virtualization system in a huge and
heterogeneous grid storage environment. It offers a very flexible and simple model that
collects nodes state information and requirements needed by the other services of our
virtualization storage system improving storage distributed data performance. It is based on
a multi-tiered hierarchical architecture the start-up of the monitoring and administration
system.
Administration and Monitoring Service for Storage Virtualization in Grid Environments 115

Future work will focus on developing interactive jobs with the storage virtualization system,
in order to study the relation between different system resources (CPU, disk, networks) and
tracing events, where Admon allows changing the storage distributed protocols (for
instance, replication) during runtime execution. Therefore, this work will be difficult since
we have many parameters that must be taken into consideration. However, this mechanism
will allow generalizing I/O workload model in order to improve applications throughput
during workload variations in grid environments.

7. References
Foster and al. (2002), The Anatomy of the Grid: Enabling Scalable Virtual Organizations, Intl J.
High-Performance Computing Applications, vol. 15, no. 3, 2001, pp. 200-222, June
2002.
Porter, G. & Katz, R.H. (2006). Effective Web Service Load Balancing Through Statistical
Monitoring. Communication of the ACM, March 2006.
Wolski, R; Spring, N.T. & Hayes, J. (1999). The Network Weather Service: A Distributed
Resource Performance Forecasting Service for Metacomputing. Future Generation
Computing Systems, Metacomputing Issue, 15(5-6): 157-768, Oct. 1999.
Gunter, D.; Tierney, B.; Crowley, B.; Holding, M. & Lee, J. (2000). NetLogger: A Toolkit for
Distributed System Performance Analysis, In Proceedings of the IEEE Mascots 2000,
Aug. 2000.
Cooke, A. et al. (2004). The Relational Grid Monitoring Architecture: Mediating information about
the grid, 2004
116 Data Storage
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 117

0
8

Power-Aware Memory Allocation for Embedded

Data-Intensive Signal Processing Applications
Florin Balasa1 , Ilie I. Luican2 , Hongwei Zhu3 and Doru V. Nasui4
1 Dept. of Computer Science, Southern Utah University, Cedar City, UT 84721
2 Dept. of Computer Science, University of Illinois at Chicago, Chicago, IL 60607
3 ARM, Inc., San Jose, CA 95134
4 American Int. Radio, Inc., Rolling Meadows, IL 60008

Abstract
Many signal processing systems, particularly in the multimedia and telecommunication do-
mains, are synthesized to execute data-intensive applications: their cost related aspects –
namely power consumption and chip area – are heavily influenced, if not dominated, by the
data access and storage aspects. This chapter presents a power-aware memory allocation
methodology. Starting from the high-level behavioral specification of a given application, this
framework performs the assignment of of the multidimensional signals to the memory layers
– the on-chip scratch-pad memory and the off-chip main memory – the goal being the reduc-
tion of the dynamic energy consumption in the memory subsystem. Based on the assignment
results, the framework subsequently performs the mapping of signals into the memory lay-
ers such that the overall amount of data storage be reduced. This software system yields a
complete allocation solution: the exact storage amount on each memory layer, the mapping
functions that determine the exact locations for any array element (scalar signal) in the spec-
ification, and, in addition, an estimation of the dynamic energy consumption in the memory
subsystem.

1. Introduction
Many multidimensional signal processing systems, particularly in the areas of multimedia
and telecommunications, are synthesized to execute data-intensive applications, the data
transfer and storage having a significant impact on both the system performance and the
major cost parameters – power and area.
In particular, the memory subsystem is, typically, a major contributor to the overall energy
budget of the entire system (8). The dynamic energy consumption is caused by memory ac-
cesses, whereas the static energy consumption is due to leakage currents. Savings of dynamic
energy can be potentially obtained by accessing frequently used data from smaller on-chip
memories rather than from the large off-chip main memory, the problem being how to op-
timally assign the data to the memory layers. Note that this problem is basically different
from caching for performance (15), (22), where the question is to find how to fill the cache
such that the needed data be loaded in advance from the main memory. As on-chip storage,
the scratch-pad memories (SPMs) – compiler-controlled static random-access memories, more
118 Data Storage

energy-efficient than the hardware-managed caches – are widely used in embedded systems,
where caches incur a significant penalty in aspects like area cost, energy consumption, hit
latency, and real-time guarantees. A detailed study (4) comparing the tradeoffs of caches as
compared to SPMs found in their experiments that the latter exhibit 34% smaller area and
40% lower power consumption than a cache of the same capacity. Even more surprisingly,
the runtime measured in cycles was 18% better with an SPM using a simple static knapsack-
based allocation algorithm. As a general conclusion, the authors of the study found absolutely
no advantage in using caches, even in high-end embedded systems in which performance is
important. 1 Different from caches, the SPM occupies a distinct part of the virtual address
space, with the rest of the address space occupied by the main memory. The consequence is
that there is no need to check for the availability of the data in the SPM. Hence, the SPM does
not possess a comparator and the miss/hit acknowledging circuitry (4). This contributes to
a significant energy (as well as area) reduction. Another consequence is that in cache mem-
ory systems, the mapping of data to the cache is done during the code execution, whereas in
SPM-based systems this can be done at compilation time, using a suitable algorithm – as this
chapter will show.
The energy-efficient assignment of signals to the on- and off-chip memories has been studied
since the late nineties. These previous works focused on partitioning the signals from the ap-
plication code into so-called copy candidates (since the on-chip memories were usually caches),
and on the optimal selection and assignment of these to different layers into the memory hier-
archy (32), (7), (18). For instance, Kandemir and Choudhary analyze and exploit the temporal
locality by inserting local copies (21). Their layer assignment builds a separate hierarchy per
loop nest and then combines them into a single hierarchy. However, the approach lacks a
global view on the lifetimes of array elements in applications having imperfect nested loops.
Brockmeyer et al. use the steering heuristic of assigning the arrays having the lowest access
number over size ratio to the lowest memory layer first, followed by incremental reassign-
ments (7). Hu et al. can use parts of arrays as copies, but they typically use cuts along the array
dimensions (18) (like rows and columns of matrices). Udayakumaran and Barua propose a
dynamic allocation model for SPM-based embedded systems (29), but the focus is global and
stack data, rather than multidimensional signals. Issenin et al. perform a data reuse analysis
in a multi-layer memory organization (19), but the mapping of the signals into the hierarchi-
cal data storage is not considered. The energy-aware partitioning of an on-chip memory in
multiple banks has been studied by several research groups, as well. Techniques of an ex-
ploratory nature analyze possible partitions, matching them against the access patterns of the
application (25), (11). Other approaches exploit the properties of the dynamic energy cost and
the resulting structure of the partitioning space to come up with algorithms able to derive the
optimal partition for a given access pattern (6), (1).
Despite many advances in memory design techniques over the past two decades, existing
computer-aided design methodologies are still ineffective in many aspects. In several previ-
ous works, the reduction of the dynamic energy consumption in hierarchical memory sub-
systems is addressed using in part enumerative approaches, simulations, profiling, heuristic
explorations of the solution space, rather than a formal methodology. Also, several models of
mapping the multidimensional signals into the physical memory were proposed in the past
(see (12) for a good overview).

1 Caches have been a big success for desktops though, where the usual approach to adding SRAM is to
configure it as a cache.
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 119

However, they all failed

(a) to provide efficient implementations,
(b) to prove their effectiveness in hierarchical memory organizations, and
(c) to provide quantitative measures of quality for the mapping solutions.
Moreover, the reduction of power consumption and the mapping of signals in hierarchical
memory subsystems were treated in the past as completely separate problems.
This chapter presents a power-aware memory allocation methodology. Starting from the high-
level behavioral specification of a given application, where the code is organized in sequences
of loop nests and the main data structures are multidimensional arrays, this framework per-
forms the assignment of of the multidimensional signals to the memory layers – the on-chip
scratch-pad memory and the off-chip main memory – the goal being the reduction of the dy-
namic energy consumption in the memory subsystem. Based on the assignment results, the
framework subsequently performs the mapping of signals into the memory layers such that
the overall amount of data storage be reduced. This software system yields a complete allo-
cation solution: the exact storage amount on each memory layer, the mapping functions that
determine the exact locations for any array element (scalar signal) in the specification, metrics
of quality for the allocation solution, and also an estimation of the dynamic energy consump-
tion in the memory subsystem using the CACTI power model (31). Extensions of the current
framework to support dynamic allocation are currently under development.
The rest of the chapter is organized as follows. Section 2 presents the algorithm that assigns
the signals to the memory layers, aiming to minimize the dynamic energy consumption in the
hierarchical memory subsystem subject to SPM size constraints. Section 3 describes the global
flow of the memory allocation approach, focusing on the mapping aspects. Section 4 discusses
on implementation and presents experimental results. Finally, Section 5 summarizes the main
conclusions of this research.

2. Power-aware signal assignment to the memory layers

The algorithms describing the functionality of real-time multimedia and telecommunication
applications are typically specified in a high-level programming language, where the code is
organized in sequences of loop nests having as boundaries linear functions of the outer loop
iterators. Conditional instructions are very common as well, and the multidimensional array
references have linear indexes (the variables being the loop iterators).2
Figure 1 shows an illustrative example whose structure is similar to the kernel of a motion
detection algorithm (9) (the actual code containing also a delay operator – not relevant in this
context). The problem is to automatically identify those parts of arrays from the given appli-
cation code that are more intensely accessed, in order to steer their assignment to the energy-
efficient data storage layer (the on-chip scratch-pad memory) such that the dynamic energy
consumption in the hierarchical memory subsystem be reduced.
The number of storage accesses for each array element can certainly be computed by the sim-
ulated execution of the code. For instance, the number of accesses was counted for every pair
of possible indexes (between 0 and 80) of signal A (see Fig. 1). The array elements near the

2 Typical piece-wise linear operations can be transformed into affine specifications (17). In addition,
pointer accesses can be converted at compilation to array accesses with explicit index functions (16).
Moreover, specifications where not all loop structures are for loops and not all array indexes are affine
functions of the loop iterators can be transformed into affine specifications that captures all the memory
references amenable to optimization (20). Extensions to support a larger class of specifications are thus
possible, but they are orthogonal to the work presented in this chapter.
120 Data Storage

Fig. 1. Code derived from a motion detection (9) kernel (m = n = 16, M = N = 64) and the
exact map of memory read accesses (obtained by simulation) for the 2-D signal A.

center of the index space are accessed with high intensity (for instance, A[40][40] is accessed
2,178 times; A[16][40] is accessed 1,650 times), whereas the array elements at the periphery
are accessed with a significantly lower intensity (for instance, A[0][40] is accessed 33 times
and A[0][0] only once).
The drawbacks of such an approach are twofold. First, the simulated execution may be com-
putationally ineffective when the number of array elements is very significant, or when the
application code contains deep loop nests. Second, even if the simulated execution were fea-
sible, such a scalar-oriented technique would not be helpful since the addressing hardware of
the data memories would result very complex. An address generation unit (AGU) is typically
implemented to compute arithmetic expressions in order to generate sequences of addresses
(26); a set of array elements is not a good input for the design of an efficient AGU.
Our proposed computation methodology for power-aware signal assignment to the memory
layers is described below, after defining a few basic concepts.
Each array reference M [ x1 (i1 , . . . , in )] · · · [ xm (i1 , . . . , in )] of an m-dimensional signal M, in the
scope of a nest of n loops having the iterators i1 , . . . , in , is characterized by an iterator
space and an index (or array) space. The iterator space signifies the set of all iterator vectors
i = (i1 , . . . , in ) ∈ Zn in the scope of the array reference, and it can be typically represented
by a so-called Z-polytope (a polyhedron bounded and closed, restricted to the set Zn ): { i ∈
Zn | A · i ≥ b }. The index space is the set of all index vectors x = ( x1 , . . . , xm ) ∈ Zm of
the array reference. When the indexes of an array reference are linear mappings with integer
coefficients of the loop iterators, the index space consists of one or several linearly bounded
lattices (27): { x = T · i + u ∈ Zm | A · i ≥ b , i ∈ Zn }. For instance, the array reference
A[i + 2 ∗ j + 3][ j + 2 ∗ k] from the loop nest
for (i=0; i<=2; i++)
for (j=0; j<=3; j++)
for (k=0; k<=4; k++)
if ( 6*i+4*j+3*k ≤ 12 ) ···
A[i+2*j+3][j+2*k] · · ·
    
   1 0 0   0 

 i i 

 0 1 0   0 
has the iterator space P =  j  ∈ Z3   j  ≥   . (The
  0 0 1   
0 

 k k 

−6 −4 −3 −12
inequalities i ≤ 2, j ≤ 3, and k ≤ 4 are redundant.)
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 121

Fig. 2. The mapping of the iterator space into the index space of the array reference A[i + 2 ∗
j + 3][ j + 2 ∗ k].

The
 A-elements of the array referencehavethe indices x, y:  
    i   i 
x 1 2 0  3
= T·i+u = j +  j  ∈ P . The points of the index
 y 0 1 2 0 
k k
space lie inside the Z-polytope { x ≥ 3 , y ≥ 0 , 3x − 4y ≤ 15 , 5x + 6y ≤ 63 , x, y ∈ Z},
whose boundary is the image of the boundary of the iterator space P (see Fig. 2). However, it
can be shown that only those points (x,y) satisfying also the inequalities −6x + 8y ≥ 19k − 30,
x − 2y ≥ −4k + 3, and y ≥ 2k ≥ 0, for some positive integer k, belong to the index space;
these are the black points in the right quadrilateral from Fig. 2. In this example, each point in
the iterator space is mapped to a distinct point of the index space; this is not always the case,
though.
Algorithm 1: Power-aware signal assignment to the SPM and off-chip memory layers
Step 1 Extract the array references from the given algorithmic specification and decompose the array
references for every indexed signal into disjoint lattices.

Fig. 3. The decomposition of the array space of signal M in 6 disjoint lattices.

122 Data Storage

The motivation of the decomposition of the array references relies on the following intuitive
idea: the disjoint lattices belonging to many array references are actually those parts of arrays
more heavily accessed during the code execution. This decomposition can be analytically per-
formed, using intersections and differences of lattices – operations quite complex (3) involving
computations of Hermite Normal Forms and solving Diophantine linear systems (24), com-
puting the vertices of Z-polytopes (2) and their supporting polyhedral cones, counting the
integral points in Z-polyhedra (5; 10), and computing integer projections of polytopes (30).
Figure 3 shows the result of such a decomposition for the three array references of signal M.
The resulting lattices have the following expressions (in non-matrix format):
L1 = { x = 0, y = t | 5 ≥ t ≥ 0}
L2 = { x = t1 , y = t2 | 5 ≥ t2 ≥ 1 , 2t2 − 1 ≥ t1 ≥ 1}
L3 = { x = 2t, y = t | 5 ≥ t ≥ 1}
L4 = { x = 2t1 + 2, y = t2 | 4 ≥ t1 ≥ t2 ≥ 1}
L5 = { x = 2t1 + 1, y = t2 | 4 ≥ t1 ≥ t2 ≥ 1}
L6 = { x = t, y = 0 | 10 ≥ t ≥ 1}
Step 2 Compute the number of memory accesses for each disjoint lattice.
The total number of memory accesses to a given linearly bounded lattice of a signal is com-
puted as follows:
Step 2.1 Select an array reference of the signal and intersect the given lattice with it. If the
intersection is not empty, then the intersection is a linearly bounded lattice as well (27).
Step 2.2 Compute the number of points in the (non-empty) intersection: this is the number of
memory accesses to the given lattice (as part of the selected array reference).
Step 2.3 Repeat steps 2.1 and 2.2 for all the signal’s array references in the code, cumulating
the numbers of accesses.
For example, let us consider one of signal A’s lattices3 { 64 ≥ x , y ≥ 16} obtained in Step
1. Intersecting it with the array reference A[k][l ] (see the code in Fig. 1), we obtain the lattice
{i = t1 , j = t2 , k = t3 , l = t4 | 64 ≥ t1 , t2 , t3 , t4 ≥ 16, t1 + 16 ≥ t3 ≥ t1 − 16, t2 + 16 ≥
t4 ≥ t2 − 16}. The size of this set is 2,614,689 , which is the number of memory accesses to the
given lattice as part of the array reference A[k][l ]. Since the given lattice is also included in
the other array reference4 in the code – A[i ][ j], a similar computation yields 1,809,025 accesses
to the same lattice as part of A[i ][ j]. Hence, the total amount of memory accesses to the given
lattice is 2,614,689+1,809,025=4,423,714.
Figure 4 displays a computed map of memory accesses for the signal A, where A’s index space
is in the horizontal plane xOy and the numbers of memory accesses are on the vertical axis Oz.
This computed map is an approximation of the exact map in Fig. 1 since the accesses within
each lattice are considered uniform, equal to the average values obtained above. The advan-
tage of this map construction is that the (usually time-expensive) simulation is not needed any
more, being replaced by algebraic computations. Note that a finer granularity in the decom-
position of the index space of a signal in disjoint lattices entails a computed map of accesses
closer to the exact map.
Step 3 Select the lattices having the highest access numbers, whose total size does not exceed the
maximum SPM size (assumed to be a design constraint), and assign them to the SPM layer. The other
lattices will be assigned to the main memory.

3 When the lattice has T=I – the identity matrix – and u=0, the lattice is actually a Z-polytope, like in this
example.
4 Note that in our case, due to Step 1, any disjoint lattice is either included in the array reference or disjoint
from it.
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 123

Fig. 4. Computed 3D map of memory read accesses for the signal A from the illustrative code
in Figure 1.

Storing on-chip all the signals is, obviously, the most desirable scenario in point of view of
dynamic energy consumption, which is typically impossible. We assume here that the SPM
size is constrained to smaller values than the overall storage requirement. In our tests, we
computed the ratio between the dynamic energy reduction and the SPM size after mapping;
the value of the SPM size maximizing this ratio was selected, the idea being to obtain the
maximum benefit (in energy point of view) for the smallest SPM size.

3. Mapping signals within memory layers

This design phase has the following goals: (a) to map the signals (already assigned to the
memory layers) into amounts of data storage as small as possible, both for the SPM and the
main memory; (b) to compute these amounts of storage after mapping on both memory layers
(allocation solution) and be able to determine the memory location of each array element from
the specification (assignment solution); (c) to use mapping functions simple enough in order
to ensure an address generation hardware of a reasonable complexity; (d) to ascertain that any
scalar signals (array elements) simultaneously alive are mapped to distinct storage locations.
Since the mapping models (13) and (28) play an important part in this section, they will be
explained and illustrated below.
To reduce the size of a multidimensional array mapped to memory, the model (13) considers
all the possible canonical5 linearizations of the array; for any linearization, the largest distance
at any time between two live elements is computed. This distance plus 1 is then the stor-
age “window” required for the mapping of the array into the data memory. More formally,
|WA | = min max { dist( Ai , A j ) } + 1, where |WA | is the size of the storage window of a
signal A, the minimum is taken over all the canonical linearizations, while the maximum is
taken over all the pairs of A-elements (Ai ,A j ) simultaneously alive.
This mapping model will be illustrated for the loop nest from Fig. 5(a). The graph above
the code represents the array (index) space of signal A. The points represent the A-elements
A[index1 ][index2 ] which are produced (and consumed as well) in the loop nest. The points
to the left of the dashed line represent the elements produced till the end of the iteration
(i = 14, j = 4), the black points being the elements still alive (i.e., produced and still used as

5 For instance, a 2-D array can be typically linearized concatenating the rows or concatenating the
columns. In addition, the elements in a given dimension can be mapped in the increasing or decreasing
order of the respective index.
124 Data Storage

Fig. 5. (a-b) Illustrative examples having a similar code structure. The mapping model by
array linearization yields a better allocation solution for the former example, whereas the
bounding window model behaves better for the latter one.

operands in the next iterations), while the circles representing A-elements already ‘dead’ (i.e.,
not needed as operands any more). The light grey points to the right of the dashed line are
A-elements still unborn (to be produced in the next iterations).
If we consider the array linearization by column concatenation in the increasing order of the
columns ((A[index1 ][index2 ], index1 =0,18), index2 =0,9), two elements simultaneously alive,
placed the farthest apart from each other, are A[9][0] and A[9][9]. The distance between them
is 9×19=171. Now, if we consider the array linearization by row concatenation in the increas-
ing order of the rows ((A[index1 ][index2 ], index2 =0,9), index1 =0,18), the maximum distance
between live elements is 99 (e.g., between A[4][5] and A[14][4]). For all the canonical lin-
earizations, the maximum distances have the values {99, 109, 171, 181}. The best canonical
linearization for the array A is the concatenation row by row, increasingly. A memory win-
dow WA of 99+1=100 successive locations (relative to a certain base address) is sufficient to
store the array without mapping conflicts: it is sufficient that any access to A[index1 ][index2 ]
be redirected to WA [(10 ∗ index1 + index2 ) mod 100].
In order to avoid the inconvenience of analyzing different linearization schemes, another
possibility is to compute a maximal bounding window WA = (w1 , . . . , wm ) large enough
to encompass at any time the simultaneously alive A-elements. An access to the element
A[index1 ] . . . [indexm ] can then be redirected without any conflict to the window location
WA [index1 mod w1 ] . . . [indexm mod wm ]; in its turn, the window is mapped, relative to a base
address, into the physical memory by a typical canonical linearization, like row or column
concatenation for 2-D arrays. Each window element wk is computed as the maximum differ-
ence in absolute value between the k-th indexes of any two A-elements (Ai ,A j ) simultaneously
alive, plus 1. More formally, wk = max { | xk ( Ai ) − xk ( A j )| } + 1, for k = 1, . . . , m. This
ensures that any two array elements simultaneously alive are mapped to distinct memory lo-
cations. The amount of data memory required for storing (after mapping) the array A is the
volume of the bounding window WA , that is, |WA | = Πm k =1 w k .
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 125

In the illustrative example shown in Fig. 5(a), the bounding window of the signal A is
WA = (11 , 10). It follows that the storage allocation for signal A is 100 locations if the
linearization model is used, and w1 × w2 =110 locations when the bounding window model is
applied. However, in the example shown in Fig. 5(b), where the code has a similar structure,
the bounding window model yields a better allocation result – 30 storage locations, since the
2-D window of A is WA = (5 , 6), whereas the linearization model yields 32 locations (the
best canonical linearization being the row concatenation in the increasing order of rows).
Our software system incorporates both mapping models, their implementation being based
on the same polyhedral framework operating with lattices, used also in Section 2. This is ad-
vantageous both from the point of view of computational efficiency and relative to the amount
of allocated data storage – since the mapping window for each signal is the smallest one of
the two models. Moreover, this methodology can be applied independently to the memory
layers, providing a complete storage allocation/assignment solution for distributed memory
organizations.
Before explaining the global flow of the algorithm, let us examine the simple case of a code
with only one array reference in it: take, for instance, the two nested loops from Fig. 5(b),
but without the second conditional statement that consumes the A-elements. In the bounding
window model, WA can be determined by computing the integer projections on the two axes
of the lattice of A[i ][ j], represented graphically by all the points inside the quadrilateral from
Fig. 5(b). It can be directly observed that the integer projections of this polygon have the
sizes: w1 = 11 and w2 = 7. In the linearization model, denoting x and y the two indexes,
the distance between two A-elements A1 ( x1 , y1 ) and A2 ( x2 , y2 ), assuming row concatenation
in the increasing order of the rows, is: dist( A1 , A2 ) = ( x2 − x1 )∆y + (y2 − y1 ), where ∆y
is the range of the second index (here, equal to 7) in the array space.6 Then, the A-elements
at a maximum distance have the minimum and, respectively, the maximum index vectors
relative to the lexicographic order. These array A-elements are represented by the points M =
A[2][7] and N = A[12][7] in Fig. 5(b), and dist( M, N ) = (12-2)×7 +(0-0)=70. Similarly, in the
linearization by column concatenation, the array elements at the maximum distance from each
other are still the elements with (lexicographically) minimum and maximum index vectors,
provided an interchange of the indexes is applied first. These are the points M = A[9][4] and
N  = A[4][10] in Fig. 5(b). More general, the maximum distance between the points of a live
lattice in a canonical linearization is the distance between the (lexicographically) minimum
and maximum index vectors, providing an index permutation is applied first. The distance
j j j
between the array elements Ai ( x1i , x2i , . . . , xm
i ) and A ( x , x , . . . , x ) is:
j 1 2 m
j j j
dist( Ai , A j ) = ( x1 − x1i )∆x2 · · · ∆xm + ( x2 − x2i )∆x3 · · · ∆xm + · · · + ( xm−1 − xm
i
−1 ) ∆xm +
j i ) where the index vector of A is lexicographically larger than of A (∆x is the range
( xm − xm j i i
of xi ).
Algorithm 2: For each memory layer (SPM and main memory) compute the mapping windows for
every indexed signal having lattices assigned to that layer.
Step 1 Compute underestimations of the window sizes on the current memory layer for each indexed
signal, taking into account only the live signals at the boundaries between the loop nests.
Let A be an m-dimensional signal in the algorithmic specification, and let P A be the set of dis-
joint lattices partitioning the index space of A. A high-level pseudo-code of the computation

6 To ensure that the distance is a nonnegative number, we shall assume that [ x2 y2 ] T  [ x1 y1 ] T relative to
the lexicographic order. The vector y = [y1 , . . . , ym ] T is larger lexicographically than x = [ x1 , . . . , xm ] T
(written y  x) if (y1 > x1 ), or (y1 = x1 and y2 > x2 ), . . . , or (y1 = x1 , . . . , ym−1 = xm−1 , and ym > xm ).
126 Data Storage

of A’s preliminary windows is given below. Preliminary window sizes for each canonical lin-
earization according to DeGreef’s model (13) are computed first, followed by the computation
of the window size underestimate according to Tronçon’s model (28) in the same framework
operating with lattices. The meaning of the variables are explained as comments.
for ( each canonical linearization C ) {
for ( each disjoint lattice L ∈ P A ) // compute the (lexicographically) minimum and
maximum ...
compute x min ( L) and x max ( L) ; // ... index vectors of L relative to C
for ( each boundary n between the loop nests n and n + 1 ) { // the start of the code is
boundary 0
let P A (n) be the collection of disjoint lattices of A, which are alive at the bound-
ary n ;
// these are disjoint lattices produced before the boundary and con-
sumed after it
let Xnmin = min L∈P A (n) { x min ( L)} and Xnmax = max L∈P A (n) { x max ( L)} ;
|WC (n)| = dist( Xnmin , Xnmax ) + 1 ; // The distance is computed in the canonical
linearization C
}
|WC | = maxn { |WC (n)| } ; // the window size according to (13) for the canonical
linearization C
} // (possibly, an underestimate)
for ( each disjoint lattice L ∈ P A )
for ( each dimension k of signal A )
compute xkmin ( L) and xkmax ( L) ; // the extremes of the integer projection of L
on the k-th axis
for ( each boundary n between the loop nests n and n + 1 ) { // the start of the code is
boundary 0
let P A (n) be the collection of disjoint lattices of A, which are alive at the boundary
n;
for ( each dimension k of signal A ) {
let Xkmin = min L∈P A (n) { xkmin ( L)} and Xkmax = max L∈P A (n) { xkmax ( L)} ;
wk (n) = Xkmax − Xkmin + 1 ; // The k-th side of A’s bounding window at
boundary n
}
}
for ( each dimension k of signal A ) wk = maxn {wk (n)} ; // k-th side of A’s window over
all boundaries
|W | = Π m k=1 wk ; // the window size according to (28) (possibly, an underestimate)
Step 1 finds the exact values of the window sizes for both mapping models only when every
loop nest either produces or consumes (but not both!) the signal’s elements. Otherwise, when
in a certain loop nest elements of the signal are both produced and consumed (see the illus-
trative example from Fig. 5(a)), then the window sizes obtained at the end of Step 1 may be
only underestimates since an increase of the storage requirement can happen inside the loop
nest. Then, an additional step is required to find the exact values of the window sizes in both
mapping models.
Step 2 Update the mapping windows for each indexed signal in every loop nest producing and con-
suming elements of the signal.
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 127

The guiding idea is that local or global maxima of the bounding window size |W | are reached
immediately before the consumption of an A-element, which may entail a shrinkage of some
side of the bounding window encompassing the live elements. Similarly, the local or global
maxima of |WC | are reached immediately before the consumption of an A-element, which
may entail a decrease of the maximum distance between live elements. Consequently, for
each A-element consumed in a loop nest which also produces A-elements, we construct the
disjoint lattices partially produced and those partially consumed until the iteration when the
A-element is consumed. Afterwards, we do a similar computation as in Step 1 which may
result in increased values for |WC | and/or |W |.
Finally, the amount of data memory allocated for signal A on the current memory layer is
|WA | = min { |W | , minC { |WC | } }, that is, the smallest data storage provided by the bound-
ing window and the linearization mapping models. In principle, the overall amount of data
memory after mapping is ∑ A |WA | – the sum of the mapping window sizes of all the sig-
nals having lattices assigned to the current memory layer. In addition, a post-processing step
attempts to further enhance the allocation solution: our polyhedral framework allows to effi-
ciently check weather two multidimensional signals have disjoint lifetimes, in which case the
signals can share the largest of the two windows. More general, an incompatibility graph (14)
is used to optimize the memory sharing among all the signals at the level of whole code.

