0% found this document useful (0 votes)
31 views14 pages

Shareswitch 2

The document discusses the history and evolution of shared versus switched network access. It describes the development of early telegraph, telephone and wireless networks that utilized shared access. The arrival of computers increased interest in networking and researchers explored using shared versus switched approaches for computer networks. Standards like Ethernet and IEEE 802 emerged to define shared network access for local area networks, but concerns about limitations like distance and guaranteed service time led to other standards as well.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
31 views14 pages

Shareswitch 2

The document discusses the history and evolution of shared versus switched network access. It describes the development of early telegraph, telephone and wireless networks that utilized shared access. The arrival of computers increased interest in networking and researchers explored using shared versus switched approaches for computer networks. Standards like Ethernet and IEEE 802 emerged to define shared network access for local area networks, but concerns about limitations like distance and guaranteed service time led to other standards as well.
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 14

Shared versus Switched munications technology that will introduce a fundamental

change in the way packets are transported.


or
1.2 How It All Started
Reports of the Death of Shared Access
Communications networks utilize one of two physical envi-
are Greatly Exaggerated ronments: (a) shared medium or (b) switched medium. In a
shared medium network a station communicates with another
Graham Campbell station by transmitting a message on a channel that is shared
Computer Science Department by all stations -- all stations receive the message but only the
intended destination station(s) copy the message. In a switched
Illinois Institute of Technology
network a station communicates with another station by utiliz-
[email protected] ing zero or more switches that connect a series of point-to-
point lines -- only the destination station(s) receives a copy of
Abstract the message.

Indications are that a shared environment supported by Both mechanisms have been in use since the advent of electri-
DQSA (Distributed Queue Switch Architecture) is superior cal-based communications. Samuel Morse’s first telegraph
to the conventional switch/router environment for most com- line between Washington, D.C. and Baltimore was a point-to-
munications. DQSA and its constituent MACs (DQRAP, point connection as were Alexander Graham Bells’s first
XDQRAP, DQLAN, PDQRAP, and CDQ), developed at the phone call to Mr. Watson and Marconi’s first transmission of a
Illinois Institute of Technology, represent a major develop- wireless message. But very quickly economics dictated shared
ment in shared access communications. A DQSA-based use of lines for the above in many scenarios: in wireless a spe-
MAC (medium access control) overcomes the limitations of cific frequency was shared by multiple users; in the nascent
Ethernet, especially those with respect to distance and QoS, telephone industry, competitors of the original Bell Telephone
thus permitting shared access, normally restricted to LANs, Co. introduced the ubiquitous party line that allowed multiple
to be applied to WANs. This paper provides an overview of customers to share a common line attached to a central switch.
the current LAN environment, presents a set of criteria In the latter case the shared/switched system has been
describing the ideal MAC, describes the DQSA constituent replaced, at least in the U.S., by connecting each subscriber via
protocols, and describes potential applications for these pro- a point-to-point line to a central switch so that a direct connec-
tocols. It closes with the argument that DQSA has the poten- tion can be established as required between any two subscrib-
tial to satisfy most communications requirements without the ers.
use of STM circuit switches, packet routers or conventional
ATM switches. The fundamental problem of utilizing a shared medium for
communications is the allocation of the medium to a specific
user. There are two basic methods: deterministic and non-
1 Introduction deterministic. The deterministic method ensures that each sta-
tion receives an equal opportunity to use the medium: the com-
1.1 Preamble mon methods are to either use a master station to poll each
station in turn or to distribute the responsibility by having a
The May 18th, 1998 issue of Network World contains an inter- token pass in turn to each station, possession of the token
view with the inventor of Ethernet, Bob Metcalfe, on the occa- granting temporary control of the medium. The deterministic
sion of the 25th anniversary of the invention of Ethernet. method can provide guarantees with respect to access time and
There was also a companion roundtable discussion with five intervals between opportunities to transmit but overall utiliza-
industry leaders. The consensus of the leaders was that, as tion is good only when: (a) every station is always ready to
speeds increased, switched Ethernet would dominate and the transmit; or (b) the message transmitted by a single station is
shared access component of Ethernet would disappear leaving very long relative to the time it takes to canvas all the stations.
only the frame structure and addressing. The feeling was
summed up in a statement by Dr. Eric Cooper, Chairman of Non-deterministic methods utilize contention-based mecha-
Fore Systems: “As the historical baggage of shared access is nisms and can make no guarantees about performance that has
left behind, there is no technological reason why Ethernet- proved to be satisfactory only when the offered traffic is light
compatible address and frame formats can’t be used on higher relative to the capacity of the medium.
and higher speed dedicated links.”
In general a shared medium network, using either determinis-
We agree with the statement that Ethernet compatible address tic or non-deterministic access methods, provides satisfactory
and frame formats will remain in use but argue that the use performance only when the length (time to transmit) of the
will still be in shared access. The difference is that the shared message is much longer than the time it takes for the message
access will be the MACs that constitute DQSA (Distributed to travel from source to destination. The desired condition is
Queue Switch Architecture). Read on and learn about a com-

© 1998d 1
that when the first bit of a message has reached the destination, service. In practice most implementations do not utilize IEEE
more than 50% of the bits in the message still remain to be 802.2 and instead use Ethernet. The sole difference between
transmitted. The ratio of the time it takes for a bit to travel IEEE 802.3 and Ethernet is that the contents of a 16 bit field in
between two stations to the time it takes to place the message the frame contains either a length signifying an IEEE 802.3
in its entirety on the line is very useful and is often given the frame or a type signifying an Ethernet frame. Systems can run
name “a” (little “a”) = Tp/Tx where Tp is the propagation in either mode.
delay and Tx is the length of the message. The units can be bits
or time but must be consistent. Stallings [1] provides a good Almost immediately there were concerns about the limitations
discussion of the fundamentals. of IEEE 802.3, specifically its lack of guaranteed service and
distance limitations. Some potential users, including General
Motors, were looking to LANs for factory automation and
1.3 The Arrival of Computers concluded that a protocol that could provide guaranteed ser-
The party lines of the telephone system and the shared use of a vice over a greater distance was the answer. The IEEE 802
frequency in wireless worked well because the time it took for Committee established IEEE 802.4 -- Token Bus. IBM also
users to sort out who next used the line was short relative to had reservations about Ethernet’s lack of guaranteed service
the time the user occupied the line, i.e., the equivalent of our and also with the fact that it could not scale to higher speeds
little “a” was very small. However the arrival of the computer while maintaining current distance and packet size and so
and the subsequent desire to network computers reintroduced IEEE 802.5 -- Token Ring came into being. The reader will
the question: to switch or to share? Initially all networking of note that if you have a protocol it appears advisable to have a
computers utilized point-to-point links but the bursty nature of patron that is a heavy hitter. Project IEEE 802.6 was estab-
computer traffic that often resulted in low utilization of the lished to address the MAN (metropolitan area networks) part
links spawned active research into the development of mecha- of the IEEE 802 official designation, specifically targeting the
nisms, termed MACs (medium access control protocols), that cable TV plant as a platform to deliver digital data to homes.
would permit multiple computers to utilize a common channel. Unfortunately there was close to zero interest on the part of the
cable companies in this concept in the early 1980s so IEEE
By the mid 1970s the theory of operation of the deterministic 802.6 drifted for awhile and then settled on making a standard
(token ring, token bus) and non-deterministic (Aloha, Slotted out of a new shared access method that had originated in Aus-
Aloha, and CSMA (Carrier Sense Multiple Access) MACs tralia -- DQDB (distributed queue dual bus).
was reasonably well understood. Very early on, Aloha, not
sensitive to the value of “a”, was and continues to be used in Once again the lack of what we now call QoS (Quality of Ser-
satellite networks. CSMA with collision detection (CSMA/ vice) was noted in Ethernet and so IEEE 802.9 -- Integrated
CD) was introduced soon after as Ethernet and to this day Services came into being. It defined a standard that “stitched”
dominates the MACs used in LANs. All are described by some 6 Mbps of synchronous channels, suitable for voice, on
Stallings [1]. top of Ethernet. The dissatisfaction with Ethernet continued
with the establishment of IEEE 802.11 whose goal was a stan-
dard for wireless applications. IEEE 802.12 -- Demand Prior-
1.4 The Standards ity, was the result of a schism in 802.3 between a group that
argued that Ethernet could run at 100 Mbps by simply reduc-
Given the subject of this paper it is well to spend some time
ing the distance covered from 2.5 Km to 250 meters. The other
discussing the IEEE 802 standards since collectively they
group argued that only a deterministic access method would be
make a fundamental statement about the status of shared
suitable for 100 Mbps speeds. The IEEE 802 Executive settled
access.
the matter by disregarding their charter of one problem, one
solution, by agreeing with both camps and authorizing 802.3
Ethernet™ was invented in 1973 by Bob Metcalfe while at
to develop a 100 Mbps version of CSMA/CD and establishing
Xerox but it was in 1979 while in a discussion with Gordon
project IEEE 802.12 to develop a deterministic shared access
Bell of DEC that he had the idea of establishing a standard as a
method, i.e., Demand Priority. In the mid 1990s there finally
mechanism that would allow DEC to contribute to the devel-
emerged an interest in using the cable TV plant to carry data.
opment of LAN technology without having to reinvent the
IEEE 802.6 had been established to define a standard for this
wheel [15]. Xerox was agreeable, Intel joined the effort and in
area but it was felt they had forfeited their franchise with
1980 the IEEE established the LAN/MAN Committee to
DQDB so a new project, IEEE 802.14 - Cable TV, was estab-
define standards for the emerging LAN market. Three projects
lished and is currently at work on a standard.
were established: IEEE 802.1- Station Management; IEEE
802.2 - Logical Link Control; and IEEE 802.3 - CSMA/CD.
An overview of the IEEE 802 activity is available at [2]. The
IEEE 802.1 provides standards for the administration of the
missing numbers: 7, 8, and 10, represent standards that are not
LAN, currently it is defining a standard for implementing
in themselves access methods. The author doesn’t know what
VLANs (Virtual LANs) that span several physical LANs inter-
happened to 13 but can guess.
connected via switches and/or bridges; IEEE 802.2 imple-
ments a mechanism that supports virtual channels and
provides for acknowledgments for the IEEE 802.3 datagram