4. Experimental results
A hierarchical memory allocation tool has been implemented in C++, incorporating the al-
gorithms described in this chapter. For the time being, the tool supports only a two-level
memory hierarchy, where an SPM is used between the main memory and the processor core.
The dynamic energy is computed based on the number of accesses to each memory layer. In
computing the dynamic energy consumptions for the SPM and the main (off-chip) memory,
the CACTI v5.3 power model (31) was used.
Table 1 summarizes the results of our experiments, carried out on a PC with an Intel Core 2
Duo 1.8 GHz processor and 512 MB RAM. The benchmarks used are: (1) a motion detection
algorithm used in the transmission of real-time video signals on data networks; (2) the kernel
of a motion estimation algorithm for moving objects (MPEG-4); (3) Durbin’s algorithm for
solving Toeplitz systems with N unknowns; (4) a singular value decomposition (SVD) up-
dating algorithm (23) used in spatial division multiplex access (SDMA) modulation in mobile
communication receivers, in beamforming, and Kalman filtering; (5) the kernel of a voice
coding application – essential component of a mobile radio terminal.
The table displays the total number of memory accesses, the data memory size (in storage
locations/bytes), and the dynamic energy consumption assuming only one (off-chip) memory
layer; in addition, the SPM size and the savings of dynamic energy applying, respectively, a
previous model steered by the total number of accesses for whole arrays (7), another previous
model steered by the most accessed array rows/columns (18), and the current model, versus
the single-layer memory scenario; the CPU times. The energy consumptions for the motion
estimation benchmark were, respectively, 1894, 1832, and 1522 µJ; the saved energies relative
to the energy in column 4 are displayed as percentages in columns 6-8. Our experiments
show that the savings of dynamic energy consumption are from 40% to over 70% relative
to the energy used in the case of a flat memory design. Although previous models produce
energy savings as well, our model led to 20%-33% better savings than them.
Different from the previous works on power-aware assignment to the memory layers, our
framework provides also the mapping functions that determine the exact locations for any
Data Storage

Application #Memory Mem. Dyn. energy SPM Dyn. energy Dyn. energy Dyn. energy CPU
parameters accesses size 1-layer [µJ] size saved (7) saved (18) saved [sec]
Motion detection 136,242 2,740 486 841 30.2% 44.5% 49.2% 4
M=N=32, m=n=4
Motion estimation 864,900 3,624 3,088 1,416 38.7% 40.7% 50.7% 23
M=32, N=16
Durbin algorithm 1,004,993 1,249 3,588 764 55.2% 58.5% 73.2% 28
N =500
SVD updating 6,227,124 34,950 22,231 12,672 35.9% 38.4% 46.0% 37
n =100
Vocoder 200,000 12,690 714 3,879 30.8% 32.5% 39.5% 8
Table 1. Experimental results.
128
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 129

array element in the specification. This provides the necessary information for the automated
design of the address generation unit, which is one of our future development directions.
Different from the previous works on signal-to-memory mapping, our framework offers a
hierarchical strategy and, also, two metrics of quality for the memory allocation solutions:
(a) the sum of the minimum array windows (that is, the optimum memory sharing between
elements of same arrays), and (b) the minimum storage requirement for the execution of the
application code (that is, the optimum memory sharing between all the scalar signals or array
elements in the code) (3).

5. Conclusions
This chapter has presented an integrated computer-aided design methodology for power-
aware memory allocation, targeting embedded data-intensive signal processing applications.
The memory management tasks – the signal assignment to the memory layers and their map-
ping to the physical memories – are efficiently addressed within a common polyhedral frame-
work.

6. References
[1] F. Angiolini, L. Benini, and A. Caprara, “An efficient profile-based algorithm for scratch-
pad memory partitioning,” IEEE Trans. Computer-Aided Design, vol. 24, no. 11, pp. 1660-
1676, Nov. 2005.
[2] D. Avis, “lrs: A revised implementation of the reverse search vertex enumeration al-
gorithm,” in Polytopes – Combinatorics and Computation, G. Kalai and G. Ziegler (eds.),
Birkhauser-Verlag, 2000, pp. 177-198.
[3] F. Balasa, H. Zhu, and I.I. Luican, “Computation of storage requirements for multi-
dimensional signal processing applications,” IEEE Trans. VLSI Systems, vol. 15, no. 4,
pp. 447-460, April 2007.
[4] R. Banakar, S. Steinke, B.-S. Lee, M. Balakrishnan, and P. Marwedel, “Scratchpad mem-
ory: A design alternative for cache on-chip memory in embedded systems,” in Proc. 10th
Int. Workshop on Hardware/Software Codesign, Estes Park CO, May 2002.
[5] A.I. Barvinok, “A polynomial time algorithm for counting integral points in polyhedra
when the dimension is fixed,” Mathematics of Operations Research, vol. 19, no. 4, pp. 769-
779, Nov. 1994.
[6] L. Benini, L. Macchiarulo, A. Macii, E. Macii, and M. Poncino, “Layout-driven memory
synthesis for embedded systems-on-chip,” IEEE Trans. VLSI Systems, vol. 10, no. 2, pp.
96-105, April 2002.
[7] E. Brockmeyer, M. Miranda, H. Corporaal, and F. Catthoor, “Layer assignment tech-
niques for low energy in multi-layered memory organisations,” in Proc. ACM/IEEE De-
sign, Automation & Test in Europe, Munich, Germany, Mar. 2003, pp. 1070-1075.
[8] F. Catthoor, S. Wuytack, E. De Greef, F. Balasa, L. Nachtergaele, and A. Vandecappelle,
Custom Memory Management Methodology: Exploration of Memory Organization for Embed-
ded Multimedia System Design, Boston: Kluwer Academic Publishers, 1998.
[9] E. Chan and S. Panchanathan, “Motion estimation architecture for video compression,”
IEEE Trans. on Consumer Electronics, vol. 39, pp. 292-297, Aug. 1993.
[10] Ph. Clauss and V. Loechner, “Parametric analysis of polyhedral iteration spaces,” J. VLSI
Signal Processing, vol. 19, no. 2, pp. 179-194, 1998.
130 Data Storage

[11] S. Coumeri and D.E. Thomas, “Memory modeling for system synthesis,” IEEE Trans.
VLSI Systems, vol. 8, no. 3, pp. 327-334, June 2000.
[12] A. Darte, R. Schreiber, and G. Villard, “Lattice-based memory allocation,” IEEE Trans.
Computers, vol. 54, pp. 1242-1257, Oct. 2005.
[13] E. De Greef, F. Catthoor, and H. De Man, “Memory size reduction through storage order
optimization for embedded parallel multimedia applications,” Parallel Computing, special
issue on “Parallel Processing and Multimedia,” Elsevier, vol. 23, no. 12, pp. 1811-1837,
Dec. 1997.
[14] G. De Micheli, Synthesis and Optimization of Digital Circuits, McGraw-Hill, 1994.
[15] J.Z. Fang and M. Lu, “An iteration partition approach for cache or local memory thrash-
ing on parallel processing,” IEEE Trans. Computers, vol. 42, no. 5, pp. 529-546, 1993.
[16] B. Franke and M. O’Boyle, “Compiler transformation of pointers to explicit array accesses
in DSP applications,” in Proc. Int. Conf. Compiler Construction, 2001.
[17] C. Ghez, M. Miranda, A. Vandecapelle, F. Catthoor, D. Verkest, “Systematic high-level
address code transformations for piece-wise linear indexing: illustration on a medical
imaging algorithm,” in Proc. IEEE Workshop on Signal Processing Systems, pp. 623-632,
Lafayette LA, Oct. 2000.
[18] Q. Hu, A. Vandecapelle, M. Palkovic, P.G. Kjeldsberg, E. Brockmeyer, and F. Catthoor,
“Hierarchical memory size estimation for loop fusion and loop shifting in data-
dominated applications,” in Proc. Asia-S. Pacific Design Automation Conf., Yokohama,
Japan, Jan. 2006, pp. 606-611.
[19] I. Issenin, E. Brockmeyer, M. Miranda, and N. Dutt, “Data reuse analysis technique for
software-controlled memory hierarchies,” in Proc. Design, Automation & Test in Europe,
2004.
[20] I. Issenin and N. Dutt, “FORAY-GEN: Automatic generation of affine functions for mem-
ory optimization,” Proc. Design, Automation & Test in Europe, 2005.
[21] M. Kandemir and A. Choudhary, “Compiler-directed scratch-pad memory hierarchy de-
sign and management,” in Proc. 39th ACM/IEEE Design Automation Conf., Las Vegas NV,
June 2002, pp. 690-695.
[22] N. Manjiakian and T. Abdelrahman, “Reduction of cache conflicts in loop nests,” Tech.
Report CSRI-318, Univ. Toronto, Canada, 1995.
[23] M. Moonen, P. V. Dooren, and J. Vandewalle, “An SVD updating algorithm for subspace
tracking,” SIAM J. Matrix Anal. Appl., vol. 13, no. 4, pp. 1015-1038, 1992.
[24] A. Schrijver, Theory of Linear and Integer Programming, New York: John Wiley, 1986.
[25] W. Shiue and C. Chakrabarti, “Memory exploration for low-power embedded systems,”
in Proc. 35th ACM/IEEE Design Automation Conf., June 1998, pp. 140-145.
[26] G. Talavera, M. Jayapala, J. Carrabina, and F. Catthoor, “Address generation optimiza-
tion for embedded high-performance processors: A survey,” J. Signal Processing Systems,
Springer, vol. 53, no. 3, pp. 271-284, Dec. 2008.
[27] L. Thiele, “Compiler techniques for massive parallel architectures,” in State-of-the-art in
Computer Science, P. Dewilde (ed.), Kluwer Acad. Publ., 1992.
[28] R. Tronçon, M. Bruynooghe, G. Janssens, and F. Catthoor, “Storage size reduction by in-
place mapping of arrays,” in Verification, Model Checking and Abstract Interpretation, A.
Coresi (ed.), 2002, pp. 167-181.
[29] S. Udayakumaran and R. Barua, “Compiler-decided dynamic memory allocation for
scratch-pad based embedded systems,” in Proc. Int. Conf. on Compilers, Architecture, and
Synthesis for Embedded Systems, pp. 276-286, New York NY, Oct. 2003.
Power-Aware Memory Allocation for Embedded Data-Intensive Signal Processing Applications 131

[30] S. Verdoolaege, K. Beyls, M. Bruynooghe, and F. Catthoor, “Experiences with enumer-

ation of integer projections of parametric polytopes,” in Compiler Construction: 14th Int.
Conf., R. Bodik (ed.), vol. 3443, pp. 91-105, Springer, 2005.
[31] S. Wilton and N. Jouppi, “CACTI: An enhanced access and cycle time model,” IEEE J.
Solid-State Circ., vol. 31, pp. 677-688, May 1996.
[32] S. Wuytack, J.-P. Diguet, F. Catthoor, and H. De Man, “Formalized methodology for data
reuse exploration for low-power hierarchical memory mappings,” IEEE Trans. VLSI Sys-
tems, vol. 6, no. 4, pp. 529-537, Dec. 1998.
132 Data Storage
High Throughput Architecture for High Performance NoC 133

X9

High Throughput Architecture for High

Performance NoC
Mohamed A. Abd El Ghany,
Magdy A. El-Moursy* and Mohammed Ismail**
Electronics Engineering Dept., German University in Cairo, Cairo, Egypt
Electronics Research Institute, Cairo, Egypt, Mentor Graphics Corporation, Cairo, Egypt*
Electrical Engineering Dept., The Ohio State University, Columbus, USA. The RaMSiS Group,
KTH, Sweden**

1. Introduction
As the number and functionality of intellectual property blocks (IPs) in System on Chips
(SoCs) increase, complexity of interconnection architectures of the SoCs have also been
increased. Different researches have been published in high performance SoCs; however,
the system scalability and bandwidth are limited. Network on Chip (NoC) is emerging as
the best replacement for the existing interconnection architectures. NoC is composed of
network of interconnects and number of temporary storage elements called switches. The
temporary storage element of different NoC architectures has different number of ports. The
main component of the port is the virtual channels. The virtual channels consist of several
buffers controlled by a multiplexer and an arbiter which grants access for only one buffer at
a time according to the request priority. When the number of buffers is increased, the
throughput increases. High throughput and low latency are the desirable characteristics of a
multi processing system. More research is needed to enhance performance of NoC
components (network of interconnects and the storage elements). Many NoC architectures
have been proposed in the past, e.g., SPIN (Guerrier & Greiner, 2000), CLICHÉ (Kumar et
al., 2002), Folded Torus (Dally & Towles, 2001), Octagon (Karim et al., 2002) and Butterfly
fat-tree (BFT) (Pande et al., 2003a). Among those, the butterfly fat tree (BFT) has found
extensive use in different parallel machines and shown to be hardware efficient (Grecu et al.,
2004a). The main advantage of the butterfly fat tree is that the number of storage elements in
the network converges to a constant irrespective of the number of levels in the tree network.
In the SPIN architecture, redundant paths contained within the fat tree structure are utilized
to reduce contention in the network. CLICHÉ (Chip-Level Integration of Communicating
Heterogeneous Elements) is simplest from a layout perspective and the local interconnections
between resources and storage elements are independent of the size of the network. In the
Octagon architecture, the communication between any two nodes takes at most two hops
within the basic Octagon unit.
After the NoC design paradigm has been proposed (Dally & Towles, 2001) ; (Kumar et al.,
2002) ; (Guerrier & Greiner, 2000) ; (Karim et al., 2002) ; (Pande et al., 2003) ; (Benini &
134 Data Storage

Micheli, 2002) ; (Grecu et al., 2004a), many researches on architectural and conceptual aspects
of NoC have been reported such as topology selection (Murali & Micheli, 2004), quality of
service (QoS) (Bolotin et al., 2004), design automation (Bertozzi et al., 2005) ; (Liang et al.,
2004) ; ( Pande et al., 2005a), performance evaluation (Pande et al., 2005b) ; (Salminen et al.,
2007) ; (Grecu et al., 2007a) and test and verification (Grecu et al., 2007b) ; (Kim et al., 2004) ;
(Murali et al., 2005). These researches have taken a top-down approach (a high level analysis
of NoC) and they didn’t touch the issues on a circuit level. However, a little research has
reported on design issues in implementation of NoC in the perspective of circuit level (Lee et
al., 2003) ; (Lee et al., 2004) ; (Lee & Kim et al., 2005) ; (Lee et al., 2006) ; (Lee; Lee & Yoo, 2005).
Although, they were implemented and verified on the silicon, they were only focusing on
implementation of limited set of architectures.
In large-scale NoC, power consumption should be minimized for cost-efficient
implementations. Although different researches have been published in NoCs, they were
only focusing on performance and scalability issues rather than power efficiency. Scaling
with power reduction is the trend in future technologies. Lowering supply voltage is the
most effective way to reduce power consumption. With lowering supply voltage, the
threshold voltage (VTH) has to be decreased to achieve high performance requirements.
Reducing VTH causes significant increase in the leakage component. Different researches
have been published in power minimization of high performance CMOS circuits (Khellah &
Elmasry, 1999) ; (Kao & Chandrakasan, 2000) ; (Kursun & Friedman, 2004).
In this chapter, different tradeoffs in designing efficient NoC including both elements of the
network (interconnects network and storage elements) are described. Building high
performance NoC is presented. In addition, a high throughput architecture is proposed. The
proposed architecture to achieve high throughput can improve the latency of the network.
The circuit implementation issues are considered in the proposed architecture. The switch
structure along with the interconnect architecture are shown in Fig. 1 for 2 IPs and 2 switches.
The proposed architecture is applied to different NoCs topologies. The efficiency and
performance are evaluated. To the best of our knowledge, this is the first in depth analysis on
circuit level to optimize performance of different NoC architectures.
This chapter is organized as follows: In Section 2, the proposed port architecture is presented.
The new High Throughput architecture is described in Section 3. In Section 4, power
characteristics for different high throughput architectures are provided. The performance and
overhead analysis of the proposed architecture are provided in Section 5. In Section 6, the
proposed design of low power NoC switch is described. Finally, conclusions are provided in
Section 7.
High Throughput Architecture for High Performance NoC 135

Fig. 1. proposed high throughput architecture.

2. Port architecture
The switch of different architectures has different number of ports. Each port of the switch
includes input virtual channels, output virtual channels, a header decoder, controller, input
arbiter and output arbiter as shown in (Pande et al., 2003a). When the number of virtual
channel is increased, the throughput increases. The input arbiter is used to allow only one
virtual channel to access a physical port. The input arbiter consists of a priority matrix and
grant circuits (Pande et al., 2003b).
The priority matrix stores the priorities of the requests. The grant circuits generate the
granted signals to allow only one virtual channel to access a physical port. The messages are
divided into fixed length flow control units (flits). When the granted virtual channel stores
one whole flit, it sends a full signal to controller. If it is a header flit, the header decoder
determines the destination. The controller checks the status of destination port. If it is
available, the path between input and output is established. All subsequent flits of the
corresponding packet are sent from input to output using the established path. The flits
from more than one input port may simultaneously try to access a particular output port.
The output arbiter is used to allow only one input port to access an output port.
Virtual channels consist of several buffers controlled by a multiplexer and an arbiter which
grants access for only one virtual channel at a time according to the request priority. Once
the request succeeds, its priority is set to be the lowest among all other requests. In the
proposed architecture, rather than using one multiplexer and one arbiter to control the
virtual channels, two multiplexer and two arbiters are employed as shown in Fig. 2. The
virtual channels are divided into two groups, each group controlled by one multiplexer and
one arbiter. Each group of virtual channels is supported by one interconnect bus as
described in Section 3. However looks trivial, this port architecture has a great influence on
the switch frequency and the throughput of the network.
Let us consider an example with the number of virtual channels of 8 channels. In the NoC
architecture, 8x8 input arbiter and 8x1 multiplexer are needed to control the input virtual
channels as shown in Fig. 2 (a). The 8x8 input arbiter consists of 8x8 grant circuit and 8x8
priority matrix. In the proposed architecture, two 4x4 input arbiters, two 4x1 multiplexers,
2x1 multiplexers and 2x2 grant circuit are integrated to allow only one virtual channel to
136 Data Storage

access a physical port as shown in Fig. 2 (b). The 4x4 input arbiter consists of 4x4 grant
circuit and 4x4 priority matrix. The values of the grant signals are determined by the
priority matrix. The number of grant signals equals to the number of requests and the
number of selection signals of the multiplexer. The area of 8x8 input arbiter is larger than
the area of two 4x4 input arbiters. Also, the area of 8x1multiplexer is larger than the area of
two 4x1 multiplexers. Consequently, the required area to implement the proposed switch
with the proposed architecture is less than the required area to implement the conventional
switch. In order to divide a 4x1 multiplexer into three 2x1 multiplexers, the 4x4 input arbiter
should be divided into three 2x2 input arbiters. The grant signals generated by three 2x2
input arbiter (6 signals) aren’t the same grant signals generated by the 4x4 input arbiter (4
signals). Therefore, the 4x4 input arbiter can’t be replaced by three 2x2 input arbiters unless
the number of interconnect buses is increased to be equal the number of virtual channels
groups. By increasing the number of interconnect buses, the metal resources and power
dissipation are increased as described in Section 5.
Without circuit optimization in BFT architecture, the change in the maximum frequency of
the switch with the number of virtual channels is shown in Fig. 3. When the number of
virtual channels is increased beyond four, the maximum frequency of the switch is
decreased. The throughput is saturated when the number of virtual channels is increased
beyond four (Pande et al., 2005b) for different number of ports. On the other hand, the
average message latency increases with the number of virtual channels. To keep the latency
low while preserving the throughput, the number of virtual channels is constrained to four
(Pande et al., 2003b),(Pande et al., 2005b). Throughput is a parameter that measures the rate
in which message traffic can be sent across a communication network. It is defined by
(Pande et al., 2005b):

������� �� �������� ���������� � �������� �������

�� � �1�
������� �� �� ������� � ������ �����
The throughput is proportional to the number of completed messages. The number of
completed messages increases with the number of virtual channels. Total transfer time of
messages decreases with the frequency of the switch. Therefore the throughput can be
improved by increasing the number of virtual channels or by increasing the frequency of
switch (Lee & Bagherzadeh, 2006). The HT-BFT switch is smaller than the BFT switch.
Therefore, the maximum frequency of the switch is improved. The change in the maximum
frequency of the proposed switch with the number of virtual channels is shown in Fig. 3 for
HT-BFT architecture. The number of virtual channels could be increased up to eight without
significant reduction in the operating frequency.
The frequency of the network switch is characterized with different number of virtual
channels for different network topologies and the proposed architectures as shown in Fig. 4.
As compared to the conventional architectures, the operating frequency of the proposed
architectures is decreased when the number of virtual channels is higher than eight rather
than four. As shown in Fig. 3 and Fig. 4, doubling the number of virtual channels does not
degrade the frequency of the switch (rather than 4 virtual channels, 8 virtual channels could
be used). However, a severe increase in the number of virtual channels (more than 8) could
degrade performance. Increasing the number of virtual channels would increase the traffic
going through the links (interconnects) between the switches, increasing the contention on
the bus and increasing the latency which each flit will experience. In order to improve
High Throughput Architecture for High Performance NoC 137

throughput, the links (interconnects) connecting the switches with each other should be
increased. Since the number of virtual channels could be doubled (from four to eight),
doubling the number of virtual channels between switches is proposed.

(a) (b)
Fig. 2. (a) Circuit diagram of switch port, (b) circuit diagram of High Throughput switch
port.

400
350
max. frequency (MHz)

300
250
200 BFT
150
100 HTBFT

50
0
2 3 4 5 6 7 8 10 12 16
Number of virtual channels

Fig. 3. Maximum frequency of a switch with different number of virtual channels for BFT
and HTBFT.
138 Data Storage

600

max. frequency (MHz) 500

400 Octagon
CLICHÉ 
300
SPIN
200 HT‐Octagon
HT‐CLICHÉ 
100
HT‐SPIN
0
2 3 4 5 6 7 8 10 12 16
Number of virtual channels

Fig. 4. Maximum frequency of a switch with different number of virtual channels for
different NoC architectures.

Let us consider an example of BFT architecture. The area required to implement the BFT
switch and HT-BFT switch is shown with different number of virtual channels in Fig. 5. The
HT-BFT architecture decreases the area of switch by 18%. Consequently, a system with eight
virtual channels achieves high throughput, high frequency and low latency while the area of
design is optimized. The architectures of different NoC topologies to achieve high
throughput network is discussed in Section 3.

40
Number of transistors 

35
30
25
(x104)

20
BFT
15
10 HTBFT
5
0
2 3 4 5 6
Number of virtual channels

Fig. 5. Area of a switch for different number of virtual channels.

3. High Throughput architecture

A novel interconnect template to integrate IP blocks using NoC architecture is proposed as
shown in Fig. 1. In the proposed architecture, rather than using a single interconnect bus
between each two elements of NoC (IP block and switch or two switches), two buses are
employed. The number of virtual channels can be doubled to get higher throughput. Each
bus will support half number of virtual channels to maintain the average latency.
High Throughput Architecture for High Performance NoC 139

Increasing the number of buses between two switches could improve the throughput by
optimizing the design of the switch on the circuit level as shown in Section II. However,
using two buses to connect two switches implies a consumption of the metal resources and
may be silicon area for the repeaters within long interconnect bus. The overhead of the
proposed architecture is discussed in Section 5. Applying the proposed high throughput
architecture on different NoC topologies is presented in the following subsections.

3.1 High Throughput BFT

The interconnect template of butterfly fat-tree topology was proposed in (Pande et al.,
2003a). This structure assumes a 4-ary tree with switches connected 4 down links and 2 up
links. Each group of 4 leaf nodes needs one switch. At the next level, half as many switches
are needed (every 4 switches on the lower level need 2 switches at the next level). This
relation continues with each succeeding level.
A novel interconnect template to integrate IP blocks using HT-BFT architecture is proposed
as shown in Fig. 6 (a). In the proposed HT-BFT architecture (Abd El Ghany et al., 2009a),
rather than using a single interconnect bus between each two switches, two buses are
employed. Each group of 4 IPs (no. 0, no. 1, no.2 and no.3) needs one switch (no.4). Each
switch in the first level (no. 4) connects to each switch in the second level (no. 5) by 2 buses
as shown in Fig. 6 (a). Each bus will support half number of virtual channels. Therefore, the
throughput can be improved while preserving the average latency.

3.2 High Throughput CLICHÉ

The mesh-interconnect topology called CLICHÉ (Chip-Level Integration of Communicating
Heterogeneous Elements) was proposed in (Kumar et al., 2002). The architecture consists of
m x n mesh of switches interconnecting the IP blocks. Every switch is connected to four
switches and one IP block. At the edges, the switches, except those at the corners, are
connected to three switches and one IP block. The number of switches equals to the number
of IP blocks. The interconnect template to integrate IP blocks using High Throughput
CLICHÉ (HT-CLICHÉ) architecture is shown in Fig. 6 (b) (Abd El Ghany et al., 2009b). The
interconnect bus between each two switches consists of two unidirectional links.

3.3 High Throughput Octagon

The interconnect template of Octagon topology was proposed in (Karim et al., 2002). The
basic unit of Octagon topology consists of eight nodes and 12 bidirectional buses. Each node
is associated with an IP block and a switch. Communication between any two nodes takes at
most two hops within the basic Octagon unit. The Octagon is extended to multidimensional
space for a system of more than eight nodes. The interconnect template to integrate IP
blocks using High Throughput Octagon (HT-Octagon) architecture is shown in Fig. 6 (c). For
the basic unit of HT-Octagon architecture, number of bidirectional buses equals to 24 rather
than 12 bidirectional buses in conventional Octagon architecture.