2
1.5 Comments on the Standards 3. Lack of QoS: Ethernet has no effective priority mecha-
nism and thus cannot support any semblance of QoS.
We stated that the existence of the IEEE 802 standards in itself This problem is being addressed by IEEE 802.1 but the
makes a fundamental statement about the state of shared proposed solution has more application to switched
access. To wit, the existence of eight standards that address Ethernet than to classic shared access Ethernet.
basically the same problem indicates that none is satisfactory.
Now if Token Bus had overtaken Ethernet and in turn been A pair of questions: “Ethernet obviously has serious limita-
taken over by Token Ring and so on this would be a natural tions but just what are desirable features of the “Perfect
progression and would account for eight standards. However MAC”? It is strange that in the years since research started on
Ethernet has met all on the field of battle and emerged victori- shared access that the boundary conditions of an ideal MAC
ous. In fact the newer the standard the shorter its expected life have not been developed so that at least the flood of standards
cycle. Both IEEE 802.9 Integrated Services and IEEE 802.12 emerging from IEEE 802 could be rated. And the second ques-
Demand Priority disappeared in the MAC pool leaving barely tion, “If a MAC existed that possessed or came close to the
a ripple while IEEE 802.14 may not be deployed anywhere desirable features could that MAC dominate communications
due to the selection by the cable operators of MCNS (Multi- for LANs, MANs, WANs, etc., in the way that Ethernet domi-
media Cable Network System). IEEE 802.11 Wireless reflects nates LANs?” We list a set of desirable features for a MAC in
a committee effort to define a standard for the wireless envi- the next section and then describe the component protocols of
ronment. It employs CSMA/CA i.e., CSMA with collision DQSA that will in the future dominate communications.
avoidance. The fact that this standard is just being completed
is somewhat ironic in that the original work in the late 1960s 2.1 Features of the Ideal MAC
and early 1970s with Aloha and the CSMA variations were
conducted with the view that they would be deployed in the 1. Immediate Access: If the channel is not busy then trans-
wireless arena. It is much too early to comment on the viability mit immediately. This is the feature that along with a
of IEEE 802.11 but the author has seen at least one ad for a good backoff algorithm makes Ethernet what it is. We
wireless system which touted as one of its major features non- shall see that it possesses little else.
compliance with IEEE 802.11. Another possibly more disturb- 2. Full Channel Utilization: System throughput is equal to
ing event is that the HomeRF Working Group, now numbering the offered traffic up to the capacity of the channel
some forty plus members, will soon release the specifications whether one station or all stations are transmitting. Ether-
for SWAP (Shared Wireless Access Protocol) for use in the net can achieve full utilization if only one station trans-
home environment [27]. SWAP can perhaps be treated as the mits but terrible things happen when all stations have
offspring of both IEEE 802.9 and IEEE 802.11 since it uses something to send. Deterministic protocols such as Token
the CSMA/CA of dot 11 with provision for TDM channels to Ring provide almost full utilization when all stations are
carry synchronous voice as in dot 9. The author contends that ready to transmit but when only one station is ready much
this is yet another proof of the failure of the existing standards of the capacity could be utilized in presenting not ready
to satisfy anything more than very specific, temporary require- stations the opportunity to transmit.
ments. 3. Minimum Delay: Packet transmission delay should be
that of an ideal M/M/1 or M/D/1 queueing system for
variable length or fixed length packets respectively when
2 What’s Wrong and What’s Right the offered traffic has a Poisson distribution.
4. Predictable Delay: Once a station is ready to transmit it
Ethernet dominates the communications market for all appli- should know how long it will be till it actually can trans-
cations excepting POTS (plain old telephone service) for dis- mit.
tances up to a couple of kilometers. Fast Ethernet (100 Mbps) 5. Fair Access: Transmission should be based on FIFO with
is restricted to 250 meters on fiber and approximately 100 priorities an option.
meters on UTP but serves areas beyond this distance by means 6. Distributed Control: Nodes manage requests/transmis-
of switches. Beyond these distances, excepting for satellite and sion independently, i.e., no central control required.
wireless networks, switching/routing rules all communications Ethernet offers this feature.
including POTS. But we are addressing the role of a shared 7. QoS (Quality of Service): Available via priorities and/or
medium in communications and whether a shared medium can via the equivalent of synchronous channels. When an
be used in distances beyond Ethernet’s realm. Let us review application desires say a 1 Mbps channel then that
the situation: Ethernet rules its ever-decreasing (geographi- should be available either approximately or exactly
cally speaking) domain despite being a three-time loser: according to the requirements of the station.
8. Topology Independent: Our ideal MAC should operate
1. Low Utilization: Theoretical utilization is high but in on all topologies. DQDB possesses many of the ideal
practice Ethernet systems are operated in the 5% - 40% characteristics in this list but restriction to a dual-bus
utilization range. topology has turned out to be a fatal weakness. Ethernet
2. Distance Sensitive: Ethernet is limited to 2.5 Km at 10 originally ran on a single bus but easily adapted to a
Mbps decreasing by a factor of ten for every tenfold tree-and-branch topology and then to the current hub and
increase in speed. star topology.

3
9. Idle Nodes: Should not utilize any network resources.
We have reached the fourth and final ideal characteristic
that Ethernet possesses. 9

10. Distance Insensitive: Performance should be indepen- 8

dent of little “a”: i.e., the distance can range from a few 7

meters to tens of thousands of kilometers, e.g., continen- 6


tal or satellite networks, at any transmission rate.
5
Delay in Slots
The first five features describe the performance of a typical 4

packet/cell switch, which is reasonable, since even though our 3

stations are distributed across some arbitrary distance our goal 2

is that the medium connecting them is utilized just as if all the 1

stations were locally connected to a switch. 0


DQRAP
0.1
0.2 Kleinrock Gremlin: b = 100
0.3
0.4
0.5
We indicated that Ethernet possesses just four of the ten Offered Load (Poisson)
0.6
0.7
0.8
M/D/1 - Ideal

0.9
desired characteristics of a MAC and yet it is “king of the hill”
for all applications excepting voice in a 2.5 Km area. There is
a question posed in the title “Shared versus Switched?” i.e., if
a MAC possesses or came close on all ten of the desired char- Figure 1. Delay Comparison: DQRAP, M/D/1 and
acteristics, especially #7 (QoS) and #10 (distance insensitive) Kleinrock Gremlin
could it then support most of or all communications that,
beyond the two km boundary, presently utilize switching/rout- Higher Layer PDU e.g. IP Datagram

ing? Needless to say the reason this paper was written and that Encapsulate PDU
you are now reading it is that the DQSA family of MACs pos-
CPCS-PDU Payload -AAL 5
sesses or comes very close to all ten desired characteristics and
Divide AAL-5 Payload into 48-byte segments
so “Yes, Virginia, there is a Santa Claus -- DQSA will bring ATM Cell
Overhead
you all the feature your little heart desires.”
Map segments into 53-byte cells and transmit using DQRAP.

3 DQSA: The Protocols


DQRAP Control Minislots (CMS)

Cell transmitted individually in slot, copied by destination station.


3.1 DQRAP: The seminal DQSA protocol Preemption supported -- low priority cell defers to high priority cell.
Recurring slots can be reserved to support QoS channel.