3.4 High Throughput SPIN

The interconnect template called SPIN (Scalable, Programmable, Integrated Network) was
proposed in (Guerrier & Greiner, 2000). This structure assumes a 4-ary tree with switches
140 Data Storage

connected 4 down links and 4 up links. Each group of 4 leaf nodes needs one switch. At the
next level, the same number of switches are needed (every 4 switches on the lower level
need 4 switches at the next level). This relation continues with each succeeding level. The
main rationale behind this approach is utilization of the redundant buses by the routers in
order to reduce contention in the network. Therefore, SPIN trades area overhead and extra
power dissipation for higher throughput. The interconnect template to integrate IP blocks
using High Throughput SPIN (HT-SPIN) architecture is shown in Fig. 6 (d). In the proposed
HT-SPIN architecture, the double number of buses is needed to connect between each two
switches or between an IP block and a switch. Due to the higher usage of on-chip resources
by the interswitch links, applying the high throughput architecture on SPIN topology is not
efficient for insignificant improvement of throughput as described in Section 5. The power
characteristics for different high throughput architectures are provided in Section 4.

(a) (b)

(c) (d)
Fig. 6. proposed interconnect architectures. (a) HTBFT. (b) HT- CLICHÉ. (c) HT-Octagon.
(d) HT-SPIN.

4. Power Characteristics
Power dissipation is a primary concern in high speed, high complexity integrated circuits
(IC). Power dissipation increases rapidly with the increase in frequency and transistor
density in integrated circuits. To achieve power efficient NoC, power dissipation need to be
characterized for different topologies. Communication network on chip contains three
primary parts; network switch, interswitch links (interconnects), and repeaters within
interswitch links as shown in Fig. 7. Including different sources of power consumption in
NoC, the total power dissipation of on chip network is defined as follows:
������ � ��������� � ����� � ���� �2�
��������� � ���������� � �������� �3�
High Throughput Architecture for High Performance NoC 141

Fig. 7. communication networks on chip

where ��������� is the total power dissipation of these switches forming the network.
��������� is the summation of switching (including dynamic and short circuit) power and
leakage power of switches. ����� is the total power dissipation of interswitch links. ���� is the
total power dissipation of the repeaters which are required for long interconnects. The
number of repeaters depends on the length of the interswich link. According to the topology
of NoC interconnects, the interswitch wire lengths, the number of repeaters and the number
of switches can be determined a priori.
The power consumption of interswitch links ����� and the power consumption of
repeaters are defined by (El-Moursy & Friedman, 2004)
�
����� � � ��� � �4�
���� � �������� � ������� � ������������ ���
�
�������� � ���� ���� �� ��� � ���
where �������� is the total dynamic power dissipation of repeaters, ���� is the number of
repeaters, ���� is the optimal repeater size and �� is the input capacitance of a minimum size
repeater. ������� is the total short-circuit power of repeaters. ������������ is the total leakage
power dissipation of repeaters. ������������ and ������� are negligible as compared to the
total dynamic power dissipation of repeaters [32 ]. The closed form equations for the power
dissipation of different high throughput NoC architectures are described in the following
subsections.

4.1 High Throughput Butterfly Fat Tree

In the HT-BFT, the interconnection is performed on levels of switching. The number of
switch levels can be expressed as ������ � ���� � � �, where N is the number of IP blocks.
The total number of switches in the first level is ��4. At each subsequent level, the number
of required switches reduces by a factor of 2 as shown in Fig. 6 (a). The interswitch wire
length and total number of switches are given by the following expression (Grecu et al.,
2004b):

√����
������ � �7�
2��������
������
� 1 � �1�2�
��������������� � � � ���
4 1 � 1�2
where ������ is the length of the wire spanning the distance between level a and a+1
switches, where a can take integer values between 0 and (levels-1). In the HT-BFT, The total
142 Data Storage

length of interconnect and the total number of repeaters can be determined from the
following equations:
√����
���������� � � � �������� � 2������ ���
2������ �����

���� 1 ���� 1 ����

��������������� � � 2������ �� �� � � � � �� ��
���� 2 ���� 4 ����
1 ����������������
� � �� �1��
2��� ����
Where ���� is the optimal length of the global interconnect (Li et al., 2005). Using the
number of switches, the total length of interconnect and the total number of repeaters, the
total power dissipation of HT-BFT architecture ( ���������� ) can be calculated using the
following expression:
������ �����
� ������� √����
���������� � � � �� ����� � � � � ��log � �� � �� �
� ����� ������� �����
�
2������ � ��� � �
� � ���� � ���� � ������� ������ ������ ����� � �11�
������ � � �
� ����
� � �
� ����
� � ��� ����
�
����
�� � 2������ ���� �� ��� �
���

4.2 High Throughput CLICHÉ Architecture

In HT-CLICHÉ architecture, the number of switches is equal to the number of IPs as shown
in Fig. 6 (b). The interswitch wire lengths can be determined from the following expression:
√����
��������� � �12�
√�
The number of horizontal interswitch links between switches equals to 2√��√� � 1�, and
the number of vertical interswich links between switches equals to 2√��√� � 1�. According
to the technology node, the optimal length of global interconnect can be obtained (Li et al.,
2005). Therefore, the total length of interconnect and the number of repeaters for HT-SPIN
can be calculated by:
������������� � 4 √���� �√� � 1� ������ �1��

√����
������������������� � 4 � � √� �√� � 1� ������ �14�
√� ����
Using the number of ports, number of switches, total length of interconnects and number
of repeaters, the total power consumption of the HT-CLICHÉ architecture can be
determined by the following expression:
�
������������� � � � ����� � 4 √���� �√� � 1� ������ � ��� �
√���� �
� 4� � √� �√� � 1� ������ ���� �� ��� � �1��
√� ����

4.3 High Throughput Octagon Architecture

For HT-Octagon architecture, there are four types of interswitch wire length as shown in
Fig. 6 (c) : First (connecting nodes 1-5 and 4-8), second (connecting nodes 2-6 and 3-7, third
High Throughput Architecture for High Performance NoC 143

(connecting nodes 1-8 and 4-5), forth (connecting nodes 1-2, 2-3, 3-4, 5-6, 6-7 and 7-8). the
interswitch wire lengths can be defined by:
��
�� � ����
4
�
�� � �� �� ������ � ����
4
�� � �� �� ������ ����
�
�� � ����
4
����
Where L is the length of four nodes; it equals to �4 � � �
�. �� is the summation of the
global interconnect width and space. Considering the interswitch wire lengths and the
optimal length of global interconnect, the total length of interconnect and number of
repeaters can be obtained by:
�������������� � �� � � ��4 �� ������ � ������ ���������� ����

�������������������� �
��� �� �� ������ � ��� �� �� ������ ��
�4 �� �� � 4 � ����
�� 4 � � � �� �� � �� ������ ��������� ����
��� ���� ���

Where ���������� is the number of basic octagon unit. The total power dissipation of the HT-
Octagon architecture can be determined by the following expression:
���� �
�������������� � � � ����� � ���� �� �
� � ��4 �� ������ � ������ ���������� � � ��� ��

���� ����
� �� � �� �� ������ � �� �
� � �� �� ������
��4 � � � 4� �� 4 � ��
���� ���� ����

����
�� �
�
�
�� � �� ������ ��������� � ���� �� ��� � ����
����

4.4 High Throughput SPIN Architecture

An interconnect template to integrate IP blocks using SPIN architecture was proposed as
shown in Fig. 6 (d). In large SPIN, the total number of switches is ���4 (Guerrier & Greiner,
2000). The interswitch wire length can be determined using eq. (7). In the HT-SPIN, The total
length of interconnect and the number of repeaters are defined by:
����������� � ���� √���� � ������ ����

√���� √���� √����

����������������� � �� � � � �� � � � �� �� � ������� ��4�
�� ��� ��� ���
The total power dissipation of the network architecture depends on the main three
parameters; the number of switches, the total length of interconnect and the number of
repeaters. The total power consumption of the HT-SPIN architecture (����������� ) can be
determined by:
144 Data Storage

�� � √���� √����
������������������������������ � � � �
��������� � � ������√������������� ������ � � ���� �� � �� � �
����� �����
√���� �
��������������������������� �� ���������� ���� ��� ��� �����������������������������������������
�����

4.5 Power Dissipation for Different NoC Architectures

According to the equations (11), (15), (22) and (25), the total power dissipation of the
network can be considered as a function of the number of IP blocks. The change in the
power consumption with the number of IP blocks for different network architectures is
shown in Fig. 8. The power consumption for different NoC architectures increases by
different rates with the number of IP blocks. The SPIN and Octagon architectures have
much higher rates of power dissipation. The BFT architecture consumes the minimum
power as compared to other NoC architectures.
100
dissipation of 16 IP blocks
Power dissipation as 
compared to power 

80

60 HT‐BFT

40 HT‐SPIN
HT‐CLICHÉ 
20
HT‐Octagon
0
16 32 64 128 256 512 1024
Number of IP blocks
Fig. 8. power dissipation of different NoC architectures

The percentage of the power dissipation of the interswitch links and repeaters is shown in
Fig. 9. For the SPIN and architecture, the power dissipation of the interswitch links and
repeaters equals to 25% of the total power dissipation of the architecture. For the BFT,
CLICHÉ and Octagon architectures, the percentage of power dissipation of the interswitch
links and repeaters decreases with the number of IP blocks.
60
interconnect and repeaters 

50
Power dissipation (%)
percentage of the 

40
HT‐BFT
30 HT‐SPIN
20 HT‐CLICHÉ 
10 HT‐Octagon
0
16 32 64 128 256 512 1024
Number of IP blocks
Fig. 9. power dissipation of interswitch links and repeaters for different NoC architectures.

The overhead analysis and simulation results are provided in Section 5.

High Throughput Architecture for High Performance NoC 145

5. Performance and Overhead analysis

The proposed high throughput architectures are implemented using the Application
Specific Integrated Circuit (ASIC) by Leonardo Spectrum synthesis tool, used for 90nm
technology node. Under uniform traffic assumption, the throughput for different NoC
architectures is calculated. The comparative analysis focuses on the frequency of the switch,
the throughput, the area of the switch and the power consumption is presented in the
following subsections.

5.1 Improvement of the Throughput

The proposed high throughput architecture trades the double number of virtual channels
for higher throughput while preserving the average latency. Therefore, the throughput of
using eight virtual channels in the HT-BFT is double the throughput of four virtual channels
in BFT. The average latency of HT-BFT with 8 virtual channels equals to the average latency
of BFT with 4 virtual channels. Considering the uniform traffic, the Maximum frequency of
the switch and the number of completed messages for HT-BFT, the throughput of HT-BFT is
determined. The variation of throughput with the number of virtual channels for HT-BFT
and BFT is shown in Fig. 10. In our architecture, when the number of virtual channels is
increased beyond eight, the throughput saturates. The architecture increases the throughput
of the network by 38%. The percentage of increasing of throughput for different high
throughput architectures is presented in Table 1. The maximum improvement in the
throughput is obtained in HT-CLICHÉ and HT-BFT. The increase in the throughput for HT-
SPIN is the minimum as compared to other high throughput architectures.
0,5

0,4
(flit/cycle/IP)
Throughput 

0,3
BFT
0,2 HTBFT

0,1
1 2 4 6 8 10 12 14 16
Number of virtual channels
Fig. 10. Throughput for different number of virtual channels.

architecture The percentage of increase in throughput

(%)
HT-BFT 38
HT-CLICHÉ 40
HT-Octagon 17
HT-SPIN 12
Table 1. the percentage of increase in the throughput for different high throughput
architectures
146 Data Storage

5.2 Overhead of High Throughput Architectures

With the advance in technology, the number of metal levels increases reaching twelve (ITRS,
2007). Metal resources on chip increase. Considering a chip size of 20 mm x 20 mm (Area),
technology node of 90 nm, and a system of 256 IP blocks, the length of interswitch links for
different NoC architectures is obtained. Given the optimal global interconnect width Wopt of
935 nm, optimal global interconnect spacing Sopt of 477 nm (Li et al., 2005), the global
interconnect pitch is Wopt + Sopt. Assuming all of global interconnects have the same line
width and line spacing, then the number of global interconnects Ngi per layer equals to
√����
��� � �
��� � ����
According to the NoC architecture, the total length of interswitch links are calculated. Using
the critical interconnect length of 2.54 mm, optimal repeater size of 174 (Li et al., 2005), the
number of repeaters is determined. The extra area and power required to implement
different high throughput NoC architectures are presented in the following subsections.

5.2.1 HT-BFT
It is possible to organize the butterfly fat tree so that it can be laid out in O(N) active area(IPs
and switches) and O(log(N)) wiring layers (Dehon, 2000). The basic strategy for wiring is to
distribute tree layers in pair of wire layers – one for horizontal wiring Ha+1,a and one for
vertical wiring Va+1,a. The length of horizontal part Ha+1,a equals to the length of vertical part
Va+1,a given that the chip is squared. More than one tree layer can share the same wiring
trace. High throughput architecture has the same number of switches, but the number of
wires and repeaters will be doubled. The length of interswitch wire depends on the number
of levels in BFT, which depends on the system size as shown in eq (7).
In the circuit implementation of HT-BFT, a bus between each two switches has 12 wires, 8
for data and 4 for control signals. Considering a system of 256 IP blocks, the length of Ha+1,a
and Va+1,a are calculated. The number of BFT levels is seven. Using the critical interconnect
length, the number of repeaters equals to 960 repeaters. The area of repeaters required to
implement the HT-BFT interswitch links equals to 20880 µm2 (it equals to the double area of
repeaters required for BFT interswitch links). The power consumption of repeaters and
switches required to implement the BFT and HT-BFT is presented in Table 2. The power
consumption required to implement HT-BFT is increased by 7% as compared with the
power consumption of BFT.
Percentage of
Power
power
dissipation of Power Total power
No. of dissipation of
Architecture repeaters and dissipation of dissipation
repeaters repeaters and
interswitch switches (mw) (mw)
interswitch
links (mw)
links (%)
BFT 960 1458.24 15663.68 17121.92 8.5
HT-BFT 1920 2916.48 15674.84 18591.32 15.7
Table 2. power consumption of repeaters and switches for BFT and HT-BFT

The horizontal wiring is distributed in the metal layer no. 11 and the vertical wiring is
distributed in the metal layer no. 12. The total length of horizontal wires needed equals to
4800 mm (it is 5 % of the total metal resources available in the metal 11). The same for total
length of vertical wires, it requires 5 % of the total metal resources available in the metal 12.
High Throughput Architecture for High Performance NoC 147

For the proposed design, the double number of interswitch links is required to achieve the
communication between each two switches. Therefore, the total metal resources required to
implement the proposed architecture will be 10%. The metal resources of HT-BFT
architecture equals to the double metal resources of BFT architecture. The extra metal
resources required to achieve the proposed architecture is negligible as compared to the
metal resources.
The percentage of the metal resources and power consumption of interswitch links and
repeaters for different technology node is shown in Table 3. With the advance in technology,
the available metal resources in the same die size are increased. Therefore, the number of
IPs could be increased. The number of switches is also increased. The required metal
resources to implement the BFT and HT-BFT are increased by fewer rates than the rates of
increase of the available metal resources with the advance in technology. The extra metal
resources and power consumption required to implement the HT-BFT decreases. The extra
power consumption required to achieve the proposed architecture is 1% of the total power
consumption of the BFT architecture. Also, the extra metal resources required for HT-BFT is
3% of the metal resources. The HT-BFT is more efficient with the advance in technology.
Percentage of Percentage of
power power
Percentage Percentage
No. consumption consumption
Technology No. of of BFT of HT-BFT
of of interswitch of interswitch
node levels metal metal
IPs links and links and
resources resources
repeaters for repeaters for
BFT HT-BFT (mw)
130 nm 500 6 10.26% 20.5% 4.95 % 9.89 %
90 nm 1000 7 4.49% 8.98% 4.02 % 8.04 %
65 nm 2500 9 1.32% 2.64% 2.55 % 5.1 %
45 nm 7500 10 0.59% 1.19% 2.97 % 5.94 %
Table 3. metal resources and power consumption of interswitch links and repeaters for HT-
BFT and BFT

5.2.2 HT-CLICHÉ
The CLICHÉ architecture with N IP blocks can be laid out in O(N) active area(IPs and
switches) and O(√�) interswitch links. In the circuit implementation of HT-CLICHÉ, a bus
between each two switches has 20 wires, 16 for data and 4 for control signals. Considering a
system of 256 IP blocks, the architecture consists of 16x16 mesh of switches interconnecting
the IPs. The length of horizontal links and vertical links equal to 1.25 mm. They are smaller
than the critical interconnect length. Therefore, no repeaters are needed within the
interswitch links. The power dissipation of the network is presented in Table 4 for CLICHÉ
and HT-CLICHÉ. The extra power dissipation required to implement HT-CLICHÉ for 256
IPs equals to 5%.
148 Data Storage

Power
Percentage of
consumption
Power power dissipation
of interswitch Total power
Architecture consumption of of repeaters and
links and dissipation (mw)
switches (mw) interswitch links
repeaters
(%)
(mw)
CLICHÉ 1398 24448 25846 5.4
HT-CLICHÉ 2796 24471 27267 10.25
Table 4. power consumption for CLICHÉ and HT-CLICHÉ architectures

Using the equation no. 12, the total length of interswitch links is calculated. Distributing the
horizontal and vertical interswitch links into metal 11 and metal 12 respectively, the metal
resources required to implement the horizontal wires equals to 7 % of the total metal
resources available in the metal 11. Also, the metal resources required to the vertical wires
equals to 7 % of the total metal resources available in the metal 12. Therefore, the total metal
resources required to implement the HT-CLICHÉ architecture will be 14%. The increasing
percentage of the metal resources for HT-CLICHÉ is negligible as compared to the metal
resources.
Since the interswitch links is short enough, there is no need for repeaters within the
interconnects, the power and metal resources consumed by CLICHÉ and HT-CLICHÉ are
shown in Table 5 for different technology nodes. With the advance in technology, the power
dissipation required to implement the HT-CLICHÉ is increased by less than 2% of the total
power consumption of the CLICHÉ architecture. The percentage of metal resources for HT-
CLICHÉ is increased by 35% as compared with the metal resources of CLICHÉ. The HT-
CLICHÉ trades extra metal resources for higher throughput.
Percentage of Percentage of
power power
Percentage of Percentage of
No. consumption of consumption of
Technology CLICHÉ HT-CLICHÉ
of interswitch interswitch
node metal metal
IPs links and links and
resources (%) resources (%)
repeaters for repeaters for
CLICHÉ (%) HT-CLICHÉ (%)
130 nm 361 7.6 14.1 21 43
90 nm 729 4.8 9.1 22 44
65 nm 1849 2.7 5.2 28 57
45 nm 5625 1.4 2.7 36 71
Table 5. Power consumption of interswitch links and repeaters for HT-CLICHÉ and
CLICHÉ

5.2.3 HT- Octagon

The HT-Octagon architecture has the same number of switches, but the number of wires and
repeaters will be doubled. A bus between each two switches has 12 wires, 8 for data and 4
for control signals. Considering a system of 256 IP blocks, the length of interswitch links is
obtained. According to the critical interconnect length (Li et al., 2005), the number of
repeaters equals to 7680 repeaters. The power consumption required to implement the
Octagon and HT-Octagon architectures is presented in Table 6. Due to the extra interswitch
High Throughput Architecture for High Performance NoC 149

links required to implement HT-Octagon architecture, the power consumption is increased

by 6% as compared with the power consumption of Octagon topology.
By distributing the wiring of HT-Octagon architecture into the metal 11, the total length
of wires needed equals to 7057.92 mm. The architecture consumes 8 % of the total metal
resources available in the metal 11. In the proposed design, the double number of
interswitch links is utilized to implement HT-Octagon architecture. Therefore, the total
metal resources required to implement the proposed architecture will be 16%.
Power
Percentage of
consumption
Power power dissipation
of interswitch Total power
Architecture consumption of of repeaters and
links and dissipation (mw)
switches (mw) interswitch links
repeaters
(%)
(mw)
Octagon 1094.12 19861.04 20955 5.2
HT-Octagon 2188.24 19844.08 22072.3 9.9
Table 6. power consumption of switches for Octagon and HT-Octagon

The percentage of power consumption and metal resources required to implement the
Octagon and HT-Octagon networks in different technologies are shown in Table 7. By
increasing the number of IP blocks with the advance in technology, the extra power
consumption required to implement the proposed architecture is decreased. The extra
power consumption is 2% of the total power consumption of the Octagon architecture. The
percentage of extra metal resources for HT-Octagon is 25% of the available metal resources.
Percentage of
Percentage of
power
power
consumption of Percentage of Percentage of
No. consumption
interswitch Octagon HT-Octagon
Technology node of of interswitch
links and metal metal
IPs links and
repeaters for resources (%) resources (%)
repeaters for
HT-Octagon
Octagon (%)
(%)
130 nm 361 7.6 14.1 13 26
90 nm 729 4.78 9.1 13 26
65 nm 1849 2.8 5.4 17 35
45 nm 5625 1.6 3.1 25 50
Table 7. Power consumption of interswitch links and repeaters for HT-Octagon and Octagon

5.2.4 HT-SPIN
By applying the high throughput architecture on SPIN topology, the length of interswitch
links and number of repeaters are calculated by eq. (22) and eq. (23) respectively.
Considering a system of 256 IP blocks, the number of repeaters equals to 12288 repeaters.
The area of repeaters required to implement the HT-SPIN interswitch links equals to 267264
µm2 (it equals to the double area of repeaters required for SPIN interswitch links). The
horizontal wires and vertical wires are distributed into metal 11 and metal 12 respectively.
The length of horizontal wires needed consumes 28 % of the total metal resources available
in the metal 11. The vertical wires needed consume 28 % of the total metal resources
available in the metal 12. The total metal resources required to implement the proposed HT-
150 Data Storage

SPIN architecture will be 56%.The power consumption of interswitch links, repeaters and
switches required to implement the SPIN and HT-SPIN is presented in Table 8. The extra
power dissipation required by the interswitch links and repeaters for HT-SPIN architecture
(with 256 IPs) equals to 15% as compared with the total power dissipation.

Percentage of
Power
power
dissipation of Power Total power
No. of dissipation of
Architecture repeaters and dissipation of dissipation
repeaters repeaters and
interswitch switches (mw) (mw)
interswitch
links (mw)
links (%)
SPIN 12288 10612.99 32263.68 42876.67 24.75
HT-SPIN 24576 21225.98 32280.96 53506.94 39.67
Table 8. power consumption of repeaters and switches for SPIN and HT-SPIN

For different technologies, the power consumption and metal resources required to
implement the SPIN and HT-SPIN are shown in Table 9. With the advance in technology,
the extra power consumption required to achieve the proposed HT-SPIN architecture is 15%
of the total power consumption of the architecture. The percentage of extra metal resources
needed is more than 100% of the metal resources. Therefore, the overhead in the HT-SPIN is
high. Applying the high throughput architecture on the SPIN topology is not recommended.
Percentage of
Percentage of
power
power Percentage Percentage
consumption
No. consumption of SPIN of HT-SPIN
Technology of
of of interswitch metal metal
node interswitch
IPs links and resources resources
links and
repeaters for (%) (%)
repeaters for
HT-SPIN (%)
SPIN (%)
130 nm 400 41.2 58.4 21 42
90 nm 800 33.6 50.3 30 59
65 nm 2000 32.6 49.1 59 118
45 nm 6000 28.1 43.8 126 253
Table 9. Power consumption of interswitch links and repeaters for HT-SPIN and SPIN

Since the proposed architecture increases the power dissipation, a low power NoC switch is
proposed in Section 6.

6. Low power noc switch design

The switch of BFT has six ports, four children ports and two parent ports. Each port can be
used as either input port or output port. If the port considers as input port, the input virtual
channels, header decoder and crossbar are active. If the port considers as output port, the
output virtual channels are active.
In the proposed design, only one part (input part or output part) is activated as shown in
Fig. 11. The stand-by transistors (M1) disconnect the input circuit from the supply voltage
High Throughput Architecture for High Performance NoC 151

during the output mode. The stand-by transistors (M2) disconnect the output circuit from
the supply voltage during the input mode. There is no need for the new control signals to
control the stand-by transistors (M1 and M2). The acknowledgment signals (Ack_in and
Ack_out) developed by the control unit are used to control the stand-by transistors M1 and
M2 respectively. Using the number of virtual channels ሺܰ௏஼ ሻ, The number of stand-by
transistors equals to ͵ ൅ ʹܰ௏஼ . The number of virtual channels is limited (it is not more 16
virtual channels (Abd El Ghany et al., 2009a)). By comparing the number of stand-by
transistors with the total number of transistors required to implement the NoC port (as
described is Section 2), the number of stand-by transistors is less than 1% of the total
number of transistors. Therefore, the area overhead in the proposed design is negligible as
compared to the area of NoC switch. The total power dissipation can be reduced by using
power gating technique.

Fig. 11. proposed design for low power NoC port.

Using the Cadence tools and 90nm technology node, the proposed low power NoC switch is
implemented. The power dissipation of BFT switch is determined. The total power
dissipation of the BFT switch equals to 41.29 mW. The total power dissipation of the port
during the input mode equals to 6.79 mW. The total power dissipation of the port during the
output mode equals to 6.57 mW. In the proposed BFT switch design of one virtual channel,
the power dissipation of the main components of the port for the active mode and sleep
mode is obtained as shown in Table 10. According to the mode of operation, the activation
of the component is determined. In the Input mode, the input FIFO, header decoder and
crossbar are activated, while the output FIFO is switched to sleep mode. The power
dissipation of the port will be 5.68 mW. In the output mode, the output FIFO is activated
while the input FIFO, the header decoder and cross bar are switched to sleep mode. The
power dissipation of the port equals to 3.79 mW. Therefore, the average power dissipation of
the proposed switch equals to 29,59 mW. The average power dissipation of the proposed
BFT switch is decreased by 28.32 % as compared to the average power dissipation of the
conventional BFT switch.
152 Data Storage

power dissipation in power dissipation in Percentage of

Component active mode sleep mode reduction in power
(mW) (µW) dissipation (%)
Input FIFO 3.618 0.1029 97.15
Header decoder 0.955 0.2157 77.41
Crossbar 0.473 0.1274 73.07
Output FIFO 3.562 0.1003 97.18
Table 10. the power dissipation of the main components of the BFT switch

The power consumption of BFT switch increases with the number of virtual channels as
shown in Fig. 12. Applying the leakage power reduction technique on the BFT with different
number of virtual channels, the power reduction increases with the number of virtual
channels. The percentage of power reduction equals to 28 % when the number of virtual
channels equals to one. The percentage of power reduction of BFT switch with 12 virtual
channels equals to 45%. Increasing the number of virtual channels can improve the
throughput in an on- chip interconnect network. By optimizing the design on the circuit
levels, the high throughput can be provided by eight virtual channels (Abd El Ghany et al.,
2009a). Using the leakage power reduction technique, the power consumption of BFT switch
with 8 virtual channels is reduced by 44 %.
With the advance in technology, the number of IPs implemented in the same system size is
increased. The effect of power gating technique on the HT-BFT is presented in Fig. 13. The
power consumption of HT-BFT architecture using the leakage power reduction technique
(HT-BFT-PR) is less than the power consumption of the conventional BFT architecture.
5
total power dissipation of one 

total power disipation of one 
switch as compared to the 

4
BFT
3
VC

2
BFT using leakage 
power reduction 
1
technique
0
1VC 2VC 3VC 4VC 6VC 8VC 10VC 12VC

Fig. 12. power dissipation of a switch with different number of virtual channels.
High Throughput Architecture for High Performance NoC 153

80
70
Power dissipation (W)
60
50
40 BFT
30 HTBFT
20
HTBFT‐PR
10
0
16 32 64 128 256 512 1024
Number of IP blocks
Fig. 13. power dissipation of the HT-BFT using the power reduction technique

The power consumption ܲ௦௪௜௧௖௛௘௦ of switches for different high throughput architectures is
obtained as shown in Table 11. The power consumption of these switches is more than 80%
of the total power consumption of the on chip network. Switching off the power supply is
an efficient technique to reduce the total power dissipation of NoC. The minimum power
consumption can be obtained by using the BFT architecture as presented in Table 11. Using
the leakage power reduction technique, the power consumption for different NoC
architectures is determined. The overall power consumption, includes the power
consumption of the interswitch links and repeaters, is decreased up to 33%.
Network Total Pswitches Total power using Percentage of
architecture power power reduction power
(mW) technique (mW) reduction
mW %
HT-BFT 18591.32 15674.84 84 15104.44 19%
HT-SPIN 53506.94 32280.96 60 46312.7 13%
HT-CLICHÉ 26148.64 24471 94 18608.16 29%
HT-Octagon 22072.32 19884.08 90 16253.04 26%
Table 11. the total power consumption of different network architectures with 256 IPs

7. Conclusions
In this chapter, the high throughput NoC architecture is proposed to increase the throughput
of the switch in NoC. The proposed architecture can also improve the latency of the network.
The proposed high throughput interconnect architecture is applied on different NoC
architectures. The architecture increases the throughput of the network by more than 38%
while preserving the average latency. The area of high throughput NoC switch is decreased
by 18% as compared to the area of BFT switch. The total metal resources required to
implement the proposed high throughput NoC is increased by less than 10 % as compared to
the metal resources required to implement the conventional NoC design.
The power characterization for different high throughput NoC architectures is developed.
The extra power consumption required to achieve the proposed high throughput NoC
architecture is less than 15% of the total power consumption of the NoC architecture. Low
power switch design is proposed. The power reduction technique is applied to different high
154 Data Storage

throughput NoC architectures. The technique reduces the overall power consumption of the
network by up to 29%.
The relation between throughput, number of virtual channels and switch frequency is
analyzed. The simulation results demonstrate the performance enhancements in terms of
throughput, number of virtual channels, switch frequency and power dissipation. It is shown
that optimizing the circuit can increase the number of virtual channels without degrading the
frequency. The throughput of different NoC architectures is also improved with the proposed
architecture. The minimum power consumption and the minimum area can be obtained by
using HT-BFT as compared to other high throughput NoC architectures. The extra metal
resources required to achieve the proposed HT-BFT is negligible as compared to the metal
resources of the network. The extra power consumption required to achieve the proposed
HT-BFT is eliminated by using the leakage power reduction technique.