We have taken a long trip to get to the gist of the paper but Figure 2. DQRAP Segmentation for ATM
here it is. DQRAP (Distributed Queueing Random Access
Protocol) is the seminal protocol of DQSA. The author resolving collisions of requests, equivalent to the problem of
admired the simplicity of DQDB and launched a research resolving collisions of data in a shared medium, a problem that
effort, the goal of which was a MAC that provided the perfor- had been the subject of essentially fruitless research since the
mance on other topologies that was provided by DQDB over a early 1970s.
dual-bus topology. The details of DQDB are available in Stall-
ings [4], but briefly, the performance of DQDB (which does The author then did what any good professor would do: he
satisfy seven of our ideal characteristics) is based upon sta- assigned his then student Wenxin Xu to “work out the details”.
tions sending requests via single bits to upstream stations, Some details! DQRAP is the result but there was only one
while at the same time keeping count (queue length) of the bits “detail” taken from DQDB.
received from downstream stations. When ready to transmit,
the station permits a number of empty slots equal to the current Briefly, in DQRAP, control minislots take the place of the
count (queue length) to pass before transmitting. Thus a ready request bit in DQDB. Ready stations select one of three CMS
station “joins a queue” of those stations already ready and (control minislots) randomly in which to transmit a request for
transmits in turn. The performance is close to ideal but being a slot. Figure 2 illustrates the relationship of the CMS and the
limited to a dual bus is what caused DQDB to enter the MAC dataslot. Global feedback indicating a success, i.e., no colli-
pool leaving barely a ripple. sion, causes all stations to increment their TQ (Transmission
Queue) by one. If a collision occurs in a minislot, the Collision
The dual bus topology allows requests to transmit on the Resolution Queue (RQ) is incremented indicating that a group
downstream channel to be sent upstream on the other bus. of stations has collided. The group at the head of the RQ is
There is no danger of colliding since all upstream stations can given exclusive use of the CMS to resolve their differences,
“see” the incoming request and thus wait, if ready, to transmit new arrivals are blocked. While this contention resolving is
their own request. In a tree-and-branch topology, requests ongoing the station at the head of the TQ transmits its data.
transmitted “upstream” can and will collide with requests The key to DQRAP is that using only three CMS the system
arriving on other branches. The problem reduced to one of will, on average, resolve collisions caused by the arrival of
requests of multiplicity N in a time span of less than N

4
compared with 6 slots delay for the M/D/1 system. Although
IP Datagram
CRC not shown, the DQRAP delay at 95% load is 13.52 as com-
Frame Relay Encapsulation pared with 11 slots for M/D/1. With DQRAP at 90% offered
traffic, a new station makes an average of 1.6 requests before
Divide frame into 64-byte segments, entering the TQ. The original DQRAP simulation and analyti-
Frame Relay Header no extra overhead. cal results, other details, and the DQRAP algorithm are avail-
able in [5] and [13]. Kleinrock’s material is in [28].

DQRAP utilizes a fixed size data payload. Therefore data


being transmitted, e.g., Internet Protocol (IP) datagrams, must
Control Minislot contains request by station for multiple slots. be sliced into fixed-size segments. If the ATM (Asynchronous
Segment transmitted individually in slot directly to destination. Transfer Method) adaptation layer (AAL) is used for the SAR
Preemption supported -- PDQRAP ensures low priority defers to high priority.
(segmentation and reassembly) the result is a distributed ATM
Figure 3. XDQRAP Segmentation for Frame Relay switch. The conventional central switch is eliminated -- all
switching is carried out on the transmission medium under the
dataslots thus ensuring 100% utilization of the dataslots. control of the distributed stations. The segmentation is illus-
trated in Figure 2.
Interestingly enough, even though the research was inspired
by DQDB, the “transmission queue” component turned out to The operation of DQRAP is totally dependent upon successful
the only DQDB “detail” utilized. DQRAP utilizes components determination of the contents of a minislot so Miramica inves-
of MACs already extant that include CMS (control minislots), tigated the robustness of DQRAP. His research showed that
immediate access, tree splitting, and blocking. The tree proto- DQRAP can suffer up to 10% misreading of minislots before
col introduced by Capetanakis [25] can be used to explain why performance is affected [14]. The reason is that while three
DQRAP works with three minislots. Capetanakis’ first tree minislots are used to resolve contention, mathematically
protocol has a throughput of 0.347, Chang shows that DQRAP speaking less than three minislots will do the job -- the differ-
using a single minislot has a throughput of 0.347, two minis- ence, especially under heavy loads, allows mistakes to go
lots provide a throughput twice that at 0.694 and Bingo! three unpunished. McPheters showed that the performance of
minislots passes the 100% barrier [22]. Thus resolving with as DQRAP in the finite-model environment moved to that of the
few as three minislots along with queues for both transmission infinite model with as few as six stations [18]. Khasawneh
of data and contention resolution are the keys to DQRAP. compared the performance of DQRAP on a dual-bus topology
with that of DQDB and showed that performances were com-
The perfect shared access method will have performance equal parable in that restricted topology [19].
to an M/D/1 queueing system. Kleinrock in an invited paper
[28] states the same and analyzed such a system assuming that Summing up, DQRAP represents a remarkable breakthrough
there was an all-knowing “gremlin” that could pass the word in that it satisfies two key ideal characteristics: QoS and dis-
to a terminal on the total number of busy terminals, thus per- tance insensitivity, and satisfies or comes close to the other
mitting that terminal to determine its position in the queue. He eight desirable characteristics. QoS is discussed in section 3.4,
was realistic in recognizing that even a gremlin requires band- distance insensitivity is discussed in section 3.5.
width to pass this information and so included this factor in his
analysis. He plots delay for ratios, b, of message size to infor-
mation of 10, 100 and 1000 against offered traffic. Figure 1 3.2 XDQRAP (Extended DQRAP)
shows the delay for the ideal M/D/1 system, Kleinrock’s
Gremlin: b = 100, and DQRAP for offered traffic ranging up Each request in a minislot in DQRAP reserves a single slot.
to 90% of capacity. DQRAP performance is impressive as What if we permit a ready station to request a multiplicity of
compared with the perfect M/D/1 system and Kleinrock’s the- slots with one request? This obviously will increase the size
oretical system. It does appear that at 90% traffic DQRAP has and complexity of the minislot (error checking must now be
a lower delay than Kleinrock, but it is not so. Kleinrock included) but what it buys is the ability to slice a variable
charges the gremlin overhead against capacity thus impacting length IP datagram into a series of fixed-size segments and to
the curves as system capacity is approached. DQRAP plots transmit the segments without any further overhead. Another
against available slots excluding the overhead of the minislots. key benefit is that the number of minislots can be reduced to
Kleinrock also assumes past history is not available to the two. In section 3.1 we pointed out that the throughput of
gremlin thus total information is sent each time. DQRAP DQRAP with two minislots is 0.69, but that assumed one
maintains a past history with the TQ and RQ thus reducing the dataslot per minislot request. If stations on the average request
amount of information sent, especially at high loading. anything greater than 1/.69 ≈1.5 dataslots then two minislots
DQRAP minislot overhead in practice will range from 1% - suffice.
2% in a synchronous system to 15% - 20% in a satellite sys- The ATM SAR (segmentation and reassembly) stage requires
tem. that each segment include enough information so that the cells
can be stitched back together into the original datagram. Let us
The delay for DQRAP at offered traffic of 90% is 8.25 slots as assume that in XDQRAP an IP datagram is passed across the