8. References
Abd El Ghany, M. A.; El-Moursy, M. & Ismail, M. (2009a) “High Throughput Architecture
for High Performance NoC” Proceedings of IEEE International Symposium on Circuits
and Systems (ISCAS), May, 2009 (in publication)
Abd El Ghany, M. A.; El-Moursy, M. & Ismail, M. (2009b) “High Throughput Architecture
for CLICHÉ Network on Chip” Proceedings of the IEEE International SoC Conference,
September, 2009
Benini, L. & Micheli, G. de (2002) “Networks on chips: A new SoC paradigm,” IEEE
Computer, vol. 35, no. 1, pp. 70–78, Jan. 2002
Bertozzi, D.; Jalabert, A. & Murali, S. et al., (2005) “NoC synthesis flow for customized
domain specific multiprocessor systems-on-chip,” IEEE transactions on Parallel and
Distributed Systems, vol. 16, no. 2, pp. 113–129, February 2005
Bolotin, E.; Cidon, I.; Ginosar, R. & Kolodny, A. (2004) “QNoC: QoS architecture and design
process for network on chip,” Journal of Systems Architecture, vol. 50, no. 2–3, pp.
105–128, February 2004
Dally, W. J. & Towles, B. (2001) “Route packets, not wires: on-chip interconnection
networks”, In Proceedings of Design Automation Conference, pp 684–689, June 2001
Dehon, A. (2000) “Compact, Multilayer layout for butterfly fat-tree”, In Proceedings of The
12th ACM Symposium on Parallel algorithm Architectures, pp. 206- 215, July 2000
El-Moursy, M. A. & Friedman, E. G. (2004) “optimum wire sizing of RLC interconnect with
repeaters”, Integration, the VLSI journal, vol. 38, no. 2, pp. 205-225, 2004
Grecu, C.; Pande, P. P.; Ivanov, A. & Saleh, R. (2004a) “Structured Interconnect Architecture:
A Solution for the Non-Scalability of Bus-Based SoCs,” Proceedings of Great Lakes
Symposium on VLSI, pp. 192-195, April 2004
Grecu, C.; Pande, P. P.; Ivanov, A. & Saleh, R. (2004b) “Ascalable Communication-Centric
SoC Interconnect Architecture”, In Proceedings of IEEE International Symposium On
Quality Electronic Design, pp. 22- 24, March, 2004
Grecu, C.; Pande, P.; Ivanov, A.; Marculescu, R.; Salminen, E. & Jantsch, A. (2007a)
“Towards open network-on-chip benchmarks,” In Proceedings of the International
Symposium on Network on Chip, pp. 205, May 2007
High Throughput Architecture for High Performance NoC 155

Grecu, C.; Ivanov, A.; Saleh, R. & Pande, P. (2007b) “Testing network-on-chip
communication fabrics,” IEEE transactions Computer-Aided Design of Integrated
Circuits and Systems, vol. 26, no. 10, pp. 2201–2014, December 2007
Guerrier, P. & Greiner, A. (2000) “A generic architecture for on-chip packet-switched
interconnections”, In Proceedings of Design, Automation and Test in Europe Conference
and Exhibition, pp. 250–256, March 2000
ITRS 2007 Documents, http://itrs.net/Links/2007ITRS/Home2007.htm
Kao, J. T. & Chandrakasan, A. P. (2000) “Dual-Threshold Voltage Techniques for Low-Power
Digital Circuits “, IEEE Journal of Solid-State Circuits, vol. 35(7), pp. 1009- 1018, July
2000
Karim, F.; Nguyen, A. & Sujit Dey, (2002) “An Interconnect Architecture for Networking
Systems on Chips,” IEEE Micro, vol. 22, no. 5, pp. 36-45, September 2002
Khellah, M. M. & Elmasry, M. I. (1999) “ Power minimization of high-performance
submicron CMOS circuit using a dual-V/sub dd/ dual-V/sub th/ (DVDV)
approach” In Proceeding of the International symposium on Low Power Electronics and
Design , pp. 106-108, 1999
Kim, J.-S.; Hwang, M.-S & Roh, S. et al., (2004) “On-chip network based embedded core
testing,” In Proceedings of the IEEE International SoC Conference, pp. 223–226,
September 2004
Kumar, S. et al., (2002) “A Network on Chip Architecture and Design Methodology,” In
Proceedings of the IEEE Computer Society Annual Symposium on VLSI, pp. 117-124,
2002
Kursun, V. & Friedman, E. G. (2004) “Sleep switch dual threshold voltage domino logic with
reduced standby leakage current,” IEEE transactions on VLSI systems, 12(5), pp. 485-
496, May 2004
Lee, S.-J.; Song, S.-J. & Lee, K. et al. (2003) “An 800MHz Star-Connected On-Chip Network
for Application to Systems on a chip”, IEEE Digest of International Solid State Circuits
Conference, vol. 1, pp. 468-489, February, 2003
Lee, K.; Lee, S.-J. & Kim, S.-E. et al. (2004) “A 51mW 1.6GHz On-Chip Network for Low
power Heterogeneous SoC Platform”, IEEE Digest of International Solid State Circuits
Conference, vol. 1, pp.152-518, February, 2004
Lee, S.-J.; Kim, K. & Kim, H. et al. (2005) “Adaptive Network-on-Chip with Wave-Front
Train Serialization Scheme”, IEEE Digest of Symposium on VLSI Circuits, pp. 104-107,
June, 2005
Lee, S.-J.; Lee, K. & Yoo, H.-J. (2005)“Analysis and Implementation of Practical Cost-
Effective Network-onChips”, IEEE Design & Test of Computers Magazine (Special
Issue for NoC), September 2005
Lee, K.; Lee, S.-J. & Yoo, H.-J. (2006) “Low-Power Networks-on-Chip for High-Performance
SoC Design”, IEEE Transactions on Very Large Scale Integration Systems, vol. 14, no.2,
pp.148-160, February 2006
Lee, S. & Bagherzadeh, N. (2006) “Increasing the Throughput of an Adaptive Router in
Network-on Chip(NoC)”, In Proceedings of 4th International Conference on
Hardware/Software Codesign and System Synthesis CODES+ISSS’06, pp. 82-87, Oct.
2006
156 Data Storage

Liang, J.; Laffely, A.; Srinivasan, S. & Tessier, R. (2004) “An architecture and compiler for
scalable on-chip communication,” IEEE transactions on VLSI Systems, vol. 12, no. 7,
pp. 711–726, July 2004
Li, X.-C.; Mao, J.-F.; Huang, H.-F. & Liu, Y. (2005) “Global interconnect width and spacing
optimization for latency, bandwidth and power dissipation,” IEEE Transactions on
Electron Devices, vol. 52, no. 10, pp. 2272–2279, Oct. 2005
Murali, S. & De Micheli, G. (2004) “SUNMAP: A Tool for Automatic Topology Selection and
Generation for NoCs”, IEEE Proceedings of Design Automation conference, pp. 914-919,
June 2004
Murali, S.; Theocharides, T. & Vijaykrishnan, N. et al., (2005) “Analysis of error recovery
schems for networks on chips,” IEEE Design and test, vol. 22, no. 5, pp. 434–442,
October 2005
Pande, P. P.; Grecu, C.; Ivanov, A. & Saleh, R. (2003a) “Design of a Switch for Network on
Chip Applications,” In Proceedings of The 2003 International Symposium on Circuits
and Systems, vol. 5, pp. 217-220, May 2003
Pande, P. P.; Grecu, C.; Ivanov, A. & Saleh, R. (2003b) ”High-Throughput Switch-Based
Interconnect for Future SoCs”, the 3rd IEEE International workshop on SoC for real-time
Applications, pp 304-310, July 2003
Pande, P. P.; Grecu, C.; Ivanov, A. & Saleh, R. (2005a) “Design, synthesis, and test of
networks on chips,” IEEE Design and Test of Computer, vol. 22, no. 5, pp. 404–413,
Aug. 2005
Pande, P. P.; Grecu, C.; Jones, M.; Ivanov, A. & Saleh, R. (2005b) “Performance Evaluation
and Design Trade-Offs for Network-on-Chip Interconnect Architectures”, IEEE
Transaction on Computers, vol. 54, no. 8, Aug. 2005
Salminen, E.; Kulmala, A. & H¨am¨al¨ainen, T. (2007) “On network-on-chip comparison,” In
Proceedings of the Euromicro conference on Digital System Design Architecture, August
2007, pp. 503–510
Non-volatile memory interface protocols for smart sensor networks and mobile devices 157

10
X

Non-volatile memory interface protocols for

smart sensor networks and mobile devices
Octavian Postolache, Pedro Silva Girão and José Miguel Dias Pereira
Instituto de Telecomunicações
Portugal

1. Introduction
Data acquisition, storage and transmission are mandatory requirements for different
applications in the area of smart sensors and sensor networks.
Different architectures and scenarios can be considered. Thus for smart sensors architectures
(Frank, 2002) based on the IEEE1451.X standard (IEEE, 2007) the acquired data can be
processed at the smart sensor level using data of the so-called standard template TEDS
(Honeywell, 2009) stored in a non-volatile memory (Brewer & Gill, 2008). The auto-
identification of the smart sensor (Yurish & Gomes, 2003) unit from a sensor network is
based on the Basic TEDS that also represents part of the stored information.
Considering smart sensor architectures (Song & Lee, 2008), the communication between the
sensor processing unit (e.g. microcontroller) and one or multiple non-volatile memory units (e.g.
Flash EEPROM memory units) is done using different communication protocols, such as SPI,
I2C, 1-wire (Kalinsky & Kalinsky, 2002); (Paret & Fenger, 1997); (Linke, 2008). These protocols are
thus frequently used in smart sensor implementations (IEEE, 2004)(Ramos et al., 2004).
As the name implies, smart sensors networks are networks of smart sensors, that is, of
devices that have an inbuilt ability to sense information, process the information and send
selected information to an external receiver (including to other sensors). A "smart sensor" is
a transducer (or actuator) that provides functions beyond what is necessary to generate a
correct representation of a sensed or controlled quantity. This means that such nodes require
memory capabilities to store data temporarily or permanently.
In an increasingly number of applications, the nodes are required to change their spatial
position (mobile nodes), which leads to wireless networks. The sensor network nodes data
management and advanced data processing are carried out by a host unit characterized by
high data processing capabilities, non-volatile data storage capabilities and data
communication capabilities. One kind of solutions that materialize the host unit is mobile
devices (e.g. phones and PDAs) with special operating systems (e.g. WindowsCE, Symbian,
BalckBerry OS) and internal and extended data storage memory capabilities (CF card
memory, SD card memory). Specific protocols, CompactFlash and Secure-Digital (Compact
Flash, 2009) (SD Association, 2009) are associated with memory card interfaces that are used
to perform the communication between the host unit processor and the memory units. In
wireless sensors networks, special attention is granted to the memory read/write operations
158 Data Storage

time interval and the associated power consumption considering the mobile device
autonomy requirements.
Considering the importance of non-volatile memory and the communication protocols
associated with memory units as parts of smart sensors, sensor networks and distributed
mobile systems (e.g. wearable sensing system for physiological parameter measurements),
the proposed chapter briefly reviews some non-volatile memory solutions more used in
those contexts.

2. Non-volatile Memory, Smart Sensors and Mobile Devices

A non-volatile memory is an important part of a smart sensor. Non-volatile memory is a
general term for all forms of a solid state memory that do not need to have their memory
contents periodically refreshed. This includes all forms of read-only memory (ROM) such as
programmable read-only memory (PROM), erasable programmable read-only memory
(EPROM), electrically erasable programmable read-only memory (EEPROM), and flash
memory. It also includes the random access memory (RAM) that is powered by a battery.
Regarding smart sensors, the non-volatile memory stores a table of parameters that identify
the transducer and are held in the transducer on an EEPROM for interrogation by external
electronics. The table’s contents can be the one defined by the IEEE committee associated
with IEEE1451 standard for smart sensors (Ulivieri et al., 2009). The transducer parameters
table is also known as TEDS (Transducer Electronic Datasheet) and according with the
above mentioned standard, brings plug-and-play capabilities to transducers. TEDS-
compatible measurement systems can auto-detect and automatically configure these “smart
sensors” for measurement, reducing setup time and eliminating transcription errors that
commonly occur during sensor configuration.
Different IEEE1451 family members include the TEDS module as part of particular smart
sensor architecture implementations. Figure 1 presents the TEDS module localization
according to IEEE1451.2, IEEE1451.3, IEEE1451.4, IEEE1451.5 and IEEE1451.6. Figure 1
shows that the TEDS that is localized at the transducer’s level sends the specific transducer
information to the Network Capable Information Processor (NCAP) using different kinds of
interfaces. The smart transducers communicate with the NCAP using the mixed mode
interface that joins the analogue signal line and the memory communication lines according
to IEEE1451.4, using CAN (Controller Area Network) (Pfeiffer et al., 2003) interface according
to IEEE1451.6, using digital point-to-point according to IEEE1451.2, using distributed bus
interfaces such as I2C (Pardo et al., 2006) according to IEEE1451.3 or using wireless
interfaces such as Bluetooth (Flittner, 2007) or ZigBee (Higuera, 2009) according to
IEEE1451.5. Additionally, an IEEE1451 standard extension for smart sensors with RFID
(IEEE1451.7) is under discussion.
From the different individuals in the IEEE1451 standard family one of the more
implemented is the mixed mode transducer interfaces IEEE1451.4. The IEEE 1451.4
standard defines a mechanism for adding self-identification technology to traditional
analogue sensors and actuators (IEEE, 2009) and it is described in details in the next section.

IEEE1451.4
IEEE 1451.4 (dot 4 from now on) defines a mechanism for adding self-describing behaviour
to traditional transducers with an analogue signal interface. The dot 4 defines the concept of
Non-volatile memory interface protocols for smart sensor networks and mobile devices 159

a transducer that supplies both analogue and digital interface, namely, mixed-mode
interface (Fig.2) where the TEDS non-volatile memory localization is better highlighted. In
this case, the non-volatile memory interface will materialize the digital interface associated
with IEEE1451.4.

Analog + Digital
TEDS Mixed-Mode
IEEE P1451.4 Transducer
Txdcr
Network-Capable CANOpen
Application TEDS

Txdcr Bus
Interface
Processor IEEE P1451.6 Intrinsically safe
(NCAP) system
Any A/D Txdcr
Network
Digital, Point-to-point
TEDS Smart Transducer

Interface
Wireless
IEEE P1451.2 Interface Module
IEEE IEEE A/D Txdcr (STIM)
1451.1 1451.0
Distributed Multidrop
Common Common
Bus
Txdcr Bus
TEDS
Interface
Object Function Transducer Bus
Model ality & IEEE P1451.3
Interface Module
TEDS A/D Txdcr

Wireless
TEDS
Interface
Wireless

IEEE P1451.4 Wireless Txdcr

A/D Txdcr

Txdcr – Transducer (Sensor or Actuator)

TEDS – Transducer Electronic Data Sheet

Fig. 1. TEDS on IEEE 1451 family of smart transducer interface standards network (Txdcr –
transducer sensor or actuator)

Mixed-Mode Transducer
NCAP
(Sensor with TEDS)

IN+ Analog Signal +

A/D
Analog GND -
IN-

1-w Data
I/0 (1-wire) TEDS

Dig. GND

Fig. 2. Non-volatile memory for TEDS associated with mixed-mode interface for smart sensors
(NCAP – network capable application processor)
160 Data Storage

The memory interfaces that are mainly used in the IEEE1451.4 implementation are 1-wire
and I2C. In the following paragraph a brief description of the 1-wire and 2-wire (I2C)
memory interface protocols particularly used in smart sensor implementation will be
presented.

3. Memory Interfaces Protocols for Smart Sensors

The 1-wire interface provides advantages over other industry interfaces such as I²C or SPI™
in applications where contacts between the host controller and the memory device are
limited or must be minimized, and/or where factory-programmed unique device
serialization is required or valuable. 1-wire devices require a total of two contacts for total
device operation, compared to four contacts for I²C memory or five contacts for SPI
memory.

3.1 1-wire interface protocol for smart sensor memories

1-wire protocol was developed by Dallas Semiconductor in order to permit digital
communications over twisted-pair cables with 1-wire components over a 1-wire network.
A 1-wire bus uses only one wire for signalling and power. Communication is asynchronous
and half-duplex, and it follows a strict master-slave scheme. One or several slave devices can
be connected to the bus at the same time. Only one master should be connected to the bus.
1-wire bus electronics implements an open-drain (wired-AND) master/slave multidrop
architecture with resistor pull-up to a nominal supply at the master. A block diagram of 1-
wire network is presented in Fig. 3.

1-wire Memory
IO Infc Array

GND 64-bit ROM

VCC
VCC
VCC VCC VCC
RPUP
IN OUT IN OUT IN OUT
MASTER DATA DATA DATA
UNIT 1 UNIT 2 UNIT n
GND GND GND
GND
GND

Fig. 3. - 1-wire network of non-volatile memories

The 1-wire network has three main components: a bus master with controlling software,
wiring and associated connectors (e.g. Esensors connectors) and 1-wire devices (unit1, unit2,
unit3,..unitn) that can be sensors, actuators and memories (e.g. DS2430A from Maxim). The
Non-volatile memory interface protocols for smart sensor networks and mobile devices 161

protocol uses conventional CMOS/TTL logic levels (maximum 0.8V for logic “zero” and a
minimum 2.2V for logic “one”) with operation specified over a supply voltage range of 2.8V
to 6V. System clock is not required; each 1-wire part is self-clocked by an internal oscillator
synchronized to the falling edge of the master.
Signalling on the 1-wire bus is divided into time slots of 60us. One data bit is transmitted on
the bus per time slot. Units are allowed to have a time base that differs significantly from the
nominal time base. This however, requires the timing of the master to be very precise, to
ensure correct communication with slaves with different time bases.
Addressing
All 1-wire devices have a unique address laser-registered into the chip. Dallas
Semiconductors guarantees that the address is unique. The individual address is expressed
by a 64-bit serial number that is stored in the device memory. It is composed of eight bytes
divided into three main sections as presented in Table 1.

Family code ID CRC8

8 bits 48 bits (unique within family) 8 bits
Table 1. – 1-wire address format. Left to right: increasing time, increasing bit order

Starting with the least significant bit (LSB), the first byte stores the 8-bit family codes that
identify the device type. For the particular case of 1-wire memories the family codes are
presented in Table 2.

1-wire memory Family code

1k memory iButton 08
4k memory iButton 06
16k memory iButton 0A
64k memory iButton 0C
4k EEPROM 23
1k EEPROM 2D
256 EEPROM 14
1k EEPROM protected with SHA-1 33
Table 2. - 1 wire address family code for several memory devices

The next six bytes store a customizable 48-bit individual address or ID that guaranteed
unique within a family. A few types of chips have sequences of IDs reserved for special
manufacturing runs, but in general, there are no special characteristics to the ID. The last
byte, the most significant byte (MSB), contains a cyclic redundancy check (CRC) with a
value based on the data contained in the first seven bytes. This allows the master to
determine if an address was read without error.
With a 248 serial number pool, conflicting or duplicate node addresses on the net are never a
problem. For maximum data security the 1-wire memories can implement US government-
certified Secure Hash Algorithm (SHA-1).
162 Data Storage

Basic bus signals

As mentioned before, 1-wire memory interface protocol uses a single wire (plus ground) to
accomplish both communication and power transmission. A single bus master can feed
multiple slaves over a single twisted-pair cable. Thus, the master initiates every
communication on the bus down to the bit-level. This means that for every bit that is to be
transmitted, regardless of direction, the master has to initiate the bit transmission. This is
always done by pulling the bus low, which will synchronize the timing logic of all units.

1-wire bus commands and operations

The communication between the master and slaves uses a set of five basic commands of the
1-wire bus: “Write 1”, “Write 0”, “Read”, “Reset” and “Presence”.

 Write 1- The master pulls the bus low for 1 to 15 μs. It then releases the bus for the
rest of the time slot (Fig.4).

Fig. 4. - Write 1 – command on the 1-wire bus

 Write 0 - The master pulls the bus low for a period of at least 60 μs, with a
maximum length of 120 μs (Fig.5).

Fig. 5. - Write 0 – command on the 1-wire bus

 Read - The master pulls the bus low for 1 to 15 μs. The slave then holds the bus
low if it wants to send a ‘0’. If it wants to send a ‘1’, it simply releases the line. The
bus should be sampled 15μs after the bus was pulled low. As seen from the
master’s side, the “Read” signal is in essence a “Write 1” signal. It is the internal
state of the slave, rather than the signal itself that dictates whether it is a“Write 1”
or “Read” signal (Fig. 6).

Fig. 6. - Read command on the 1-wire bus

 Reset & Presence - The master pulls the bus low for at least 8 time slots or 480μs
and then releases it. This long low period is called the “Reset” signal. If there is a
slave present, it should then pull the bus low within 60μs after it was released by
the master and hold it low for at least 60μs. This response is called a “Presence”
signal. If no presence signal is issued on the bus, the master must assume that no
device is present on the bus, and further communication is not possible (Fig.7).
Non-volatile memory interface protocols for smart sensor networks and mobile devices 163

Reset Presence
Fig. 7 .- Reset & Presence – command on the 1-wire bus

The communication between the master and slave is performed using the 1-wire commands
according with the flowchart that is presented in Fig.8.

Reset 1-wire bus

Select 1-wire
device

Perform a single
Device operation

Fig. 8. - 1-wire master-slave communication flowchart

The first step of the communication is materialized by the „reset” command that is delivered
by the master synchronizing the entire 1-wire bus. One of the unit1, unit2, ...., unit n (slave
devices) is selected for the next communication. The selection of the specific slave is done
using the serial number of the device or using a binary search algorithm (Maxim, 2002).
Once a specific device has been selected, all other devices drop out and ignore subsequent
communications until the next reset is carried out.
Because each device type performs different functions and serves a different purpose, each
has a unique protocol once it has been selected. For the particular case of a non-volatile
memory, a set of particular commands are mentioned:
 Write Scratchpad [0Fh] - applies to the data memory and the writable addresses in
the register page. After issuing the Write Scratchpad command, the master must
first provide the 2-byte target address, followed by the data to be written to the
scratchpad;
 Read Scratchpad command [AAh] - allows verifying the target address and the
integrity of the scratchpad data. After issuing the command code, the master
begins reading. The master should read through the end of the scratchpad, after
which it receives an inverted CRC16, based on data as it was sent by the 1-wire
memory.
 Copy Scratchpad command [55h] - is used to copy data from the scratchpad to the
data memory and the writable sections of the register page.
 Read Memory command [F0h] - is the general function to read from the 1-wire
memory. After issuing the command, the master must provide a 2-byte target
164 Data Storage

address, which should be in the range of 0000h to 0A3Fh. If the target address is
higher than 0A3Fh, for the particular case of DS28EC20 (1-wire memory), the
upper four address bits are changed to “0”. After the address is transmitted, the
master reads data starting at the (modified) target address and can continue until
address 0A3Fh. If the master continues reading, the result is FFh. The Read
Memory command sequence can be ended at any point by issuing a reset pulse.
 Extended read memory [A5h]- works essentially the same way as Read Memory,
except for the 16-bit CRC that the DS28EC20 generates and transmits following the
last data byte of a memory page. The Extended Read Memory command sequence
can be ended at any point by issuing a reset pulse.

Master host computer interfacing

In a 1-wire network associated with smart sensors the master is generally a microcontroller.
An example of DS28EC20 connection to the microcontroller as part of an IEEE1451.4
implementation is presented in Fig. 9. It presents the advantage of compactness but requires
1-wire implementation protocol at the microcontroller level. 1-wire protocol implementation
for Atmel microcontroller (Atmel, 2004) and PIC microcontroller (Maxim, 2003) are referred
in the literature.
Using a serial 1-wire line driver, the 1-wire memory device can be connected to the UART
port of the microcontroller reducing the design time. A solution in this field is the MAXIM’s
DS2480B chip. It connects directly to UARTs and 5V RS232 systems. Interfacing to RS232C
(±12V levels) requires a passive clamping circuit and one 5V to ±12V level translator such as
MAX232. Internal timers relieve the host of the burden of generating the time-critical 1-wire
communication waveforms. The DS2480B can be set to communicate at four different data
rates, including 115.2kbps, 57.6kbps, and 19.2kbps, with 9.6kbps being the power-on default.
Command codes received from the host's crystal controlled UART serve as a reference to
continuously calibrate the on-chip timing generator. The various control functions of the
DS2480B are optimized for MicroLAN 1-wire networks and support the special needs
EPROM-based add-only memories and EEPROM devices as well as the other 1-wire devices
(e.g. ibutton, 1-wire thermometers).
SmartSensor
(IEEE1451.4)

AO
VCC
RPUP (300
to 2.2 k)
1-w I/O
Mem
AI (DS28EC20)

C GND

Master Slave
Fig. 9. - The interfacing of memory as part of IEEE1451.4 smart sensor to the microcontroller
(C)
Non-volatile memory interface protocols for smart sensor networks and mobile devices 165

Referring to the implementation of 1-wire software that works under Windows OS, an 1-
wire Software Development KIT (SDK) can be used to develop applications associated with
IEEE1451.4 smart sensor data management. The 1-wire SDK includes 1-wire API for .NET,
API examples in VB.NET and C#, along with TMEX API examples in C, C++, Pascal
(Borland Delphi), and Microsoft Visual Basic that assures high degree of flexibility and
reduced time of software design and implementation.

3.2 2-wire (I2C) interface protocol for smart sensor memories

In the early 1980's, Philips Semiconductors developed a simple bi-directional 2-wire bus for
developing networks of integrated circuits (IC), including memories. This bus is called the
Inter-Integrated Circuit or I2C-bus. At present, Philips’ IC range includes more than
150 CMOS and bipolar I2C-bus compatible types for performing communication functions
between intelligent control devices (e.g. microcontrollers), general-purpose circuits (e.g.
LCD drivers, remote I/O ports, memories) and application-oriented circuits (e.g. digital
tuning and signal processing circuits for radio and video systems).
Taking into account the capabilities of I2C bus on device interfacing, it was considered as an
interesting solution for smart sensors (SmartS) interfacing, including those compatibles with
the IEEE1451 protocol. An example of I2C network, including a set of memories as part of
IEEE1451.4 smart sensor network implementation, is presented in Fig. 10.

DAQ device

TR
SDA SCL TRi SDA SCL (i+1) SDA SCL TRn
unit i unit i+1 unit n
I2C Master I2C Slave I2C Slave
(e.g. microcontroller) (e.g. serial EEPROM (e.g. serial EEPROM
memory) memory)

Fig. 10. - I2C device network including 1 master device and n – slave devices ( TRi, TR(i+1),
TRn – analog transducers, DAQ device – analog signal acquisition devices)

Each device has a particular address. I2C bus supports two addressing schemes: 7-bit
address and 10-bit address. Up to 1024 devices are allowed to be connected to the bus. The
7-bit address scheme has shorter message length and requires less complex hardware.
Devices with 7 and 10 bit addresses can be mixed in the same system.
The bus can operate in three modes with different data rates. Data on the bus can be
transferred at rates of up to 100kbit/s in the standard-mode, up to 400 kbit/s in the fast-
mode, or up to 3.4 Mbit/s in the high-speed mode. The number of interfaces connected to
the bus is dependent on the bus capacitance limit of 400pf (Phillips, 2000).
166 Data Storage

Basic bus signals

The I2C bus wires are called SDA (Serial DAta line) and SCL (Serial CLock line). They are
both bi-directional and assure the communication between the network devices through the
SDA and SCL signals (Fig.11). As can be observed, the communication signals are
characterized by: the start bit, the 7 bits (B6...B0) I2C device address, the RD/nWR bit
associated with master-slave read or write operations signalling, the ACKNOWLEDGE bit,
the 8 (D7...D0) data bits, and the STOP bit.
The SCL signal is an explicit clock signal on which the communication synchronizes and is
imposed by the master unit. When the slave unit is not able to follow with the clock rate
given by the master a mechanism clock stretching is activated on the slave level (Leens,
2009).

SDA B6 B5 B4 B3 B2 B1 B0 R/W ACK

SCL
start
Comm.

D7 D6 D5 D4 D3 D2 D1 D0 ACK

stop
Comm.
Fig. 11. - The SDA and SCL signals associated with the I2C bus communication

The I2C bus does not support plug-n-play or interrupt functions that are important for
many sensor networks or memory networks. To save energy, some sensor nodes should be
in sleep mode most of the time and woken up by a timer or sensing event. For each
transported information, the microcontroller should initiate the request and provide the
clock to the sensors. To get the updated information from the sensor, the microcontroller has
to poll each sensor node connected on the bus very often to make sure it will not miss new
information or unexpected events. These features make I2C unsuitable for the applications
which have strict requirement for the power efficiency and emergency processing. Although
the sensor node could be clipped to or from the system easily, the microcontroller can not
detect an event or configure the system during operation. This limits the application of I2C
in sensor networks which are usually dynamically reconfigurable and/or demand high
energy-efficiency.
Non-volatile memory interface protocols for smart sensor networks and mobile devices 167

Commands and operations

Every device connected to the bus has its own unique address and can act as a receiver
and/or transmitter, depending on its own functionality.
The I2C bus is a multi-master bus. Thus, more than one device with I2C interface are
capable of initiating a data transfer. The I2C protocol specification states that the unit that
initiates a data transfer on the bus is considered the Bus Master. Consequently, at that time,
all the other units are regarded to be Bus Slaves. The master can send data to a slave or
receive data from a slave. Slaves do not transfer data between themselves.
Considering the particular scenario of a microcontroller as master of I2C bus that includes
memories associated with smart sensor TEDS as bus slaves, the commands and operation
associated with master-slave, slave-master communication are next presented.
First, the master will send out a START signal. This acts as an “Attention” signal sent to the
bus units. All of the units connected to the bus will listen for incoming data. Then, the
master sends the ADDRESS of the device it wants to access, along with an indication of
whether the access is a Read or Write operation (“1” logic for read operation and “0” for
write operation).
Having received the address, all ICs will compare it with their own address. If it does not
match, they simply wait until the STOP command from the master.
Considering that the received address by the slave matches its proper address, the response is
materialized by the ACKNOWLEDGE signal. Once the master unit receives the
ACKNOWLEDGE, it can start transmitting or receiving DATA according to the RD/nWR bit.
When all the data transmission is done the master will send the STOP command. This is the
signal that the bus has been released and that the connected units may expect another
transmission to start any moment.
Referring to memories as I2C bus units, there are several important manufactures of very
broad range of memory capacities with I2C bus interface. As common used memories in smart
sensor implementation can be mentioned the Microchip memories, that are characterized by
capacities from 16 bytes, for the 24C00, up to 32k bytes, for the 24C256 family.

4. Memory Interfaces Protocols for Mobile Devices

Mobile devices must have memories that assure the capability to store applications and data
files. However, data logging tasks are always related to the utilization of memory extension
cards, such as CompactFlash memories (CF card) or Secure Digital memories (SD card).
Considering the importance of mobile devices external memories, two of the
communications protocols associated with this kind of memories are now described.

4.1 Compact Flash memory protocol

CompactFlash® (CF) has established itself as a dominant flash memory card technology in
applications where small form factor, low-power dissipation and ease-of-design are crucial
considerations. Introduced by SanDisk Corporation in 1994, CF Storage Cards provide the
capability to transfer all types of digital information and software between different types of
digital systems and are compatible with future CF designs, eliminating interoperability
restrictions. As a result, CF is a powerful solution that can be found in digital SLR cameras,
PDAs (e.g. iPAQ 2700 series), embedded systems, single-board computers and data recorders.
168 Data Storage

CF based data storage is used especially at the master device level such as PDAs
(Postolache, 2006) or touch panel computers (e.g. TPC2106T) (Postolache, 2007).

Two types of CF non-volatile memory cards are considered:

 Type I cards or CF, dimensionally characterized by 43mm (1.7") x 36mm (1.4")
x 3.3mm (0.13"));
 Type II cards or CF2, dimensionally characterized by 43mm (1.7") x 36mm (1.4")
x 5mm (0.19").
CompactFlash cards are designed with flash technology [(Kingmax Digital Inc., 2009)].
According to the CompactFlash card specification version 4.1 it can be characterized by data
storage capacities up to 137GB. However, the current commercial CF solutions have data
storage capacities up to only 64GB (e.g. Pretec CF memory).
The CF classification by speed is:
 original CF;
 CF High Speed (using CF+/CF2.0);
 CF3.0 standard.
Compact Flash supports data rates up to 133MB/sec. However, values associated with
current commercially CF devices are up to only 45Mb/s (e.g. 8GB ScanDisk Extreme).

CF electrical interface
The internal configuration of the CF protocol associated to CF non-volatile memory cards
includes a CF controller connected through I/O digital lines to a host interface that can be
associated with PDAs, laptop computers, tablet PC, etc. The CF block diagram and the CF
connector pin-out are presented in Fig. 12.
Non-volatile memory interface protocols for smart sensor networks and mobile devices 169

Data
In/Out
Storage
Host Module(s)
Controller
Interface
(Flash, Disk
Control
Drive, etc.)

CFast Storage Card

26 50
CF connector
1 25
Fig. 12. - Compact Flash block diagram and CF connector pin-out

Pins 13 and 38 correspond to power supply, which according to the standard can be either
3.3 volts or 5 volts. Pins 1 and 50 correspond to GND. The data stored on the CF can be
accessed through 8 or 16 bit data bus associated to the CF connector. The data and address
bits for 8 and 16 bits memory access bus are shown in Table 3.

Pins 21 22 23 2 3 4 5 6
Data D00 D01 D02 D03 D04 D05 D06 D07
lines
Pins 47 48 49 27 28 29 30 31
Data D08 D09 D10 D11 D12 D13 D14 D15
lines
Table 3. - CF pins and data line correspondence for 8bit and 16 bit data bus for memory access.

The address bus (A0 to A10 lines) pin assignment is given in Fig. 13.
170 Data Storage

A9 A7 A5 A3 A1

26 50
CF
connector
1 25

A10 A8 A6 A4 A2 A0
Fig. 13. - CF’s connector address bus pin out

4.2 Secure digital (SD) memory protocol

The SD card standard is a standard for removable memory storage designed and licensed by
the SD Card Association (Compact Flash, 2009). It is the result of the collaboration between
three manufacturers, Toshiba, SanDisk, and MEI (Panasonic) and is specifically designed to
meet the security, capacity, performance, and environment requirements inherent in newly
emerging audio and video consumer electronic devices as well as an important component
of smart sensing networks that include smart mobile devices such as smart phones and
PDAs, assuring data logging capabilities for individual nodes.
At the same time, the SD standard is not limited to removable memory storage devices and
has been adapted to many different classes of devices, including 802.11 cards, Bluetooth
devices, and modems.
The SD Memory Card includes a content protection mechanism that complies with the
security of the SDMI (Secure Digital Music Initiative) standard and is faster and capable of
higher memory capacity. Nowadays, SD high-capacity (SDHC) cards start at 4GB and going
up to 32GB.
In what concerns the operating rate of the SD card working in the default mode, typical
values are up to 12.5 MB/sec interface speed for a variable clock rate 0 - 25 MHz, while for
high-speed working mode are reported values of up to 25 MB/sec interface speed for a
variable clock rate from 0 - 50 MHz.

SD electrical interface
A block diagram of an SD card interface associated with a SD card non-volatile memory is
presented in Fig. 14 and the corresponding pin-out, together with the SD function are
presented in Table 4.
The SD card is clocked by an internal clock generator. The interface driver unit synchronizes
the DAT and CMD signals from external CLK to the internal used clock signal. The card is
controlled by the six line SD card interface containing the signals: CMD, CLK,DAT0~DAT3.
Non-volatile memory interface protocols for smart sensor networks and mobile devices 171

For the identification of the SD card in a stack of SD card, a card identification register (CID)
and a relative and address register (RCA) is foreseen [(Kingmax Digital Inc., 2009)].
An additional register, (CSD), contains different types of operation parameter.

VDD
DAT2 CMD CLK DAT0
CD/DAT3 Interface driver DAT1
Internal Clock

OCR [31:0]
CID [127:0]
Card

Power on detection
RCA [15:0]
Interface
DSR [15:0] controller
CSD [127:0]
SCR [630]

Memory interface

Memory core

Fig. 14. - SD card block diagram

Pin Name SD function

1 DATA3/CS Data Line 3

2 CMD/DI Command Line
3 VSS1 Ground
4 VDD Supply Voltage
5 CLK Clock (SCK)
6 VSS2 Ground
7 DAT0/D9 Data Line 0
8 DAT1/IRQ Data Line 1
9 DAT2/NC Data Line 2
Table 4. – SD pin out
172 Data Storage

The card has its own power on detected unit. No additional master reset signal is required
to setup the card after power on. It is protected against short circuit during insertion and
removal while the SD card system is powered up
The communication using the SD card lines to access either the memory field or the register
is defined by the SD card standard. Different protocols are supported by SD cards.

SD 1 bit protocol
It is a synchronous serial protocol with one data line, used for bulk data transfers, one clock
line for synchronization, and one command line, used for sending command frames. The SD
1-bit protocol explicitly supports bus sharing. A simple single-master arbitration scheme
allows multiple SD cards to share a single clock and DAT0 line.

SD 4 bit protocol
It is nearly identical to the SD 1-bit protocol. The main difference is that bulk data transfers
use a 4-bit parallel bus instead of a single wire. With proper design, this has the potential to
quadruple the throughput for bulk data transfers. Both the SD 1-bit and 4-bit protocols by
default require a Cyclic Redundancy Check (CRC) protection of bulk data transfers.
Cyclic Redundancy Check is a simple method for detecting the presence of simple bit-
inversion errors in a transmitted block of data. In SD 4-bit mode, the input data is
multiplexed over the four bus (DAT) lines and the 16-bit CRC is calculated independently
for each of the four lines, which imply an increased software complexity of CRC calculation.
Hardware implementation of 4-bit parallel CRC calculation represents an interesting
alternative and can be materialized using an Application Specific Integrated Circuit (ASIC)
or field programmable gate arrays (FPGA).

SPI mode – SD card protocol

It is distinct from the 1-bit and 4-bit protocols in that the protocol operates over a generic
and well-known bus interface, Serial Peripheral Interface (SPI). SPI is a synchronous serial
protocol that is extremely popular for interfacing peripheral devices with microcontrollers
(Leens, 2009). Most modern microcontrollers, including the MSP430, support SPI natively at
relatively high data rates. The SPI communications mode supports only a subset of the full
SD Card protocol. However, most of the unsupported command sets are simply not needed
in SPI mode. A fully-functional SD Card implementation can be realized using only SPI.
This flexibility of electrical interfaces is a significant advantage to a designer. A designer
may opt for a fast parallel interface, or depending on the application, may prefer a slower
implementation using SPI. Due to the popularity of the SPI protocol and its efficient
implementation on the MSP430, we mention here only the SPI mode of the SD Card
protocol. Minor differences in the initialization sequence exist between the two major
modes. For more information, the interested reader should consult the full SD Card
Specification available from the SD Card Association.

5. References
Atmel (2004). AVR318: Dallas 1-Wire master 8-bit Microcontrollers, on-line at:
www.atmel.com/dyn/resources/prod_documents/doc2579.pdf, Sept-2004
Non-volatile memory interface protocols for smart sensor networks and mobile devices 173

Brewer, J. & Gill, M. (2008). Nonvolatile Memory Technologies with Emphasis on Flash: A
Comprehensive Guide to Understanding and Using Flash Memory Devices, Wiley-IEEE
Press, February 2008
Compact Flash Org. (2009). CF+ & CompactFlash Specification Revision 4.1, on line at
http://www.compactflash.org/spec_download.htm, May 2009
Flittner P. (2007). CSR (Bluetooth Subgroup Chair) Thurston Brooks, 3eTI, IEEE P1451.5
Wireless Sensor Interface Working Group Bluetooth Subgroup. Proposal on-line at:
grouper.ieee.org/groups/1451/5/.../P1451.5_Bluetooth2.pdf
Frank, R. (2002). Understanding Smart Sensors, Artech House Publishers, April 2002
Higuera J.; Polo J.; Gasulla M. (2009). A Zigbee wireless sensor network compliant with the
IEEE1451 standard, Proceedings of IEEE Sensors Applications Symposium, SAS 2009,
pp.309-313, 2009
Honeywell (2009). “TEDS - plug and play sensor configuration”, on-line at
http://www.sensotec.com/pnpterms.shtml, June 2009
IEEE Std 1451.4-2004 (2004), Standard for a Smart Transducer Interface for Sensors and
Actuators- Mixed-Mode Communication Protocols and Transducer Electronic Data
Sheet (TEDS) Formats, IEEE Standards Association, Piscataway, NJ, sub classe 5.1.1,
2004
IEEE STD 1451.0-2007 (2007). Standard for a Smart Transducer Interface for Sensors and
Actuators – Common Functions, Communication Protocols, and Transducer
Electronic Data Sheet (TEDS) Formats, IEEE Instrumentation and Measurement
Society, TC-9, The Institute of Electrical and Electronics Engineers, Inc., New York,
N.Y. 10016, SH99684, October 5, 2007
IEEE Standard Association (2009). IEEE Standard 1451.4-2004 Tutorials, on-line at:
http://standards.ieee.org/regauth/1451/Tutorials.html, 2009
Kalinsky D. & Kalinsky R. (2002). Introduction to Serial Peripheral Interface, on-line at
http://www.embedded.com/story/OEG20020124S0116, January ‘02
Kingmax Digital Inc. (2009). SD Card Specification” on line at::
downloads.amilda.org/MODs/SDCard/SD.pdf
Leens F. (2009). An introduction to I2C and SPI protocols, Instrumentation and Measurement
Magazine, vol.12, No. 1, pp. 8-13, February 2009
Linke B. (2008). Overview of 1-Wire® Technology and Its Use, on-line at
http://www.maxim-ic.com/appnotes.cfm/an_pk/1796, June 2008
Maxim Inc. (2002), 1-Wire Search Algorithm, on-line at http://www.maxim
ic.com/appnotes.cfm/an_pk/187, March 2002
Maxim Inc. (2003), 1-Wire Communication with a Microchip PICmicro Microcontroller on-line at
http://www.maximic.com/appnotes.cfm/an_pk/2420, Sept. 2003
Pardo A.; Camara L.; Cabre J.; Perera A.; Cano X.; Marco S.; Bosch J. (2006). Gas
measurement systems based on IEEE1451.2 standard, Sensors and Actuators B:
Chemical, Vol.116, Issues 1-2, pp. 11-16, July 2006
Paret D.& Fenger, C. (1997). The I2C Bus: From Theory to Practice, Wiley, John & Sons, 1997
Pfeiffer O.; Ayre A.; Keydel C. (2003) Embedded Networking with CAN and CANopen,
Annabooks Publisher, 2003
Phillips Semiconductor (2000). The I2C-BUS specification, Version 2.1, 2000, online at:
i2c2p.twibright.com/spec/i2c.pdf. 2000
174 Data Storage

Postolache O.; Postolache G.; Girão P. M. (2006). Non-invasive mobile homeostasis

Instrument", Proc IEEE International Workshop on Medical Measurements and
Applications, Benevento, Italy,.Vol. I, pp. 94 – 98, May 2006
Postolache O.; Dias Pereira J. M.; Girão P. M. (2007). Optical Fiber Based Turbidity Sensing
System, Proc IMEKO TC19 Symp, Iasi, Romania, Vol. II, pp. 71 - 75, September, 2007
Ramos H.; Pereira J.M.; Viegas V.; Postolache O.; Girão P. M. (2004). A Virtual Instrument
to Test Smart Transducer Interface Modules (STIMs), IEEE Transactions on
Instrumentation and Measurement, vol. 53, no. 4, pp. August 2004
SD Association (2009). SD Card Compact and Powerful, Ideal for Small Products, on-line at:
http://www.sdcard.org/developers/tech/sdcard/, May 2009
Song E. Y. & Lee K (2008). Understanding IEEE 1451-Networked smart transducer interface
standard, IEEE Instrumentation & Measurement Magazine, Vol.11, No. 2, pp.11-17,
February 2008
Ulivieri N.; Distante C.; Luca T.; Rocchi S.; Siciliano P. (2006). IEEE1451.4: A way to
standardize gas sensor, Sensors and Actuators B: Chemical, Volume 114, Issue 1, pp.
141-151, March 2006
Yurish, S. & Gomes M. (2003). Smart Sensors and MEMS, NATO Science Series Mathematics,
Physics and Chemistry, September 2003
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 175

11
X

Electronic Nose System and Artificial Intelligent

Techniques for Gases Identification
Iman Morsi
Arab Academy for Science and Technology,
Electronics and Communications Department
Alexandria-Egypt

1. Introduction
Electronic nose is the intelligent design to identify food flavors, cosmetics and different gas
odors, depending on sensors. The continuous developing of these sensors permit advanced
control of air quality, as well as, high sensitivity to chemical odors. Accordingly, a group of
scientists have worked on developing the properties of sensors, while others have modified
ways of manufacturing ultra low-cost design (Josphine & Subramanian , 2008); (Wilson et
al., 2001).
In the design of an electronic nose, sampling, filtering and sensors module, signal
transducers, data preprocessing, feature extraction and feature classification are applied.
(Getino et al., 1995) is used as an integrated sensor array for gas analysis in combustion
atmosphere in the presence of humidity and variation in temperature from 150-350oC. The
sensor array exposed to a gas mixture formed by N2, O2, CO2, H2S , HCL and water vapour
with a constant flow rate of 500 ml/ min was studied. (Marco et al., 1998). The gas
identification with tin oxide sensor array is investigated, in addition, the several undesirable
characteristics such as slow response, non-linearties, long term drifts are studied. Correction
of the sensor’s drift with adaptive self organizing maps permit success in gas classification
problems.(Wilson et al., 2001) is introduced as a review of three commonly used gas sensors
which are, solid state gas sensor, chemical sensors and optical sensors. Comparisons are
deducted among them in terms of their ability to operate at low power, small size and
relatively low cost with numerous interference and variable ambient conditions.(Dong Lee
& Sik Lee, 2001) depended on solid state gas sensor, thus the pollutants of environment are
controlled relative to the sensing mechanism, the sensing properties of solid – state gas
sensors to environmental gases, such as No, Co and volatile organic compounds.(Guardado
et al. , 2001) is used as a neural network efficiency for the detection of incipient fault in
power transformers. The NN was trained according to five diagnosis criteria and then tested
by using a new set of data.
This study shows that NN rate of successful diagnosis is dependent on specific criterion
under consideration with values in the range of 87-100 %.( Zylka & Mazurek, 2002)
introduced a rapid analysis of gases by means of a portable analyzer fitted with
176 Data Storage

electrochemical gas sensor. The analyzer, which was built, is controlled by a microprocessor
and the system incorporates only two gases which are Co and H2.
The drawback is the lack of sensors selectivity which is disadvantageous in most
applications.(Belhovari et al., 2004; Belhovari et al., 2005) used sensors array with gas
identification and Gaussian mixture models. Some problems are studied such as drift
problem and slow response is introduced. Robust detection is applied through a drift
counteraction approach which is based on extending the training data set using a simulated
drift. (Belhovari et al., 2006) gas identification is introduced using sensors array, and
different neural networks algorithms. Different classifiers are used MLP, RBF, KNN, GMM
and PPCA are compared with each other using the same gas data set allowed performance
up to 97%. Electronic gas sensors based on tin oxide films are used for the identification of
gases, detection of toxic contaminants and separation of mixture of gases (Kolen, 1994);
(Belhovari et al., 2005); (Marco et al., 1998); (Getino et al., 1995); (Becker et al., 2000);
(Amigoni et al., 2006).
The problem here is to identify or to discriminate different gases such as, methane, propane,
butane, carbon dioxide and hydrogen. Using different concentrations. Taguchi gas sensors
(TGS) are used, which is a metal oxide semiconductor sensor, based on tin oxide that has
been commercially available from Figaro engineering company (Figaro sensor.com “on –
line”). In the design of electronic nose systems, power consumption directly related to
temperature operation, selectivity, sensitivity, and stability typically has the most influence
on the choice of metal oxide films for a particular application (Fleming, 2001); (Carullo,
2006). For electronic nose applications (Morsi-b, 2008); (Luo et al, 2002); (Bourgeois et al.,
2003); (Carullo, 2006) metal oxide semiconductors are largely hampered by their power
consumption demands. Thermal isolation and intermittent operation of the heaters reduce
the power consumption of the sensors themselves to facilitate their use of importable
applications. However, it also presents significant obstacles in terms of noise, drift, aging,
and sensitivity to environmental parameters. The Feed Forward Back Propagation of Neural
Network using the multilayer perceptron is used to separate between them. Fuzzy logic is
used to discriminate different gases and to detect the concentration of each gas. Electronic
nose design provides rapid responses, ease of operation and sufficient detection limits. Data
quality objectives (DQOs) of gases must be considered as a part of technology development
and a focus should be made on the most urgent problems.

Organization of the chapter

1. Introduction.
2. Experimental setup.
3. Results and analysis.
4. Surface response modeling algorithms and analysis of variance (ANOVA).
5. Separation of butane and propane as a gas mixture using an artificial intelligent
technique of Neural Networks.
6. On-line identification of gases using artificial intelligent techniques of Fuzzy Logic.
7. Conclusion.
8. Acknowledgement.
9. References.
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 177

2. Experimental Setup
The analysis and the characterization of gases are acquired by building a prototype of multi-
sensors monitoring system (electronic nose), which are TGS 822, TGS 3870, TGS 4160 and
TGS 2600, from Figaro sensor industry, temperature sensor, humidity sensor and supply
voltage equal to 5 V. However, current monitoring methods are costly and time intensive,
likewise limitations and analytical techniques exist. Clearly, a need exists for accurate,
inexpensive, long-term monitoring using sensors. TGS 822 is tin dioxide (Sno2) as
semiconductor. The sensor’s conductivity increases, depending on the gas concentration. It
has high sensitivity to the vapors of organic solvents, as well as, combustible gases. TGS
2600 is comprised of metal oxide semiconductor layer formed on an alumina substrate of a
sensing chip together with an integrated heater. TGS 3870 is a metal oxide semiconductor
gas sensor, embedding a micro-bead gas sensing structure. TGS 4160 is a hybrid sensor unit
composed of a carbon dioxide sensitive element and a thermistor. All presented sensors
have features of long life, low cost, small size, simple electrical circuit, low power
consumption and are available in commercial application. The experimental equipment
consists basically of the gas bottle mass flow controllers, sensor chamber with volume 475
cm3, supply voltage of 5 V and a heating system (Morsi, 2007); (Morsi-a, 2008).Gases used in
the experiment are carbon dioxide, hydrogen, methane, propane, butane and a mixture of
propane and butane. All measurements are presented at 45% relative humidity. All sensors
are connected as an array and covered by a chamber, which has “in” and “outlet” ports. The
input is connected with the mass flow controllers to control the concentration of input gas
after purging with humidified air. All sensors are subjected to variation in temperatures
from ambient temperatures and up (Clifford et al., 2005); (Fleming, 2001). Four variable
resistances are connected in series to the four sensors, placed out of chamber, then are
followed by the microcontroller, to control and monitor the output of each sensor (Smulko,
2006). The output of the microcontroller is monitored and recorded every 20 sec. Different
gases concentrations are applied 100 ppm, 400 ppm, 700 ppm, 1000 ppm with different
environmental temperatures between 20oC to 50oC with different variable resistances for
each sensor RL = 1 k, 3 k, 5 k, and 7 k . Variable load resistance is used to control the
conductivity and to increase the selectivity of each gas than other gases. Sensitivity is used
to refer either to the lowest level of chemical concentration that can be detected or to the
smallest increment of concentration that can be detected in the sensing environment. While
selectivity refers to the ratio of the sensor’s ability to detect what is of interest over the
sensor’s ability to detect what is not of interest as the interferents. Sensors for use in
electronic nose need partial selectivity, mimicking the responses of the olfactory receptors in
the biological nose (Belhovari et al., 2006). Figure (1) shows the electronic nose gas system.
The hardware requirements for the system implementation include a microcontroller Pic 16
F 877A with embedded (A/D) converter. It is chosen for the implementation of this task due
to the on chip memory resources, as well as, its high speed. The output data is transferred to
a PC via a serial port RS 232 with a Baud rate of 2400 from the microcontroller. The software
is developed in C language and is complied, assembled, and downloaded to the system. The
output volt of each sensor is collected, stored in memory and transferred to a
microcontroller to be ready for the processing work and the temperature is also monitored
via a temperature sensor and is recorded (Ishida et al., 2005); (Weigong et al.,
2006);(Smulko,2006).
178 Data Storage

Fig. 1-a. Electronic nose of gas detection Fig. 1-b. Electronic nose of gas detection
with chamber without chamber

Measurements using the electronic nose gas system detector have been done according to
three parts: measurement part, mathematical analysis part, and presentation part. The
system is supported by a collection of methods to improve the uncertainty and reliability.
Different processing techniques like self calibration, self validation and statistical analysis
methods are included. Data averaging standard deviation calculation are used to test and
evaluate the performance of the whole measuring system in order to minimize error (Morsi,
2007); (Morsi – a, 2008).