5
Frame Relay interface and thence to XDQRAP. The HDLC XDQRAP was proposed to IEEE 803.14 [26] and the three
frame is 1500 bytes. The SAR stage of XDQRAP requests minislot feature of DQRAP was picked up by the IEEE 802.14
1500/64 = 24 slots. All stations receive the request and add 24 project and stitched on top of a tree protocol [16]. They would
to their copy of the distributed TQ to make it 57 (this assumes have come up with a much cleaner MAC with better perfor-
that the current value was 33). Every station is now aware that mance if they had just used DQRAP or XDQRAP.
after 33 slots have passed they should then treat the subsequent
20 slots as carrying a single datagram. Thus there is no XDQRAP is the basis for the DQSA distributed Ethernet,
requirement to identify the individual segments. Figure 3 illus- Frame Relay, and IP switches. Simulation results are available
trates the segmentation process. in [6], the detailed algorithm is available in [17].

A priority mechanism, described in Section 3.4, allows higher


priority messages, for instance control messages or voice 3.3 DQLAN: Variable Length Frames
packets, to preempt the ongoing transmission of a longer DQLAN (Distributed Queue Local Area Network) is an imple-
packet. Figure 6 illustrates the simulated performance of mentation of DQRAP that is restricted to the same geographic
XDQRAP and priorities where the offered load consists of a area as an equivalent speed Ethernet system. Variable length
mix of 50% single-slot (64-octet) packets and 50% 30-slot frames are carried without segmentation. The performance at
(1920 octets) packets. This type of load is representative of an lighter offered loads is almost identical to that of Ethernet, i.e.,
environment where there are many long packets representing immediate access -- the station finding the medium not in use
ongoing file transfers with an equal number of short packets transmits immediately. However, as the offered load
sent as acknowledgments, etc. Note that the average delay of increases, the number of collisions occurring in Ethernet
the short packets, assigned high priority status, is almost con- increases to the point where operation is unreliable. DQLAN
stant at approximately 2.5 slots over offered loads ranging moves seamlessly from immediate access to a reservation sys-
from 10% to 95%. Note also that the average delay of the tem thus providing reliable operation up to the full capacity of
longer packets at 95% offered load is approximately 300 slots, the channel. DQLAN cannot provide a guaranteed bandwidth
approximately ten times the packet length. This compares to the exactness provided by the slotted versions of DQSA, but
favorably with the ideal delay of an M/D/1 system at 95% load implemented with PDQRAP it provides a “controlled” level of
and is a result of both the efficiency of XDQRAP in that when service that approaches that of guaranteed bandwidth. Table 1
the average packet length is several times the slot length then shows the performance of DQLAN over a range of offered
contention in the minislots is minimal and the fact that a loads at 10 Mbps. Note that throughput tracks the offered load
packet of length 30 slots has a built-in latency of 15 slots just linearly and that at 90% load factor the delay is still reason-
to access the system. able. Details of the algorithm and simulation results compar-
ing DQLAN and Ethernet are presented by Wu and Campbell
Obviously the shorter packets did not queue behind longer [7]
packets, a key bragging point of ATM but unlike ATM
XDQRAP does not have to encapsulate each segment -- the DQLAN can be used in wired applications but its major use
segments can be sent “naked”. Figure 3 illustrates the segmen- will be in the wireless environment where it can often be sub-
tation process when XDQRAP supports a distributed frame stituted for CSMA/CD or ISMA (Inhibit Sense Medium
relay switch carrying IP traffic. Access) implementations by simply modifying the code in a
DSP. ISMA is the access method commonly employed in mes-