3. Results and Analysis:

The design of electronic nose depends on physical connectivity of the sensors, relating to the
data management, computing management and knowledge discovery management, which
are associated with the sensors and the data they generate, and how they can be addressed
within an open computing environment. Eventually, the issue relates to the integrated
analysis of the sensor data depending on the variation of gases respectively. Moreover, there
is a correlation and interaction of data.
Hence, the use of standardized data access and integration techniques to asses and integrate
such data is essential. Furthermore, if the analysis is to proceed over large data sets, it is
essential to provide high performance computing resources to allow rapid computation to
proceed. Measurements have been done using the described experimental setup. The gas is
injected inside the chamber and the concentration in ppm is controlled through the mass
flow controller. Extensive measurements are performed and data collected from gas sensor
via microcontroller undergoes a processing stage. Due to variation in temperature from
20oC to 50oC, the load resistance is adjusted to 1 k , 3 k, 5 k and 7 k, respectively and
the gas concentrations are adjusted to 100 ppm, 400 ppm, 700 ppm and 1000 ppm
respectively. There is an output voltage of each sensor relative to the variation of these
parameters. The extensive numbers of measured values have been extracted at each
resistance with different concentrations. The measurement values of different parameters
with different gases at RL = 7 k are described by figs 2 till 13 as a specimen deducted from
enormous measured data due to different variable resistances with different concentrations
(Morsi –a , 2008) .
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 179

Fig. 3. Methane gas with TGS 822 at RL 7

Fig. 2. Methane gas with TGS 3870 at RL
k
7 k

Fig. 4. Methane gas with TGS 2600 at RL 7 Fig. 5. Methane gas with TGS 4160 at RL 7
k k

Fig. 6. Hydrogen gas with TGS 3870 at RL 7 Fig. 7. Hydrogen gas with TGS 822 at RL 7
k k

Fig. 8. Hydrogen gas with TGS 2600 at RL 7k Fig. 9. Hydrogen gas with TGS 4160 at RL 7
k
180 Data Storage

Figures 14, 15 and 16 show resistances fluctuations with different sensors, different gases
and constant concentration at 700 ppm (Morsi – a, 2008).

Fig. 10. Carbon dioxide gas with TGS 3870 Fig. 11. Carbon dioxide gas with TGS 822 at RL 7
at RL 7 k k

Fig. 12. Carbon dioxide gas with TGS 2600 Fig. 13. Carbon dioxide gas with TGS 4160 at
at RL 7 k RL 7 k

Fig. 14. Methane gas at concentration 700 Fig. 15. Hydrogen gas at Concentration 700 ppm
ppm
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 181

Fig. 16. Carbon dioxide gas at concentration 700 ppm

The Electronic nose system used for gas detection depends on the resistance variation of the
gas sensor, which given the possibility increases the selectivity and sensibility. The portable
and cheap electronic nose is based on commercially available gas sensors. The stability of
the sensors, as well as, their sensitivity depends on the ratio between the change in sensor
resistance in the presence or absence of the gas, which was experimentally characterized.
Moreover, the method can reduce the number of gas sensors to limit power consumption
and maintenance costs. From figs 14 , 15 and 16 it can be noticed that, by increasing load
resistance from 1 k to 7 k, the output volt of TGS 822, TGS 2600 is directly proportional
(i.e. increase with load resistance), However, TGS 3870 is inversely proportional by
increasing load resistance incase of hydrogen and carbon dioxide. TGS 4160 remains
constant with average output volt in the range (0.2-0.4) V incase of hydrogen, but incase of
carbon dioxide TGS 4160 gives variation in the output voltage from (0.4 - 1.8) V. This
variation is inversely proportional to the load resistance variation, which concludes that this
sensor is preferable to detect carbon dioxide than other gases.
It can also illustrate the sensitivity of each sensor due to the resistance fluctuation. The
output results include non-linear response, drift and slow response time. The main
problem is the drift, which causes significant temporal variations of the sensor response
when exposed to the same gas under identical conditions (Clifford et al., 2005);
(Fleming, 2001). It is noticed that response times depend on many parameters, such as
the material type, operating temperature, thickness of the semiconductor, variable
resistance, humidity as well as gas concentration. The sensor array reacts slowly and
takes an average of l0 min to reach the stationary state. This time is a combination of the
time to fill the chamber and the sensor response time. To achieve robust and fast
identification of combustion gases with an array of sensors, a recent study suggested
three main methods for reducing response time:
1. Increasing operating temperature.
2. Reducing the film thickness.
3. Using variable resistance to reduce the number of sensors and the power
consumption of the system.
182 Data Storage

4. Surface Response Modeling Algorithms and Analysis of Variance (ANOVA)

In the system design, surface response models investigated are very useful in understanding
the functional relationships between a set of independent factors.
The relationship can be expressed as a mathematical equation which describes the response
of the independent factors and a set of parameters. The surface response models depend on
observing input and output values of the actual and predicted parameters, which is known
by the system model identification (SMI). The resulting empirical models of SMI can be used
to: (1) Predict the system behaviour over a specified designed space, (2) Illustrate the
mathematical relationships and interactions between input and output of the system (3)
Optimize the system performance (4) Select the appropriate factor values that satisfy system
performance goals to investigate the performance prediction accuracy of different
algorithms, thus the predicted results generated by each empirical model is compared with
actual measured results to provide a comparison and identify the advantages of each
algorithm.
In the present work, different empirical models are used to detect the response of the surface
of each gas. By entering the voltage of each sensor, load resistance and temperature as an
input thus predicts the concentration as an output. The following are the used four
empirical models:
Linear

Y = b0 + b1 x1 + b2 x2 + b3 x3. (1)
Interaction

Y= b0 + b1 x1+ b2 x2+ b3 x3+ b12 x1 x2+ b13 x1 x3+ b23 x2 x3. (2)
Pure Quadratic

Y= b0 + b1 x1 + b2 x 2 + b3 x3 + b11 x12 + b22 x22 + b33 x33. (3)

Full Quadratic

Y = b0 + b1 x1 + b2 x2 + b3 x3 + b12 x1 x2 + b13 x1 x3 + b23 x2 x3 + b11 x12 + b22 x 22 + b33 x23. (4)

Where
Y is the predicting concentration of each gas.
x1, x2, x3 are voltages of each sensor, load resistance, and temperature, respectively.
Bij is the effect of element i on element j or the effect of first input on the second input, i and j
have values from 1 to 3.
More than 300 readings for each input are used to detect the error. The concentrations of each
gas are stored in matrices, which are related to the output voltage of each sensor through a
regression coefficient matrix and the equations can be solved using surface response
empirical models to predict the concentrations. Then, the percentage error can be calculated
between actual concentration and predicted concentration to determine the best empirical
modeling algorithm, which describes the surface response of each gas and has the least error.
Table1 shows the percentage error for the different gases with different surface response
empirical modeling algorithms (Morsi – a, 2008).
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 183

Percentage
Percentage

Percentage
Gases

Hydrogen

Methane
Dioxide
Error of

Error of

Error of
Carbon
The

The

The
Empirical Modeling
Algorithms
Sensor Type TGS 822
Linear 21.322% 13.745% 20.048%
Interactions 19.613% 0.2439% 21.175%
Pure Quadratic 17.7612% 11.2664% 17.9546%
Full Quadratic 14.2644% 0.096.029% 14.426%
Sensor Type TGS 2600
Linear 18% 9.2% 2.2%
Interactions 11% 7.7% 5.5%
Pure Quadratic 30.5% 12.95% 5.8%
Full Quadratic 8% 5.5% 5.7%
Sensor Type TGS 3870
Linear 6.41% 1.66% 3.47%
Interactions 3.03% 1.1% 0.98%
Pure Quadratic 2.08% 2.3% 0.28%
Full Quadratic 1.68% 0.641% 0.02%
Sensor Type TGS 4160

Linear 6.6534% 49.925% 51.572%

Interactions 8.1904% 53.159% 51.253%
Pure Quadratic 8.9625% 45.646% 44.386%
Full Quadratic 7.0078% 40.353% 44.358%
Table 1. Comparisons of different empirical modeling algorithms with different sensors
and different gases.

It can be noticed that, for the three different gases, four different algorithms are used. Thus,
full quadratic algorithms can predict the concentration of each gas with less error than other
algorithms. From the percentage error, it is clear that in the case of TGS 822, TGS 2600 and
TGS 3870 gas sensors provide the least error in the case of methane than hydrogen. It is
preferred to use this sensor to detect hydrogen and methane. However, it is unpreferable in
the case of carbon dioxide. It is also noticed that in TGS 4160 gas sensor is preferable to be
carbon dioxide, where as the highest errors are recorded in the case of hydrogen and
methane.
184 Data Storage

From the above results, it can be concluded that the surface response modeling algorithms
provide accurate detection for different concentrations of gases depending on solving
regression matrices using different equations while detecting the percentage error between
the actual and the predicted measurements. The key challenges for building regression
algorithms determine the significant factors that are included in the final mathematical
equation and quantify the effects of those factors. Tables 2, 3 and 4 show the ANOVA results
for each sensor with different gases. The second column is the degree of freedom (DF). The
mean square (MS) is the variance of the data for each factor interaction. The sum of squares
(SS) is computed as MS x DF. F-statistic is defined as the ratio of MS to Error. P-value is the
smallest probability of rejecting the null hypothesis. Using the analysis of variance
(ANOVA), it is possible to identify those effects that are statistically significant. It can be
noticed that, the variance is not constant and if the output voltage has a high value, the
variance also has a high value for different gases. The resulting algorithm includes only
those independent factors that are statistically significant (P-value < 0.05). Quantifying the
main and two-way interaction effects of the independent factors are equivalent to using the
well- known method of least squares fitting method in order to compute the regression
coefficients (Morsi-a, 2008) .

Table 2-a. Anova of methane gas with TGS 3870 Table 2-c. Anova of methane gas with TGS 2600
gas sensor gas sensor

Table 2-b. Anova of methane gas with TGS 822 Table 2-d. Anova of methane gas with TGS 4160
gas sensor gas sensor
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 185

Table 3-a. Anova of hydrogen gas with TGS Table 4-a. Anova of carbon dioxide gas with TGS
3870 gas sensor 3870 gas sensor

Table 3-b. Anova of hydrogen gas with TGS Table 4-b. Anova of carbon dioxide gas with TGS
822 gas sensor 822 gas sensor

Table 3-c. Anova of hydrogen gas with TGS Table 4-c. Anova of carbon dioxide gas with
2600 gas sensor TGS 2600 gas sensor

Table .3-d. Anova of hydrogen gas with TGS Table 4-d. Anova of carbon dioxide gas with TGS
4160 gas sensor 4160 gas sensor
186 Data Storage

5. Separation of Butane and Propane as a Gas Mixture Using an Artificial

Intelligent Neural Network
The use of natural gas has been increasingly adapted as butane and propane exist in both
industry and as a fuel in domestic applications. One of the most important problems in the
gas detection field is that there is a strong demand to detect butane and propane as pure
gases which are used as fuel; however, both are extracted from the natural gas mixed with
each other. The calibration of both gases in the pure case and also as a mixture at different
temperatures using electronic nose system is studied. Also, a study is done for the
efficiency of feed forward back propagation neural network for the detection of gases using
the multilayer pereception (MLP) methods to differentiate between them depending on the
data driven from the electronic nose gas system (Morsi – a, 2008).
Butane (C4H10): a normally gaseous straight-chain or branch-chain hydrocarbon is derived
from natural gas or from refinery gas streams. It includes isobutane and normal butane and
is used for cooking, heating, as a household fuel, propellant or refrigerant, where, it is
commonly used in UK. It is slightly toxic by inhalation, as it causes central nervous system
depression at high concentrations, and is used in the manufacture of rubber and fuels
(Morsi –a, 2008).
Propane (C3H8): a normally gaseous straight- chain hydrocarbon is a colorless paraffinic gas
that boils at a temperature of – 42 oC. It is extracted from natural gas or refinery gas streams.
Under normal atmospheric pressure and temperature, it exists in a gaseous state. However,
propane is usually liquified through pressurization for transportation and storage. It is
primarily used for heating or cooking, as a fuel gas in areas not serviced by natural gas, and
as a petrochemical fuel stock. Propane is used in forklifts because it offers the best
performance in power, speed and saves money on fuel costs. It is also used in agriculture as
an alternative to conventional pesticides and herbicides and in outdoor grills
(Morsi –a, 2008).
Neural Networks with feed forward back propagation training algorithm is used to detect
each gas with different sensors. It is combined with gas sensors to address these issues.
After a processing stage, the resulting feature vector is used to solve separation problem in
case of a mixture, by identifying an unknown sample as one from a set of previously known
gases. Multi Layer Perceptron MLP improves linear discrimination techniques in case of a
mixture, depending on the data driven from each sensor (Morsi –a, 2008); (Guardado et al.,
2001).
Gases used are butane and propane and their mixtures. They are injected into a gas
chamber. The sensor detects sequentially the variation in the propane and butane gas
concentration and its resistance is modified accordingly. Figures 17 till 32 depict the output
voltage of sensors TGS 822, TGS 3870, TGS 2600 and TGS 4160 as functions of temperature,
with load resistances RL = 1 k and RL = 7 k. The relations are given for butane and
propane at concentrations of 100 ppm, 400 ppm, 700 ppm and 1000 ppm respectively. It can
be noticed that the detection of both gases can be carried out by TGS 822, TGS 3870, TGS
2600, but not by TGS 4160. It is clear noticed from figs 17-20, that for the TGS 822 gas sensor,
the conductivity and sensitivity increase for both butane and propane by increasing
concentration and load resistance. The sensitivity is directly proportional to the load
resistance. For TGS 3870 gas sensor, Figs 21-24, it can be noticed that the sensitivity and
conductivity are inversely proportional with load resistance.
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 187

By increasing load resistance, the sensitivity and conductivity will decrease. The difference
between the relationship of sensitivity, conductivity and load resistance between both
sensors, allows the discrimination between propane and butane depending on the variation
of the resistance. For the TGS 2600 gas sensor, figs 25 till 28, the voltage increases with the
increase of concentrations, temperature and load resistance for both gases.
With TGS 4160 gas sensor, figs 29-32, for both butane and propane, there is no change in the
output volt in load resistance, therefore, we can not depend on this sensor discrimination.
The calibration of pure gases among their semiconductor sensor, predicts the correct sensor
that should be used in classification. Figures 33 and 34 depict the mixture of both propane
and butane injected inside the chamber. Propane has a concentration of 600 ppm while
butane has a concentration of 400 ppm. It can be noticed that the output of both gases is
unstable, which leads to difficulties in discrimination. Neural Networks (NN) have been
used extensively in applications where pattern recognition is needed. They are adaptive,
capable of handling highly non-linear relationships and generalizing solutions for a new set
of data. In fact, NN do not need a predefined correspondence function, which means that
there is no need for a physical model. A neuron model is the most basic information
processing unit in a neural network. Depending on the problem complexity, they are
organized in three or more layers: the input layer, the output layer and one or several
hidden layers. Each neuron model receives input signals, which are multiplied by synaptic
weights. An activation function transforms these signals into an output signal to the next
neuron model. The Back Propagation learning algorithm is used due to its ability of pattern
recognition. A sigmoid activation function was also used because of two reasons: it is highly
non-linear and has been reported to have a good performance when working with the back
propagation learning algorithms [45][46] . In order to avoid slow training, it is decided to
use only three layers. During the training process, a vector from a training set (xi)
representing a gas pattern is presented to the net. The winning neuron (the closet to the
pattern with an Euclidean), and its neighbours, the neighborhood area, change their
position, becoming closer to the input pattern according to the following learning rule:
Wjinew = Wjiold +  (t). nb (t,d). (xi – xjiold) (5)
where Wji are the weights of the neurons inside, the neighborhood area, j is the index of the
neuron, i is the index of pattern , t is the time step,  (t) is the learning rate, nb (t,d) is the
neighborhood function, and d is the distance between the neuron and the winner measured
over the net. The learning rate and the neighborhood are monotonically decreasing
functions along the training (Eberhart et al. 1996); (Wesley, 1997). Both patterns constitute an
NN training set. In case NN training is slow or shows little convergence, then both patterns
are either poorly correlated or incorrect. This study is performed on pure butane and
propane gases, and a mixture of them. The data set is divided in two parts. The first is used
to train the net, while the second is used to test it. Training patterns are chosen from
different concentrations, different times and different temperatures. A large number of
patterns have been selected from the extremes of the concentrated range and from initial
parts of the dynamic response to give a larger weight to more early difficult cases. The
Neural Network is constructed as a feed forwarded back propagation network that is
composed of three layers: input, hidden and output layers. The input layer has three
neurons corresponding to the output voltages of each sensor (x1), temperature (x2) and
variable resistance (x3).
188 Data Storage

The hidden layer has five neurons. The hidden layer neurons use a transfer function of
tansig which is a hyperbolic tangent sigmoid function used to calculate the layer’s output
from its net input. One hidden layer with 5 neurons is used which gives the least mean
square error (MSE) between the actual and the predicted data. The output layer has one
neuron corresponding to the concentration of gas. The predicted performance metric, y,
given by the neural network model is as follows.

 5  3  
y  Purelin   w 2i1, Tansig   w x j  θ1j   θ 21  (6)
 i 1  j1 1ji  
   
Where: xj is the input of node j in the input layer.
W1ji is the weight between node j in the input layer to node i in the hidden layer. θ1i is the
bias of node i in the hidden layer. W2i1 is the weight between node i in the hidden layer to
the node in the output layer. θ21 is the bias of the node in the output layer. The numbers 5
and 4 are the numbers of nodes in the hidden layer and in the input layer, respectively,
using a simple linear transformation. All performance data are scaled to provide values
between –1 and 1. Scaling is performed for two reasons: to provide commensurate data
ranges, so that the regressions are not dominated by any variable that happened to be
expressed in large number, and to avoid the asymptotes of the sigmoid function. During the
training, the weights of the neural network are iteratively adjusted to minimize the network
performance function MSE. The validation set is used to stop training early if the network
performance on the validation set fails to improve. Test set is used to provide an
independent assessment of the model predictive ability. The percentage error is important;
100% error means a zero prediction accuracy and error close to 0% means an increasing
prediction accuracy. For the proposed Neural Network model, the percentage error is found
to be 0.662%, 0.031%, 0.162%, 1.5% for sensors TGS 822, TGS 3870, TGS 2600 and TGS 4160,
respectively. MLP provides an optimized structure which provides linear discrimination
between both gases. Figs 35, 36, 37, and 38 depict the MLP results in separating butane and
propane gas by using TGS 822, TGS 3870 , TGS 2600 and TGS 4160 sensors respectively. The
sign circle indicates butane gas where as the sign plus indicates propane gas. Table 5 depicts
the results of classification for different gases among the different semiconductor sensors.
From the presented results, TGS 3870, TGS 2600 and TGS 822 gas sensor can discriminate
both gases rather than TGS 4160 which does not give different response with different
conditions. This conclusion is obtained from the multiplayer perception of Neural Network.
The Neural Network model does not yield a mathematical equation that can be
manipulated. However, its strength lies in its ability to accurately predict system
performance over the entire design space and its ability to compensate for the inherent
information inadequacy by requiring large and well spread training sets. Neural Network
with feed forward back propagation is used to detect the concentration of each gas. MLP is
able to separate between the mixtures with linear discrimination. TGS 3870 gives the
optimum classification with a percentage error of 0.031%, then, TGS 2600 gas sensor can be
classified between them with a percentage error 0.162%, then, TGS 822 gas sensor gives
percentage error of classification 0.662%. However, TGS 4160 gas sensor failed to
discriminate both gases and gives a percentage error of 40%.
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 189

Neural Network with MLP method is very robust against sensor nonlinearities, time effects,
and error rate depending on the selection of data, which is acquired by different
semiconductor gas sensors (Morsi –a, 2008) .

Fig. 17. Butane with TGS 822 at RL 1 k Fig. 18. Butane with TGS 822 at RL 7 k

Fig. 19. Propane with TGS 822 at RL 1 k Fig. 20. Propane with TGS 822 at RL 7 k

Fig. 21. Butane with TGS 3870 at RL 1 Fig. 22. Butane with TGS 3870 at RL 7 k

Fig. 23. Propane with TGS 3870 at RL 1k Fig. 24. Propane with TGS 3870 at RL 7 k
190 Data Storage

Fig. 23. Propane with TGS 3870 at RL Fig. 24. Propane with TGS 3870 at RL

Fig. 25. Butane with TGS 2600 at RL Fig .26. Butane with TGS 2600 at RL 7
k k

Fig. 27. Propane with TGS 2600 at RL Fig. 28. Propane with TGS 2600 at RL

Fig. 29. Butane with TGS 4160 at RL Fig. 30. Butane with TGS 4160 at RL
k

Fig. 31. Propane with TGS 4160 at RL 1 Fig. 32. Propane with TGS 4160 at RL 7
k k
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 191

Fig. 33. Mixture of propane and butane at RL 1 Fig. 34. Mixture of propane and butane at RL 7
k k

Vectors to be classified
P(2)

Fig. 35. Separation between propane and butane using NN(MLP) with TGS 822

Vectors to be classified
P(2)

P(1)

Fig. 36. Separation between propane and butane using NN (MLP) with TGS 3870
192 Data Storage

P(2) Vectors to be classified

P(1)

Fig. 37. Separation between propane and butane using NN (MLP) with TGS 2600

Vectors to be classified
P(2)

P(1)

Fig. 38. Separation between propane and butane using NN (MLP) With TGS 4160

%
Classificatio
Sensors No. of inputs No. of output Test Epoch % Error Unflassificatio
n
n

TGS 822 3 (100 sample) 1 (100 sample) 30 sample 1000 0.662% 60% 40%
TGS 3870 3 (100 sample) 1 (100 sample) 30 sample 1000 0.031% 97% 3%
TGS 2600 3 (100 sample) 1 (100 sample) 30 sample 1000 0.162% 96% 14%
TGS 4160 3 (100 sample) 1 (100 sample) 30 sample 1000 1.5% 40% 60%

Table. 5. The results of classification by using Neural Networks (MLP) for four different
sensors .
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 193

6. On- Line Identification of Gases Using Artificial Intelligent Technique of

Fuzzy Logic:
Fuzzy set theory and fuzzy logic theory are the preferred choices for the models. Fuzzy sets
introduced were particularly designed to mathematically represent fuzziness and
vagueness, and to provide the fundamental concept for handling the imprecision intrinsic to
the problems of subjective evaluation and measurement. Fuzzy set is based on possibility
instead of probability where as fuzzy logic is based on fuzzy set. Fuzzy logic is unlike
classical logical systems in that it aims to modeling the imprecise modes of reasoning that
plays an essential role in the remarkable human ability to make rational decisions in gases of
uncertainty and imprecision. This ability depends, in turn, on the ability to infer an
approximate answer to a question based on a store of knowledge that is inexact, incomplete,
or not totally reliable. The whole approach is based on measurements taken from an
experimental set up with certain typical commercial sensors. The outputs of sensors are
monitored by a microcontroller, and then a proper intelligent processing. Fuzzy logic has
been chosen because it gives better results and enhances discrimination techniques among
sensed gases. Fuzzy logic systems encode human reasoning to make decisions and control
dynamical systems. Fuzzy logic comprises fuzzy sets which are methods of representing
nonstatistical uncertainty and approximate reasoning, including the operations used to
make inferences. It is a tool for mapping the input features to the output, based on data in
the form of “IF – Then” rules. An implementation of a fuzzy expert system depends on
Mamdani type fuzzy controller as shown in fig. 39 (Tanaka, 1996); (Timothy, 1995); (Wesley,
1997). The objective of the controller is to discriminate different gases and to detect the
concentration of each gas according to the input variables as shown in fig. 40. There are five
steps to construct a Mamdani type fuzzy controller:

Fig. 39. Mamdani type fuzzy controller

Step 1: Identify and name the input linguistic variables, the output linguistic variables and
their numerical ranges. There are three input variables which are temperature, output
voltage of the microcontroller, and the variable resistance related to each sensor. The output
variable is the concentration of each gas. There are five identified ranges for each variable
(Morsi , 2007)
194 Data Storage

Temperature (oC) Output volt of microcontroller (V)

Linguistic Range Linguistic Range
Low 20 < T < 30 V. Low 0.5 < V < 1.5
Moderate 25< T < 35 Low 1<V<2
Medium 30 < T < 40 Medium 1. 5 < V < 2.5
High 35 < T < 45 High 2<V<3
V. High 40 < T < 50 V. High 2.5 < V < 3.5
Variable Resistance (K) The concentration of each gas in (ppm)
Linguistic Range Linguistic Range
V. Low 1 < RL < 3 V. Low 100 <concentration < 400
Low 2 < RL < 4 Low 250 <concentration < 550
Medium 3 < RL < 5 Medium 400 <concentration < 700
High 4 < RL < 6 High 550 <concentration < 850
V. High 5 < RL < 7 V. High 700 <concentration <
1000

Step 2: Define a set of fuzzy membership functions for each of the input and the output,
variables. The low and high values are used to define a triangular membership function. The
height of each function is one and the function bounds do not exceed the high and low
ranges listed above for each range. The membership functions must cover the dynamic
ranges related to the minimum and maximum values of inputs and outputs that represent
the universe of discourse.

Step 3: Construct the rule base that will govern the controller’s operation. The rule base is
represented as a matrix of input and output variables. At each matrix row different input
variable ranges with one of the output variable range. All rules are activated and fired in
parallel whether they are relevant or not and the duplicate ones are removed to conserve
computing time. Each rule base is defined by ANDing together with the inputs to produce
each individual output response. For example, If temperature is low AND, if voltage is low
AND if, RL is low THEN concentration is low.

Step 4: The control actions will be combined to form the excited interface. The most common
rules combination method is the centroid defuzzification to get the crisp output value. This
step is a repeated process, after all adjustments are made, which allows the fuzzy expert
system to be able to discriminate and classify the data set patterns of the different gases.
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 195

Fig. 40. Fuzzy membership functions with three input variables (Temperature, variable
resistance, voltage of each sensor) and output variable (concentration).
196 Data Storage

Mamdani type fuzzy controller is used to construct the rules, which are extracted from the
data driven from the microcontroller. It can be used to discriminate and classify the data set
patterns for different gases according to the variation in different parameters, such as gas
concentrations, variation in sensor’s resistance and output voltage of microcontroller at
different temperatures and to improve the sensor’s selectivity for gas identification.
Due to the abundant number of membership function figures, the results are limited to
representing the fuzzy logic output surface for each sensor. Fuzzy logic gives on line
prediction of the concentration depending on the behavior of each gas with different sensors
which are extracted from the experiments. The feature of each gas is detected, based on the
fuzzy system. The input and output surface of the fuzzy inference system is illustrated in
figs 41-60 (Morsi, 2007).

Fig (41) The behavior of TGS Fig (42) The behavior of TGS Fig (46) The behavior of TGS Fig (47) The behavior of TGS
822 822 2600 2600
output volt with Butane gas output volt with Carbon output volt with Butane gas output volt with Carbon dioxide

Fig (43) The behavior of Fig (44) The behavior of

TGS 822 output volt with TGS 822 output volt with Fig (48) The behavior of Fig (49) The behavior of
TGS 2600 output volt with TGS 2600 output volt with
Hydrogen gas Methane gas
with different Hydrogen gas Methane gas
with different with different
concentration with different
concentration concentration concentration

Fig (50) The

behavior of
TGS 2600
Fig (45) The behavior of TGS 822 output volt with With Propane gas with
Propane gas with different concentration different concentration
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 197

7. Conclusion
The large scale of data is able to provide high-level information to make decisions about
each gas achieved by using efficient and affordable sensor systems that show autonomous
and intelligent capabilities. Measurements have been done using an electronic nose design,
depending on commercially available gas sensors. An enormous data collected allows
analyzing and characterizing of different gases. Several tests have been carried out by
eliminating from the complete time series, a subset of data in a random way to evaluate the
reliability of the developed model. The data set is characterized by several problems which
can generate errors during processing and prediction phases inducing:
 Missing data, caused by the periodic setting or the stop function of
instruments.
 Incorrect recording data occurred by errors in transmission, recording and
non-setting of equipment.
198 Data Storage

Some interpolation and validation techniques to address and solve the above problems
depend on building a mathematical model based on a suitable regression analysis and
interpolation, which is able to describe the behaviour of daily average concentration. The
aim of modeling for gases studies is to describe the peculiar characteristics of gases as alarm
situations or risk events. Electronic nose provides, low cost, low maintenance small size, and
in some cases low power consumption. The system can handle problems such as sensor
drift, noise and non-linearity. It is used as a simple alarm level based on index data. This
could be used as a stepping stone for the development of more complex systems, which are
needed for more demanding applications.

8. Acknowledgement:
I would like to express my sincere gratefulness and gratitude to Prof. Dr. Moustafa Hussein
Professor of Optical Communications and OSA member. Vice Dean for education and
research affairs, College of Engineering and Technology in Arab Academy for Science,
Technology and Maritime Transport, e-mail [email protected], for his tolerance in revising
the chapter as well as his endless guidance and support.

9. References:
Amigoni, F.; Brandolini, A.; Caglioti, V.; Lecce, V.; Guerriero, A.; Lazzaroni, M.; Lombardi,
F.; Ottoboni, R.; Pasero, E.; Piuri, V.; Scotti, O. & Somenzi, D. (2006). Agencies for
perception in environmental monitoring. IEEE Transactions on Instrumentation and
Measurement, Vol. 55. No.4. (august 2006), pp. (1038 – 1050), DOI 10.1109/ TIM,
2006. 877747.
Becker, I.; Mihlberger, S.; Bosch, C.; Braunmuhl, V., Muller, G.; Meckes, A. & Benecke, W.
(2000). Gas mixture analysis using Micro – reactor systems. IEEE Journal of micro
electromechanical systems, Vol.9, No.4, (december 2000), pp. (478 – 484), DOI S1057 –
7157 (00) 10867-4.
Bourgeois, W.; Romain, A.; Nicolas, J. & Stuetz, R. (2003). The use of sensors arrays for
environmental monitoring: interests and limitations. J. Environmental, Monitoring,
Vol.5, (october 2003), pp (852 – 860), ISSN: 10. 1039 / b 307905h.
Belhovari, S.; Bermak, A.; Wei, G. & Chan, P. (2004), Gas identification algorithms for
microelectronic gas sensor, Proceeding of instrumentation and Measurement
Technology, pp. 584 – 587, ISBN 0-7803 – 8248 Como, (may 2004), IMTC 2004, Italy.
Belhovari, S.; Shi, M.; Bermark, A. & Chan, P. (2005). Fast and robust gas identification
system using an integrated gas sensor technology and gaussian mixture models.
IEEE Sensors Journal, Vol.5, No.6, (december – 2005), pp. (1433 – 1444), DOI
510.1109/ JSEN, 2005. 858926.
Belhovari. S.; Shi, M.; Bermark, A. & Chan, P. (2006). Gas identification based on committee
machine for micro electronic gas sensor. IEEE Transactions on Instrumentation and
measurement, Vol. 55, No.5, (october 2006), pp (1786 – 1793), DOI S10.1109/TIM.
2006. 880956.
Clifford, K.H.; Robinson, A.; David, R. & Mary, J. (2005) Overview of sensors and needs for
environmental monitoring. Sensors, Vol. 5, No. 28, (february 2005), pp. (4-37), ISSN
1424 – 8220.
Electronic Nose System and Artificial Intelligent Techniques for Gases Identification 199

Carullo, A. (2006). Metrological management of large – scale measuring systems. IEEE

Transactions on Instrumentation and Measurement, Vol. 55, No.2, (april 2006), pp. (471-
476), DOI 10.1109/ TIM. 2006. 870125.
Dong Lee, D. & Sik Lee, D. (2001). Environmental gas sensor. IEEE Sensors Journal, Vol.1,
No.3, (october 2001), pp. (214 – 217), ISSN 1530 – 437 x (01) 09470-2.
Eberhart, R.; Simpson, P. & Dobbings, R. (1996). Computational Intelligence PC Tools,
APRROFESSIONAL, ISBN 0-12-228630-8, USA>
Fleming, W. (2001). Overview of automotive sensors. IEEE Sensors Journal, Vol. 1, No.4,
(december 2001), pp. (296 – 308), ISSN S1530-437X (01) 11158-9.
Getino, J.; Gutierrez, J.; Ares, L.; Robla, J.; Sayago, C.; Horrillo & Gap, J. (1995). Integrated
sensor array for gas analysis in combustion atmosphere, Proceeding on solid- state
sensors and Actuators, pp. 803 – 806, Stockholm, (june 1995), Transducers 95 Euro
sensors 1x, Sweden.
Guardado J.; Naredo, J.; Moreno, P. & Fuerte, C. (2001). A Comparative Study of neural
network efficiency in power transformers diagnosis using dissolved gas analysis,
IEEE Transactions on Power Delivery, Vol. 16, No.4, (october 2001), pp. 643-647, ISSN
0885 – 8977 (01) 08497-7.
Ishida, H.; Nakayama, G., Nakamoto, T. & Morizumi, T. (2005). Controlling a gas/odor
plume tracking robot based on transient responses of gas sensors. IEEE Sensors
Journal, Vol.5, No, 3. (june 2005), pp. (537 – 545), DOI 10.1109 / ISEN, 2051 – 839597.
Josphine, B. & Subramanian, V. (2008). Electrinic noses sniff success. IEEE Spectrum, Vol.45,
No.3, (march 2008), pp. (47 – 53), ISSN 0018 – 9235.
Kolen, P. (1994). Self- calibration / compensation technique for microcontroller based
sensors arrays. IEEE Transactions on Instrumentation and measurement, Vol. 43, No.4,
(august 1994), pp. (620 – 623), ISSN 0018 – 9456 / 94.
Luo, R.; Chen, C. & Luo, Y. (2002). Multisensor fusion and integration: approaches,
applications and future research directions. IEEE Sensors Journal. Vol. 2, No.2, (april
2002), pp. 107-119. ISSN S 1530 – 437x (02) 03941 – 6.
Marco, S.; Ortega, A.; Pardo, A. & Samitier, J. (1998). Gas identification with tin oxide sensor
array and self organizing maps: adaptive correction of sensor drifts. IEEE
Transactions on Instrumentation and Measurement, Vol. 47, No.1, (february 1998), pp.
(316-321), ISSN 0018 – 9456..
Morsi, I. (2007). A microcontroller based on multisensors data fusion and artificial intelligent
technique for gas identification. Proceeding of IEEE Industrial Electronics Society, pp
2203 – 2208, ISBN 1-4244 – 0783 – 4107, Taipei, (november 2007), industrial
electronics society, Taiwan.
Morsi-a, I. (2008). Discrimination of some atmospheric gases using an integrated sensor
array, surface response modeling algorithms and analysis of variance (ANOVA),
Proceeding of IEEE Sensors and Application Symposium (SAS), pp. 140 – 145, ISBN 978
– 1 – 4244 – 1963 – 0108, Atlanta, (february 2008), IEEE Sensors, USA.
Morsi-b, I. (2008). Discrimination between butane and propane in a gas mixture using
Semiconductor gas sensors and neural networks, Proceeding of IEEE Sensors
Application Symposium, pp. 134-139, ISBN 978 – 1- 4244 – 1963 – 0108, Atlanta,
(february 2008), IEEE Sensors, USA.
Morsi – b, I. (2008). Electronic noses for monitoring environmental pollution and building
regression model, Proceeding of IEEE Industrial Electronics Society, pp. 1730 – 1735.
200 Data Storage

ISBN 978 – 1 – 4244 – 1766 – 7108 Orlando, (november 2008), IEEE industrial
electronics society, Florida.
http://www.figarosensor.com (on line).
Smulko, J. (2006). The measurement setup for gas detection by resistance fluctuations of gas
sensors, Proceeding on Instrumentation and Measurement Technology, pp. 2028 – 2031,
ISBN 0-7803, 9360, 0106, Sorrento, (april 2006), IEEE instrumentation and
measurement society, Italy.
Timothy, J. (1995). Fuzzy Logic with Engineering Applications, McGraw- Hill INC, ISBN 0-07-
053917-0, New York.
Tanaka, K. (1996). An Introduction to Fuzzy Logic for Practical Applications, Springer, ISBN 0-
387-94804-4, New York.
Wesley, J. (1997). Fuzzy and Neural Approaches in Engineering, John Wiley & Sons INC, ISBN
0-471-19247-3, 1, USA.
Wilson, D.; Hoyt, S.; Janata, H.; Booksh, K. & Obando, L. (2001). Chemical sensors for
portable handheld field instruments. IEEE Sensors Journal, Vol.1, No.4 (december
2001), pp. (256-274), ISSN 1530 – 437X (01) 11120 – 6.
Weigong, J.; Chen, Q., Reu, M.; Liu, N. & Daoust, C. (2006). Temperature feedback control
for improving the stability of a semiconductor metal oxide (SMO) gas sensor, IEEE
Sensor Journal, Vol. 6, No.1, (february 2006), pp. 139 – 145, ISSN 10.1109/ JSEN,
2005. 844353.
Zylka, P. & Mazurek, B. (2002). Rapid dissolved gas analysis by means of electrochemical
gas sensor, Proceeding of 14th International conference on Dielectric Liquids, pp. 325 –
328, ISBN 0-7803 – 7350 – 2102, Graz, (july 2002), ICDL 2002, Austria.
Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 201

12
X

Medical Image Intelligent Access Integrated with

Electronic Medical Records System for Brain
Degenerative Disease
Mei-Ju Sua, Po-Hsun Chengb, Sao-Jie Chena, Chung-Yi Yangd,
Ping-Kung Yipf, Daniel Racoceanugh and Heng-Shuen Chence*
aGraduateInstitute of Electronics Engineering, College of Electrical Engineering and
Computer
Science, National Taiwan University
Department of Software Engineering, National Kaohsiung Normal University
bDepartment of Medical Informatics, College of Medicine, National Taiwan University
cDepartment of Family Medicine, National Taiwan University Hospital
dDepartment of Medical imaging, National Taiwan University Hospital
eInformation System Office, National Taiwan University Hospital
fDepartment of Neurology, National Taiwan University Hospital
gIPAL-UMI CNRS 2955, French National Research Center , I2R, A*STAR, NUS,

Singapore
hFaculty of Sciences, University of Besancon, France

*Email: [email protected]
National Taiwan University, Taipei, Taiwan

Introduction
As the computer data storage capacity increases and technology of digital imaging
progresses rapidly, today we can access and manipulate massive image database on the
Internet. How to search and access intelligently on content based image database become a
prominent focus in the field of multimedia research. In this chapter, to support the
diagnostic decision making, a new idea of grid-distributed, contextual, representation, use
of specific visual medical knowledge, intelligent information access framework for medical
images database was integrated with radiology reports and clinical department information.
It will assist reducing the human, legal and financing consequences of these medical errors.
A report by the American Hospital Association suggests that US hospitals use only
1.5%~2.5% of their Hospital Medical budgets on data information systems, which is less
than the 5-10% of similar funds dedicated to such systems found in the budgets of other
industries. Moreover, as computer data storage capacity increases and the technology
involving digital imaging progresses rapidly, the traditional image has been replaced by
Computer Radiography (CR) or Digital Radiography (DR) derived imagery, Computed
Tomography (CT), Magnetic Resonance Imaging (MRI), and Digital Subtraction
202 Data Storage

Angiography (DSA). This advance reduces the space and the cost associated with X-ray
films, speeds up hospital management procedures improving also the quality of the medical
imagery-based diagnosis quality. American hospitals, in particular, often consult image
diagnostic professionals from India via Internet due to budget consideration. The clinical
information gathered by this study will alleviate the budget issue because the
tele-consultation regarding image diagnostics will then be feasible.
For the current medical image database system, image retrieval can be performed by using
keywords for searching in pre-interpreted reports. However, the characteristics of free-text
reports may compromise with the effort for clarification from low level features such as
color, shape, texture, to medium level features describing by collection and spatial
relationships, or highest abstract features which are semantic or contextual. In 1990s, the
content-based image retrieval (CBIR) started [1]and grew steadily over the last ten years. [2]
There is growing interest in CBIR because of the limitations inherent in metadata-based
systems. Textual information about images can be easily searched using existing technology,
but requires humans to personally describe every image in the database. However, CBIR
query uses a lot of computer power to process, and it can economize a large manpower.
Therefore, the content-based retrieval of the medical image also becomes a significant field
for medical assistance, medical education and medical research. At clinics, when a physician
diagnoses a patient through the patient’s medical images, he has to entry the PACS and RIS.
And he makes a diagnosis in another system, EMR (Electronic Medical Records). [3]
In Taiwan, the National Taiwan University Hospital (NTUH), which is traditionally the
leading medical center, has recently adopted a complete health recording system, including
an Electronic Medical Record (EMR) system linked to Hospital Information System (HIS),
Research and Information System (RIS), Picture Archiving and Communication System
(PACS), and other clinical information systems. National Taiwan University (NTU), one of
the research partners of ONtology and COntext related MEdical image Distributed
Intelligent Access (ONCO-MEDIA) project1 is leading a study of the integration between
the Content-Based Medical Image Retrieval (CBIR) and EMR systems as well as on its
implications in providing improved clinical diagnostic and therapeutic decision support.
Dementia is the loss of mental functions -- such as thinking, memory, and reasoning -- that
is severe enough to interfere with a person's daily functioning. Dementia is not a disease
itself, but rather a group of symptoms that are caused by various diseases or conditions.
Symptoms can also include changes in personality, mood, and behavior. In some cases, the
dementia can be treated and cured because the cause is treatable. Dementia develops when
the parts of the brain that are involved with learning, memory, decision-making, and
language are affected by one or more of a variety of infections or diseases. The most
common cause of dementia is Alzheimer's disease, but there are as many as 50 other known
causes. Most of these causes are very rare. [5] Dementia is a word for a group of symptoms
caused by disorders that affect the brain. People with dementia may not be able to think
well enough to do normal activities, such as getting dressed or eating. They may lose their
ability to solve problems or control their emotions. Their personalities may change. They
may become agitated or see things that are not there. [6] Dementia has become more and
more prevalent in recent years. In the United States, there are approximately 5 million
people suffering from dementia and that number is projected to rise above 16 million by the
Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 203

year of 2050. Presently, Americans pay US$5000 per patient per year for dementia
medication and associated nursing care costs [9] [10].
In Taiwan, there were approximately 140 thousand people suffering from dementia in 2005
and there will be 650 thousand dementia cases by the year of 2050. MMSE (mini-mental
state examination) is commonly used in medicine to screen for dementia.The MMSE test is a
brief 30-point questionnaire test that is used to screen for cognitive impairment. It is also
used to estimate the severity of cognitive impairment at a given point in time and to follow
the course of cognitive changes in an individual over time, thus making it an effective way
to document an individual's response to treatment.
Since Dementia, as a disease, is an important and long term problem that causes significant
burdens for families and societies, this study has endeavored to find a viable procedure to
ameliorate the treatment of dementia patients and to enhance the early diagnosis and
monitoring of its progression. [11]
In this chapter, we chose Dementia to establish a system for CBIR with EMR. Dementia is a
neurological disease, usually the clinical course is long, and it represents a variety of
characteristics in brain image such as CT (computed tomography), MRI (magnetic resonance
imaging) or PET (Positron Emission Tomography). If the doctor wants to diagnosis a
symptom, he needs a series of images to diagnose or decide for therapeutic strategies.
Therefore, the image database correlated with clinical information would be crucial in care
of a dementia patient. In addition, usually it is an intensive collaboration among
neurologists, radiologists and other clinical specialties. This study focused on Dementia as a
pathology model, in order to elaborate a prototyping system for CBIR with EMR. Dementia
is a neurological disease, usually characterized by a slow histopathology, and presents itself
with a variety of characteristic abnormalities in brain imagery such as CT, MRI, or PET. In
the course of the treatment, a doctor may need a series of images to make the proper
diagnosis or to make critical decision for therapeutic strategies. Therefore, an image
database infused with clinical information could become a major component for the
improvement of the dementia patient care. The chapter will describe a novel concept of
Medical Image Distributed Intelligent Access Integrated with Electronic Medical Records is
expected to enhance the early diagnosis and monitoring of disease progress.

2. Methods
Integration of RIS, EMR, PACS and Clinics to Support Diagnosis
At present, most hospitals store the medical images from CT, MRI, DSA and X-ray film in
PACS. The clinicians make differential diagnosis of a patient in EMR system with references
to laboratory results and image reports. Therefore, we have to provide the essential
information from EMR, PACS and RIS to clinicians, such as neurologists, to support their
decision. On the other hand, in the department of medical imaging, the radiologists also
need to refer to medical information recorded by other specialties to interpret medical image
for their reports on the RIS [7]. The image report by the radiologists could assist the
clinicians to make correct diagnoses; however, the correct image interpretation also depends
on the crucial medical information that clinical doctors must input. This co-dependence
demonstrates the need for two-way communications between imaging professionals such as
radiologists and their clinical counterparts that are treating the patients. Therefore, in this
study, the integration of EMR, RIS and PACS and clinics input was implemented to
204 Data Storage

establish a prototype model for intelligent access of medical image database, and retrieve
clinical information automatically. (Fig.1.) The integrated user interface used the ICD-9
(International Classification of Disease, Ninth Revision) 331.0 code to identify and query its
relative medical information and medical images from HIS, PACS or RIS. Therefore, the
important thing is to define what the essential clinical information is.

Fig. 1. System Structure of Prototype Model

Inter-disciplinary collaboration
This important process was the collaboration of the interdisciplinary among neurology,
medical informatics and radiology experts. In this study, a prototype model was designed to
assist the procedure involving the physician-in-the-loop approach [4].In this prototype
model, a group of neurologists and radiologists collaborated to establish a common
language involving the image diagnosis. In the EMR integration with RIS and PACS, we
based on the diagnostic criteria and practical guideline to satisfy the diagnosis procedure
and requirement. Moreover, the establishment of the ontology of dementia was in
accordance with the essential clinical information of dementia. The process flow chart was
as follow:
Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 205

Fig. 2. Research flow chart

Establishing a Definition of Essential Clinical Information in Dementia

The Literature review is essential course before experts to define the essential clinical
information. Therefore, journals, textbooks and clinical reports will be our guidelines. The
UMLS [12] is a long-term project realized from 1986 by Library National of Medicine (LNM),
the largest world's library in the biomedical domain of Maryland. This system is the result
of the combination of 140 multilingual terminology databases, knowledge sources as well as
tools in order to facilitate the tasks of accessing, researching, integrating and aggregating of
biomedical and health information. The technical purpose of this system is to combine into a
unified structure of different knowledge sources in the biomedical domain.[13] The purpose
of NLM's Unified Medical Language System (UMLS) is to facilitate the development of
computer systems that behave as if they "understand" the meaning of the language of
biomedicine and health. Bio-medical information as a result of a large number of constantly
increasing and are scattered in various database systems, so you want to search complete
and no new information for easy, UMLS Thus came into being for the purpose of enhancing
the capacity of the system, allowing the system to understand the readers biomedical
aspects of meaning, and thus help readers information retrieval and integration. Therefore,
we also have to depend on UMLS to support us to develop the dementia model. And the
methods and steps, the experts discussed as follow:
Literature review from journals, textbooks and clinical reports
Selecting relevant medical image reports from department of medical imaging
Using representative cases to simulate medical image distributed intelligent access
integrated with EMR to select crucial clinical information
Finalizing structure and items of essential clinical information regarding dementia
Dependence on the Unified Medical Language System (UMLS) to facilitate the development
of a prototype model that follows the international language of biomedicine and health.
206 Data Storage

3. Results
Essential Clinical information
After the inter-disciplinary collaboration, we got the final definition of essential clinical
information in dementia when the neurologist diagnoses in a clinic. In the model, the
essential clinical information of dementia included is summarized as follows:
A. Base information a. Sex
b. Age
c. Country
d. Residency
e. Education (yr)
f. Occupation (pre-retired work)
g. Language (dialect)
B. Clinical history a. Handedness
b. Age of onset
c. Initial symptom sequence (multi-choice):i. memory, ii.
personality, iii .language, iv. gait and v. bradykinesia
d. Course: i. rapid progression (< 1 year), ii. chronic
progression, iii. stepwise and iv. fluctuated
e. Risk factors (multi-choice): i.CVA, ii.HTN, iii.DM, iv. cardiac
disease, v. hyperlipidemia, vi. obesity, vii. physical
inactivity, viii. vegetarian, ix. parkinsonism and x. family
history
C. Clinical a. Normal
diagnosis b. MCI (Mild Cognitive Impairment)
c. AD (Alzeimer Disease)
d. VaD (Vascular Dementia)
e. Mixed type
f. FTD (Frontotemporal Dementia)
g. DLB (Dementia with Lewy Bodies)
h. Dementia, other types
D. Lab a. CBC
b. Electrolyte
c. BUN/Cre
d. GOT/GPT
e. T4/TSH
f. B12/folate
g. Lipid profile
h. VDRL
i. Hachinski ischemic score
j. MMSE
k. CDR
The prototype system is designed to recognize a positive dementia diagnosis and
reconfigure its patient information presentation accordingly. International Statistical
Classification of Diseases and Related Health Problems (ICD) provides codes to classify
diseases and a wide variety of signs, symptoms, abnormal findings, complaints, social
circumstances and external causes of injury or disease.[14] The ICD-9(International
Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 207

Classification of Disease, Ninth Revision) was published by the WHO in 1977.[15] In EMR of
National Taiwan University, we use ICD-9. Each code of ICD-9 codes has a precise meaning,
and as a whole it covers practically all known symptoms. ICD-9 codes 290-319 are mental
disorders. And ICD-9 of dementia is 290. ICD-9 codes 320-359 are Diseases of the nervous
system and code 331.0 is Alzheimer’s disease. In diagnosis flow with the model of the
research, an input of the ICD-9 331.0 code for Alzheimer’s disease will alter and display the
patient information, such as: base information, clinical history, lab results and medical
images. (Fig. 3.)

Neurology clinic
Research case

Before a positive After a positive

diagnosis of dementia diagnosis of dementia

ICD 331.0
ICD 290.XX

Open base information,

Open base information, Clinical history, Clinical
Clinical history diagnosis and

Set search condition

Ye
Dementia
suspected?
Open relative images
No and reports to support
RIS 、HIS
PACS
End
Clinical input

N
Diagnose
fi i h d?

Yes

End

Fig. 3. Flow Diagram

On the other hand, if dementia has not been yet, the neurologists would still like to know
the patient’s base information, clinical history, and lab results. Using this clinical history, the
208 Data Storage

neurologists may make a choice reflecting the initial symptom sequence, course, and risk
factors, which will also support him to make an early diagnosis in the future. (Fig. 3.)
With the established prototype model, the data schema of the essential clinical information
of dementia will be integrated into the EMR user interface in the clinical department in the
EMR system of NTUH. Clinical doctors can input the essential clinical information, such as
base information, clinical history, clinical diagnosis, upon initiation of dementia diagnoses
and the system will present a prompt with a pop-up menu for easy data entry. At the same
time, the prototype model will automatically retrieve the relative medical images and image
reports by the radiologists in order to support the diagnosis. (Fig.4)
Moreover, the radiologists get images from PACS with differential diagnosis of dementia
and the RIS can simultaneously automatically retrieve the essential clinical information
related to specific image characteristics defined by a consensus of the neurology and
radiology experts (Fig.5.). In this way, both user interfaces are designed to open a pop-up
window with the pull down bar when the clinician or radiologist is ready to input data into
the system. (Figs 4-5) [8]

Fig. 4. Prototype Model: Clinical History

Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 209

Fig. 5. Prototype Model: CT Image

4. Discussion
In this study, the intelligent information access framework for medical image databases was
designed to integrate radiological reports and clinical information. The most important
concern in this approach is the interdisciplinary collaboration among neurology, medical
informatics and radiology experts. The second important concern would be the
implementation with the critical and service-oriented hospital information system.
Therefore, we will test the system by the physician-in-the-loop approach to enhance
diagnosis practically and revise the system.
Moreover, we focused on decision support for dementia diagnosis, teaching and research.
Therefore, we would retrieve images of similar patients via a medical grid by querying
keywords of the base information, clinical History, clinical diagnose, Lab. we get the new
retrieved image database for the patients. Therefore, we will analysis their medical image
correlation by image processing and the relevance of the clinical data and images in order to
assist diagnosis, research and teaching.
Security and privacy is also a very important issue in this field of research. First, we build
trusted electronic relationships between healthcare customers, employees, businesses,
trading partners and stakeholders. Therefore, when we use the patient’s anamnesis,
examination data, or medical images, whether we have the patient’s permission or not, there
must be a set of procedures to follow accordingly. We plan to consider the security and
privacy function in the system, in the future.
The implementation of this prototyping system must be well organized and the initial
testing done on an offline system. The clinical data could be backed-up and copied to a
210 Data Storage

separate system for the purposes of the trial. When the first prototype model becomes fully
implemented, the system could be expanded by adding other neurodegenerative diseases
one by one to enhance the power and comprehensiveness of the intelligent retrieval for
clinical practice.
In the future, the system will be continuously developed to extend its spectrum of diseases
and clinical specialties.

5. Acknowledgements
This research work was partially supported by the u-Hospital project in National Taiwan
University funded by the Ministry of Education and the National Science Council
(95R0062-AE00-05), Taiwan, and by the ICT-Asia ONCO-MEDIA project, France.