60
0.9

0.8
50

0.7

0.6 40

0.5
Throughput Delay: ms 30
0.4

0.3 20

0.2
DQLAN 1024
DQLAN 512 10 CSMA/CD 128
0.1 DQLAN 256
CSMA/CD 256
CSMA/CD 512
DQLAN 128
CSMA/CD 1024
0 CSMA/CD 1024
0 DQLAN 128
CSMA/CD 512 DQLAN 256
0.9
0.8 CSMA/CD 256 0.9
0.7 0.8 DQLAN 512
0.6 0.7
0.5 CSMA/CD 128 0.6 0.5 DQLAN 1024
0.4 0.4
Load 0.3 0.3
0.2 0.2
0.1 Load 0.1

Figure 4 Throughput: DQLAN vs Ethernet Figure 5 Delay: DQLAN vs Ethernet

6
sage systems used by parcel delivery services and has the same
limitations with respect to throughput that affect CSMA/CD
(Ethernet). 350

300

3.4 PDQRAP and QoS 250

The buzzword (or buzz acronym) nowadays is QoS (Quality of 200


Delay in Slots
Service). In fact it would not matter how good DQSA was 150

with respect to delay, throughput, etc., it would not fly if it


100
could not support QoS. A simple definition of Qos would be
“the ability to give ‘em what they want!” We will be more pre- 50

cise and use the classes of service as defined by Microsoft for 0

their NDIS (Network Device Interface Specifications), to be 0.1


0.2
0.3 30-Slot Packets
0.4
available in Windows NT Version 5.0 [24]. The three classes 0.5
0.6
0.7
0.8
Single-Slot Packets
0.9
are: Load: 50% 30-Slot Packets, 50% Single-Slot Packets 0.95

Figure 6. XDQRAP Delay: High and Normal Priority


1) Best Effort:
- Default flow of priority levels is easily extended by increasing the size of
- Not typically requested by applications the priority field to accommodate 2M priorities where M is the
- Low priority number of bits in the field.
- Typically borrows from other flows
“Guaranteed Service” is provided when DQSA is imple-
2) Controlled Load mented in a slotted system thus allowing a network manager to
- Gets service equivalent to lightly loaded network allocate recurring slots to a requesting station. These recurring
- Medium priority slots reflect a specific requested bandwidth and thus Microsoft
Class 3 service is supported. This allocation of “real” band-
3) Guaranteed Service width to a user is contrasted with the ATM approach of con-
-Guaranteed delay bounds trolling the total traffic entering the network and then trusting
-Highest priority that the limit in traffic plus priority for CBR cells at intermedi-
ate switches will satisfy performance requirements. Band-
We address each in turn. DQSA without priorities provides a width guaranteed by DQSA is still there even when the other
“best effort” that as shown in Figure 1 is superior to other two classes of traffic swamp the network.
MACs right up to full utilization.
DQSA is unique in its ability to support the three Microsoft
“Controlled Load” is taken care of by PDQ-RAP, a terrible classes of service and thence CBR, VBR, and ABR in ATM.
play on acronyms but PDQ-RAP is Priority DQRAP. In
DQRAP a ready station transmits in a CMS and upon receipt PDQRAP is described by Lin and Campbell [10]. Wu and
of feedback indicating success it joins the transmission queue, Campbell describe how CBR (constant bit rate) channels are
TQ. All requests are treated on what is, based on slots, a FCFS supported in DQSA [11]. Lin and Campbell [12] describe
(first come, first served) basis. In PDQRAP, a priority bit is what a great job DQRAP does for packet voice using only
added to the minislot and if a station’s request is high priority “best effort”.
the station sets the priority bit in the minislot when transmit-
ting. All stations receive the CMS and if the priority bit is set
then the high priority queue, HTQ, is bumped instead of the 3.5 The Long and the Short of It: How
TQ. The two queues operate in parallel but stations on the TQ Does DQSA Spread Out a Switch?
only transmit when HTQ is zero. Simulation and analysis
The basic operation of the DQSA protocols assumes that the
shows that when a PDQRAP system is 90% loaded the aver-
feedback of the minislots arrives back at the stations before it
age delay for high priority traffic is reduced to that of a lightly
is time to launch the next transmission but the claim is made
loaded system. This is ideal characteristic #5. This also satis-
that DQSA allows a common channel to be shared by hun-
fies the Class 2 service requirements described above for Con-
dreds of stations spread over thousands of kilometers. This is
trolled Load. This performance is graphically illustrated in the
accomplished by the use of interleaving, first described by
chart of Figure 6 that was referred to in the section on
Massey [8]. Interleaving requires the implementation of multi-
XDQRAP. The delay of the high priority traffic is nearly con-
ple protocol engines that operate in parallel, with each
stant across all loads with a delay that is comparable to the
“engine” operating on a slot, the number of slots being equal
delay encountered by normal traffic entering a lightly-loaded
to the round trip distance plus one. This technique will not
network, see Figure 1.
work with the carrier sense protocols, e.g., CSMA/CD, where
PDQRAP is not a self-standing access method in itself but is close to instantaneous feedback is required but is applicable to
implemented in the other DQSA access methods. The number most other protocols.