6. Reference
[1]H. Muller, N. Michoux, D. Bandon ,A. Geissbuhler; “A Review of Content Based Image
Retrieval Systems in Medical”, International journal of medical informatics, vol. 73,
page(s): 1-23, 2004.
[2]W. Niblack, R. Barber, W. Equitz, M. D. Flickner, E. H. Glasman, D. Petkovic, P. Yanker, C.
Faloutsos, and G. Taubin, QBIC project: querying images by content, using color,
texture, and shape, in: W. Niblack (Ed.), Storage and Retrieval for Image and Video
Databases, Vol. 1908 of SPIE Proceedings, pp. 173-187, 1993.
[3]Lei Zheng; Wetzel, A.W.; Gilbertson, J.; Becich, M.J.; “Design and analysis of a
content-based pathology image retrieval system”, IEEE Trans. Information
Technology in Biomedicine,vol. 7, no. 4, Dec. 2003, page(s):240-255.
[4]Kak, A.; Pavlopoulou, C ;”Content-based image retrieval from large medical databases”,
Proceedings of the First International Symposium on 3D Data Processing
Visualization and Transmission, 19-21 June 2002 Page(s):138 – 147.
[5]Joseph R Carcione, DO, MBA on January 01, 2007), “Alzheimer's Disease Guide”,
http://www.webmd.com/alzheimers/guide/alzheimers-dementia
[ 6]“Medline Plus: Dementia”, http://www.nlm.nih.gov/medlineplus/dementia.html
[7]Bueno, J.M.; Chino, F.; Traina, A.J.M.; Traina, C., Jr.; Azevedo-Marques, P.M., “How to
add content-based image retrieval capability in a PACS”, Proceedings of the 15th
IEEE Symposium on Computer-Based Medical Systems, 4-7 June 2002, page(s):321 –
326.
[8]Traina, A.; Rosa, N.A.; Traina, C., Jr, “Integrating images to patient electronic medical
records through content-based retrieval techniques”, Proceedings of the 16th IEEE
Symposium on Computer-Based Medical Systems, 26-27 June 2003, page(s):163 –
168.
[9]“1.7m 'will have dementia by 2051”, http://news.bbc.co.uk/1/hi/health/6389977.stm, 27
February 2007, BBC NEWS.
[10]“Report of Alzheimer’s Disease Facts and Figures 2007”, Alzheimer's
Association,http://www.alz.org/
[11]“Taiwan dementia policy advocacy 2006”, Taiwan Alzheimer’s Disease Association,
http://www.tada2002.org.tw/index.html
[12] http://www.nlm.nih.gov/research/umls/
Medical Image Intelligent Access Integrated
with Electronic Medical Records System for Brain Degenerative Disease 211

[13]Le Thi Hoang Diem; Chevallet, J.-P.; Dong Thi Bich Thuy;” Thesaurus-based query and
document expansion in conceptual indexing with UMLS: Application in medical
information retrieval”, Research, Innovation and Vision for the Future, 2007 IEEE
International Conference on 5-9 March 2007 Page(s):242 - 246
[14 ] http://www.cdc.gov/nchs/about/major/dvs/icd9des.htm
[15] Yao-Yang, Shieh; Shaw, B.C., Jr.; Roberson, G.H.; ”Internet Web-based teaching for
cross-sectional anatomy and radiological imaging”, Computer-Based Medical
Systems, 1998. Proceedings. 11th IEEE Symposium, 12-14 June 1998 Page(s):192 - 197
212 Data Storage
Use of RFID tags for data storage on quality control in cheese industries 213

13
X

Use of RFID tags for data storage on quality

control in cheese industries
Raquel Pérez-Aloe(1), José M. Valverde(1), Antonio Lara(1), Fernando
Castaño(1), Juan M. Carrillo(1), José González(2) and Isidro Roa(2)
Universidad de Extremadura(1). Instituto Tecnológico Agroalimentario(2)
Spain

1. Introduction
The current laws that regulate the food security in the European Union lay down the
principles and general requirements of the food legislation. In the article 3 of the Regulation
178/2002 the traceability is defined as the “ability to trace and follow a food, feed, food-
producing animal or substance intended to be, or expected to be incorporated into a food or
feed, through all stages of production, processing and distribution”. In addition, “the food
traceability must be guaranteed in all the stages described above: production,
transformation and distribution” which implies the obligation of being able of identifying
every product in the company providing a complete information about it (Regulation 178,
2002).
Depending on the activity in the food chain three different types of traceability can be
distinguished:
1) Back traceability, called “suppliers traceability” as well, refers to the possibility of having
knowledge of what products are coming into the company, where are they coming
from as well as which farmers are their suppliers.
2) Internal traceability or “process traceability” refers to the information about what is
made, from what it is made, how and when it is made and the identification of the
product.
3) Forward traceability or “client traceability” means the possibility of knowing what
products the company delivers, when and to whom they have been supplied (Coscarón
et al. 2007; Spanish Agency for Food Security, 2004).
Although the law imposes a generic obligation of traceability, it is not mentioned the way in
which companies can achieve that goal (Decree 1808, 1991; Decree 217, 2004).
Nowadays, the sector of cheese production has no a procedure that exhaustively guarantees
a proper traceability throughout all the fabrication stages. The main problem is the cheese
ripening, done in special chambers as shown in Figure 1. The surrounding conditions in
these rooms, such as humidity, temperature, product handling (turning), mould and flora
growing avoid the individual labeling of the products. Is for this that the production and
quality control are performed by batches and the data storage of the information is done
manually by the company staff.
214 Data Storage

Moreover, if the cheese produced by a specific manufacturer has a Protected Denomination

of Origin (P.D.O.) then the product must satisfy stringent requirements about the milk
(origin, characteristics and preserving) and the elaboration, ripening and physical and
physicochemical characteristics of the final result.

Fig. 1. A traditional ripening chamber

To have an idea, for example, the “Ibores” cheese (Spanish Southwest P.D.O.) has to be
made with milk of mountain goats, with a preservation temperature less than 6º C, a
minimum protein content of 6%, fat around 4%, pH higher than 6,5 and so on. One of the
elaboration steps is the salting , which can be wet or dry, using only sodium chloride. In
case of wet salting, the maximum stay time shall be 24 hours in a saline solution with a
maximum concentration of 20%. Regarding the ripening time, it should be, at least, sixty
days, being the finished product cylindrical with a height from 6 to 9 cm, diameter
comprised between 12 and 15 cm and a weight from 750 to 1200 g. In the most traditional
presentation, the cheese is coated with paprika or smeared with olive oil, appearing diverse
colorations due to different moulds (Order 25/4, 1997). This is only a brief summary of the
conditions that “Ibores" cheeses should meet in order to be qualified as product with P.D.O.
The agency responsible for accrediting the P.D.O. is the Regulatory Board. Usually a
technical expert visits the cheese industry requesting to the manufacturer a collection of data
corresponding to the elaboration process to certify, if appropriate, the guarantee of origin
and quality. Besides, the Regulatory Boards of P.D.O. are recognized as entities of product
certification according to the European Standard EN 45011. This regulation shows in detail
the criteria to be applied by these entities to perform a certification of products reliable and
transparent. For instance, in the document that details the criteria to be evaluated by the
technical staff of the Regulatory Board of P.D.O. “Torta del Casar”, other Spanish Southwest
P.D.O. (Order 11/1, 1999; Order 9/10, 2001; Order 6/5, 2002; Regulation 1491, 2003),
appears that the Regulatory Board lays down a control based on sampling and analysis of
milk and cheese. The results will be used in the development of a set of measurable data
which will serve as basis for the accreditation process and to verify the adequacy of the
dairy industry to the certification system. This system, in fact, is based on a procedure of
self-control that includes several actions to be done by the person in charge of quality
Use of RFID tags for data storage on quality control in cheese industries 215

control in the company. These actions must be the periodic checking, as determined by the
regulations, of the physical, physicochemical, microbiological and organoleptic parameters
of the batches to be issued to the market, having a record of the results (CC/3, 2005; CC/4,
2007).
As it is clear from above, it is desirable that the cheese industry and the Regulatory Board of
P.D.O. interact effectively and efficiently to achieve that the final product reaches the
consumer with full guarantee of its origin, manufacturing process or quality. Therefore the
main objective here will be to join efforts between the dairy companies, in terms of
traceability, and the Regulatory Boards of P.D.O. with respect to their quality systems to
certificate products. In this way, the binomial Company-Regulatory Board would have the
necessary tools (hardware-software) that would enable compliance with current legislation
regarding traceability, as well as a continuous quality improvement thanks to the
optimization of the technological process according to the information obtained,
respectively. The key points here would be the speed on getting the required information,
the time saved and the simplification of tasks that would provide the implemented tool
compared to the manual records of data usually done by both, the employees of the cheese
industry and the technical staff of the Regulatory Board of P.D.O. in charge of the quality
certification.
To achieve this goal, a system which deals with the use of Radio Frequency IDentification
(RFID) has been developed. The RFID is a contactless method for data transfer in object
identification. The transmission of data is carried-out through electromagnetic waves. The
tag (transponder) consists of a microchip, as well as an antenna, which are usually put in
some form of casing for protection against different environments. The RFID read/write
device creates a weak electromagnetic field. If a RFID tag passes this field, the microchip of
the transponder wakes up and can send or receive data without any contact to the reader. If
the tag leaves the field, the communication gets interrupted and the chip on the tag stops
working, but the data on the tag keeps stored (Finkenzeller, 2003). To provide the system
with portability a PDA (Personal Digital Assistant) with reader and embedded antenna can
be used. Figure 2 shows the usual components of a RFID system.

Fig. 2. Usual components of a RFID system

216 Data Storage

There are different RFID technologies available: passive tags have no own power source and
take their energy directly from the magnetic field of the reader. Passive RFID tags do not
need any maintenance, but the reading distance depends on the size and frequency of the
transponder and antenna. Active tags are much more complex than passive tags and have
an internal battery to increase the reading distances. The life time of active tags is limited
through the battery. The semi-passive tags contain a power source, such as a battery, to
power the microchip’s circuitry. Unlike active tags, semi-passive tags do not use the battery
to communicate with the reader. Communication is done in the same manner than passive
tags do. Semi-passive tags might be dormant until activated by a signal from a reader. This
conserves battery power and can lengthen the life of the tag. There are systems with 134
kHz, 125 kHz, 13.56 MHz, 868 MHz, 915 MHz, 2.45 GHz available, having different
properties and reading distances depending on the environmental conditions.
Hence the main characteristics of a RFID system are:
- Communication done without physical contact between the reader and the tags.
- Soil resistant.
- No line of sight required.
- Read/Write memory.
- Possibility to store production and/or product data into the tag´s memory.
- Multi read /write capability.
Obviously, the use in dairy products of these “smart cards or tags” allows obtaining the
benefits that come from the fact of providing the product with a “certain kind of
intelligence”, such as:
- To have only one identity.
- To be able to communicate with his environment in a efficient way.
- To be capable of obtaining and retaining the information about itself.
Also in this context, the RFID solution offers, versus the „traditional” bar code, the
following advantages:
- Huge number of data in a reduced physical space.
- Automatic writing/reading in either individual or by batch mode. This is performed
thanks to the use of anti-collision algorithms, which allow reading several RFID tags
without interferences.
- Bar code needs visibility to work properly and the information stored can not be changed
using the same tag.
- Bar code identifies a type of object. The RFID tag identifies an only object in an only way.
- Bar code is easily damaged in wet environments like are the ripening chambers where
mould growing exists. In this case a RFID tag can be wrapped, for example, by biofilm
without affecting the reading/writing process.
Barcodes are of course cheap to create, but they are limited in their storage capacity and are
not flexible, when data needs to be changed.
Taking into account all the features mentioned above, a system based on the use of RFID
transponders seems to be the most appropriate to carry-out the proposed tasks. The idea is
to use RFID tags as physical support for storing the information required to perform a
“complete traceability” in cheese industries, as well as facilitate the collection of data
required by the technical experts of the Regulatory Boards in charge of quality certification
(Pérez-Aloe et al., 2007). The application will perform the three types of traceability (vendor,
process and customer) and an individual register of analysis and controls done in all the
Use of RFID tags for data storage on quality control in cheese industries 217

production stages (milk reception, storage, fabrication, ripening, quality and yield control
and pH measures) as well. In this way, the collected data corresponding to the process
conditions (humidity, temperature, pressure, ventilation) including microorganism,
biochemical and pH analysis and the connection of these data with the products provided
by the different suppliers are increased. Then, the traceability will be granted, both on
batches and on individual cheeses. This contributes to assure and certificate the quality,
making easier the location, immobilization and in some cases the effective and selective
recall when problems arise.
In this way, the device RFID would operate as an interface making easier and more efficient
the exchange of data and information between the company and the technical staff of the
Regulatory Board of the P.D.O. A sketch of this procedure can be seen in Figure 3.

Company Regulatory Board

WRITING READING

READING WRITING

Diary Technical staff RFID tag Technical staff

manufacturer P.D.O.

Fig. 3. Use of RFID tags in the integrated system of quality control Company-Regulatory
Boards of P.D.O.

2. System implementation
The system developed is similar to that shown in Figure 2. It consists of two complementary
systems; one is based on a personal computer (PC) and the other in a PocketPC (PPC). The
PocketPC utility was implemented as a complement to the PC due to its portability. The
Figure 4 presents a simplified block diagram of the two systems implemented.

Software Hardware
Interface Interface

Data to write
PPC Platform CF Module

Data Analysis External RFID

User
Software Antenna Tags
PC Platform +
RS232-
Reader
Data read

Fig. 4. A simplified block diagram of the two systems implemented

218 Data Storage

The PC system includes the S6350 midrange reader and the gate antenna series 6000 (Texas
Instruments_RFID, 2002). The reader operates at 13.56 MHz and is able to communicate
with tags that accomplish ISO 15693 protocol. The communication to the reader is done
through a PC serial port, using a RS-232 data transmission protocol, with one start bit, 8 data
bits, 1 stop bit and no parity. Several speeds can be selected within the range from 9600 bps
to 57600 bps. The PC starts the communication with the reader, through a pair of
request/response sequences accomplishing ISO 15693 standard, which establishes the
request stream format as well as the fields size. Some of the commands supported by the
reader can be used with addressing, i.e., read, write and lock. If addressing is used, the
command will be sent to a single tag, and in the other case, the command will be
broadcasted to all tags in the visible range of the reader. For example, Figure 5 shows the
display of the reader utility after sending the command inventory. As it can be seen, as a
response to this command, all transponders visible to the reader appear on the screen.

Fig. 5. Display of the reader utility in response to the command inventory

The system is also able to detect errors in such a way that if an error occurs during the
communication, the reader will send an error code in the response stream to the PC. Some
errors are tags not found in the vision range, a write attempt to a read-only block or an
addressing to a non existing block.
Regarding the PPC system, it consists of:
- PDA with Windows PocketPC 2003® (Hewlett Packard iPAQ, 2003).
- Compact Flash reader with built-in antenna (ACG, 2006).
The hardware interface runs also at 13.56 MHz. Both ASCII and binary transmission modes
are supported by the reader, but only ASCII mode has been developed because it simplifies
the process. The transmission protocol is ASCII mode at 9600 bps. The PocketPC and the PC
version are fully compatible, so that a tag written by one application can be read by the
other without problems. The software was implemented with Microsoft® Embedded Visual
C++ environment, and the source code includes the library supplied by the reader
manufacturer. All the data stored in a tag are accessed and read in just 8 seconds.
Use of RFID tags for data storage on quality control in cheese industries 219

Two different types of labels running at 13,56 MHz have been used in this application:
- Individual tags to be used to identify each cheese.
- Batch tags which store all data set and parameters related to its manufacturing process as
well as information about the cheeses that belong to it.
Besides, two different rounded-shape tags with different sizes have been tested for
individual labels, both on casein plate in order to improve the adherence to the cheese. The
smaller one has a diameter of 9 mm and uses the chip I-Code SL2 ICS20 (Philips
Semiconductors, 2003). The other one consists of a similar chip and has a size of 20 mm
(Labelys, 2007). I-Code chip has a user memory of 128 bytes, divided into blocks of 32x32
bits. Of these 128 bytes, only 108 are addressable bytes (27x32 bits).
The tags HF-I ISO 15963 (Texas Instruments, 2005) have been selected to be utilized as batch
labels due to their characteristics of thinness and flexibility. The tags have a 2 kBs user
memory organized in 64 blocks x 32 bits as shown in Figure 6 (a).

(a)

(b)
Fig. 6. Characteristics of the tags used
(a) Memory organization. (b) List of commands.

Tags can contain read-only data (ROM) and read/write data (R/W). The stored data on
blocks can be locked on factory or by user on an irreversible process, so data can not be
modified any more. In other cases, tags can be reused for future utilizations.
All the tags have a locked field, the individual identification code UID (Unique IDentifier of
the transponder) which is a 64 bits code, provided by the manufacturer and defined in ISO
15693 standard. AFI (Application Family Identifier, such as “transportation”,” finance” …)
220 Data Storage

code allows storing the type of application and DSFID (Data Storage Format IDentifier) the
data format. Both of them are 1 byte blocks.
The HF-I tags commands list is seen in Figure 6 (b). Basically, two modes of reading/writing
tags are available: single and multiple tag operation. In single tag operation, the first action
is to detect a single tag, then the reader identifies the UID, and the subsequent
reading/writing commands are referred exclusively to it. On the other hand, in the multiple
tag operation the information is broadcasted to all the tags in the reader range.
The PPC utility includes routines for data interpretation and the communication protocol
between the compact flash reader and the tags (AGC DLL, 2006). The procedure to establish
communication between both devices and start sending commands is shown in Figure 7 (a).
As it can be seen, the port where the reader is connected has to be detected and opened. If
the port has been opened successfully then the reader is activated and ready to send a set of
commands. Figure 7 (b) details the software code that performs the steps previously
mentioned.
The “writing tag” command consists of 11 bytes in ASCII mode, so that the first two bytes
indicate the address of the block to write, the following 8 bytes are the data to write and the
last byte is the NULL or stop chain with all bits to zero. However, the “reading tag”
command contains only 3 bytes, which refer to the block number to be read and the NULL
byte. Moreover, the reader has a 512 bytes buffer, so once the command is sent and the
reader receives the data, the utility can extract the required word from the buffer. The 512
bytes size allows the hardware to read all the tag blocks. Data are sent to the tags by means
of a software application. This application is easy-to-use, so that any employee of the

Detect serial port

Open a communication device

Open a reader

Send command
Get data

Close a reader

Close a communication device

company of Regulatory Boards will be able to use it without difficulties. company or

Regulatory Boards will be able to use it without difficulties.

// Initialise and open the communication port

presetSettings* settings=new presetSettings();
settings->baudRate =460800;
settings->protocol=0; // ASCII mode

char puerto[30]; // Detects the PDA port where the // CF reader is connected

RDR_DetectSerialPort(puerto);
Use of RFID tags for data storage on quality control in cheese industries 221

if (RDR_OpenComm(puerto,0,settings)==0){ //and check if successful. // Open the port

If //command returns a 0 an error //has
occurred
MessageBox((CString)"Error.",(CString)"Información",MB_OK);
//no need to specify the first //parameter in
ASCII mode
}else{
//If port has been opened //successfully then
activate //the reader
RDR_OpenReader(1,6);
RDR_AbortContinuousRead(); //Stop the continuous
//readout.Reading only on request
}
(a) (b)
Fig. 7. Procedure to establish the communication between the reader and tags
(a) Flow diagram. (b) Software code

The routine includes a “format tag option” which stores zeros in all blocks in order the tag
to be reused for future activities. A scheme representing the mode in which data are stored
in the tag memory and the data appearance on the PPC screen can be observed in Figure 8.
Regarding the data to be recorded in the tags, the writing process has to be optimized in
order to store as much as possible cheese production parameters. The fields that will be
saved on tags can contain numerical values (temperature, fat, etc.), alphanumerical values
(type of milk, manufacturer, Batch, batch qualification, etc.) and data concerning key dates
on production (elaboration, reception, ripening, etc.). Numerical data will be codified in
binary, differentiating between integer and decimal part. The number of bits required for
each part depends on the range of values and the precision used. In some cases, a data
manipulation is performed. For example, if a variable varies from 3 to 5, for a value of 4.57,
the integer and the decimal part are codified separately.

Data in PPC
Data provided by the factory
Stage 6 Milk Reception
Variable Units Format Range
Date 00/00/00 01/01/11-31/12/99

Temperature ºC 00,0 02,0 – 12,0

Fields associated on the tag

Fig. 8. Data recorded in tag and its representation on the PPC screen

Effective range of integer part has only 3 values (3, 4 and 5), and an offset of 3 is subtracted,
so only 2 bits are needed (0 to 2) instead of the 4 bits needed if offset is not considered (0 to
5). On the other hand, alphanumerical values will be stored as a simple database, that is,
assigning a numerical value to each data, i.e., type of milk can contain 4 values: “null”,
222 Data Storage

“cow”, “goat” and “sheep”, that will be linked to numerical values 0, 1, 2 and 3,
respectively.
Once all the fields and its associated number of bits are defined, the next step is to arrange
the whole tag. The purpose is that each stage of the production process will be associated
with a collection of blocks. Around two hundred variables related to the parameters
involved in the different phases of the cheese fabrication can be recorded in a tag. Some of
them appear in Table 1.
Variable Range
General Data
Product Code 1000-2000
Lot 010111-311299
Volume 500 lts-2500 lts
Pieces 50 uds-300 uds
Fit for consumption? yes/no
Number of tags 000000-999999
Specific data for the different stages
Reception date 01/01/11-31/12/99
Raw milk supplier Alphabetical
Tank Alphanumerical
pH 4,00 upH-7,00 upH
Acidity 13,00 ºD-18,00 ºD
Temperature 02,0 ºC-12,0 ºC
Chemical analysis
Fat 3,00 %- 8,00 %
Protein matter 2.50 %-7,00 %
Lactose 3,00 %- 6,00 %
sodium chloride 0,80 %-2,20 %
Microbiological analysis
Listeria monocytogenes yes/no
Salmonella spp. yes/no
Staphylococcus aureus 0-20.000 ufc/gr
Yield control
Lot weight 50,0 kgr-400.0 kgr
Lot yield
4,00-12,00
(Volume/Weight)
Average unit weight 750 gr-1400gr
Control of Clients
Client Reference Alphanumerical
Order number
1-300/0,80-400
(units/kgr)
Table 1. Some parameters stored in the tags

3. Experimental results
The systems have been tested in the industries that collaborate in this work. Previous to
place the prototypes in the companies, exhaustive tests were done on tags in laboratory in
Use of RFID tags for data storage on quality control in cheese industries 223

order to prove proper working under identical operation conditions than in factory. The
simulated conditions were temperature, humidity, biological contamination, acid corrosion,
ammoniacal gases and immersion on saline solutions, inhibiting substances, sugars, colorant
pigments, preservative substances, oils action (Figure 9 (a)) and mould growing (Figure 9
(b)).

(a)

(b)
Fig. 9. Test on tags in laboratory
(a) Immersion on saline solution, olive oil and paprika. (b) Mould growing.

Besides, physical tests have been also made which include friction and flexibility because in
ripening chambers, cheeses with its corresponding tags are subject to turns, shelving
changes, frictions and personnel manipulations. In all these cases, no significant negative
effects have been reported in communication with tags, with the exception of data reading
with metallic materials in the range of the reader, which reported erroneous reading.
Finally, in order to confirm that tags were suitable to be attached to the products from the
start of the fabrication to the end, the tags were placed on the surface of the cheese at the
224 Data Storage

beginning of the production. As can be seen in Figure 10, which corresponds to photographs
taken at different stages of production, both types of tags used as individual labels remained

(a)

Individual tags

Batch tag

(b)

Batch tag Individual tags

(c)
Fig. 10. Tags in different stages of production
Use of RFID tags for data storage on quality control in cheese industries 225

(a) Before molding. (b) After pressing. (c) At the end of production attached to the rind of
the cheese until the end of the process without reporting errors in the reading/writing
process. Figure 10 also shows the containers with their batch labels.

Regarding the signal range, for the PC application the system is able to read/write tags
inside a radius of around 20 cm, whereas the PPC reader has a limited range of around 25
mm, in both cases, enough for our purposes.
Finally, Figure 11 displays the way in which data are updated using the PPC system. The
application is user-friendly so that employees of factories in charge of quality control and
technical staff of Regulatory Board of P.D.O. responsible for quality certification could use
the system easily.

Fig. 11. Data update in cheeses with attached RFID tags

4. Conclusion
Two different systems that perform the reading/writing task with RFID tags in a cheese
industry have been implemented. One of them is based on a Personal Computer whereas
the other solution uses a PocketPC providing the application with the required portability.
The main objective has been to make available to the factory the complete traceability of the
products, in individual and by batches mode, as well as to provide the technical staff of
Regulatory Boards of P. D. O. with a tool that facilitates the process of quality certification.
The tags have been tested under different conditions of temperature, humidity, acid
corrosion, ammoniacal gases and immerse in saline solutions with inhibiting substances,
sugars, colorant pigments, preservative substances and oils. Besides, physical tests have
been also made which include friction and flexibility. In all the situations, no significant
negative effects have been reported in communication with tags, excepting for metallic
materials in the range of the reader. Around two hundred variables related to the
parameters involved in the different stages of the cheese production can be stored in the tag,
which improve considerably the quality and yield control of the production plant.
226 Data Storage

5. Acknowledgements
This work has been financially supported by the Ministry of Infrastructures and
Technological Development of the Government of Extremadura, under Grant PDT05A042
and PDT08A041 with the economic assistance of the European Union (FEDER).

6. References
ACG Multi ISO Plug-In Reader, (2006). ACG Identification Technologies, RDHP-0206P0-02
ACG RFID Reader DLL, (2006). User Manual, Rev 1.0, January 2006
Certification Criteria CC/3, (2005), Requirements to be met by the dairy industry, Manual of
quality of the Regulatory Board of P.D.O., Torta del Casar, (4), 1-8, Casar de Cáceres
Certification Criteria CC/4, (2007), Characteristics of the "Torta del Casar" Produced. Manual
of quality of the Regulatory Board of P.D.O., Torta del Casar, (5), 1-6, Casar de Cáceres
Coscarón, C.; Gil, M. & Legaz, E. (2007). Revista Alimentaria, No. 383, May 2007, 48-55, ISSN
03005755
Decree 1808, (1991) Which governs the indications or marks identifying the batch to which a
foodstuff belongs
Decree 217, (2004), Which regulate the identification and registration of staff, facilities and
containers involved in the dairy sector, and record the movements of the milk
Finkenzeller, K. (2003). RFID Handbook: Fundamentals and Applications in contactless smart
cards and identification, John Wiley and Sons, ISBN 0-470-84402-7
Hewlett Packard iPAQ h5550 with Windows Pocket PC, (2003)
Labelys traçabilité, (2007). RFID casein plate
Order 25/4, (1997). Laying down the regulations of the Denomination of Origin “Ibores
cheese” and its Regulatory Board
Order 11/1, (1999). Official recognition of the Designation of Origin "Torta del Casar"
Order 9/10, (2001). Adoption of the Rules of the Designation of Origin "Torta del Casar"
Order 6/5, (2002). Ratification of the Rules of the Designation of Origin "Torta del Casar"
Pérez-Aloe, R.; Valverde, J. M.; Lara, A; Carrillo, J. M.; Roa, I. & González, J. (2007).
Application of RFID tags for the overall traceability of products in cheese
industries, Proceedings of 1st Annual RFID Eurasia Conference, pp. 268-272, ISBN 978-
975-01566-0-1, Turkey, September 2007, Istanbul
Philips Semiconductors, (2003). I-Code SL2 ICS20, Smart Label IC, Rev. 3.0, January 2003
Regulation 178, (2002). Laying down the principles and requirements of food law,
establishing the European Food Safety Authority and laying down procedures
relating to food safety
Regulation 1491, (2003). Registration of Protected Designation of Origin "Torta del Casar" in
the Register of Protected Designations of Origin and Protected Geographical
Indications
Spanish Agency for Food Security (2004). Guide for the application of traceability in food
companies. Coimán S. L., NIPO 355-04-001-9, Madrid
Texas Instruments_RFID, (2002). HF Reader System Series 6000, S6350 Midrange Reader
Module, September 2002
Texas Instruments_RFID, (2002). HF Reader System Series 6000, Gate Antenna, September 2002
Texas Instruments, (2005). Tag-itTM HF-I Plus Transponder Inlays, December 2005