7
Economics of QHub
similar to Ethernet Hub
Bridge/Router
QHub

File Server

100 - 200 meters UTP


or
Several Km Fiber
Print Server
Stacking not
limited by distance

QHub XDQRAP transmits Ethernet frames QHub


in 64-byte segments.
100% utilization of dataslots,
priority preemption, and
Voice circuits use either guaranteed bandw idth Video circuit uses guaranteed
Voice over IP or Isochronous bandwidth channel
DSO channels. that is broadcast

NICs w ith XDQRAP MAC

Figure 7. DQSA LAN Implementation

The interleaving technique developed for DQSA is highly due to framing and switching in the underlying physical cir-
optimized as compared with conventional interleaving. Wu cuits that would require an increase in the interleaving factor.
developed a method that effectively reduced the number of DQSA delay performance departs from the M/D/1 ideal as the
parallel protocol engines from N to 1 thus making for efficient distance increases. A rule of thumb is that the average access
implementation. A DQSA DS1 system utilizing a 560-bit slot delay in a DQSA networking using interleaving is roughly 1.6
(64 bytes payload plus minislot overhead) results in a slot times the round trip delay at an offered load of 90%. DQSA
length of 363 µsecs. Referring to Figure 7, a WAN set up in interleaving is described by Wu and Campbell [9], Lee investi-
the same manner as a LAN, (there will be more discussion of gated the effect of interleaving on DQRAP and the benefits
this network in a later section) the maximum distance from achieved using a global TQ as compared with separate TQs
HQ in Chicago to the most distant stations is approximately [20].
2000 miles. Using a propagation delay of 8 µsecs per mile the
interleaving factor for this network operating at DS1 speed
would be at least (8*2000*2/363) + 1 ≈ 100. The same net- 3.6 CDQ - Cascaded Distributed Queue
work operating at DS3 speed would require an interleaving Network
factor of about 2800. In both cases there could be added delay We take this opportunity to introduce what could be the final

HQ

East N
etwork
k
3 Trun
/NT1/T
FT1/T1 West N
etwork
Centra
l Netwo
rk

Detail
FT1/T1/NT1/T3 Trunk

Branch Office

QTap at
Carrier CO
FT1/T1/NT1/T3 Local Loop

Figure 8. DQSA Network Functioning as Routerless WAN for Ethernet, IP, Frame Relay, or ATM

8
member of the DQSA family and one that fills a gap in DQSA. The second set of simulations included extensive overloading
What is this gap? DQRAP and XDQRAP each allow a net- of the system by introducing “bumps” in traffic on several
work of arbitrary size and speed to be organized using hub- intermediate segments that brought total offered traffic to well
and-spoke topology, as shown in Figures 7 and 8. But no one over 100% of capacity. TRAFC, the flow control program
including the author suggests that “one big network” with a developed by Chang ensured that (a) priority traffic was not
single hub can serve the entire country. Fedex, where business affected and (b) congestion problems are avoided, i.e., there is
justifies, establishes satellite hubs such that parcels originating no cell loss other than that due to BER [22].
in and destined for the catchment area served by the satellite
hub never leave that area. We introduce CDQ (Cascaded Dis- More research remains to be done on CDQ but the results so
tributed Queue Network) to provide this same function in a far confirm our belief that CDQ can successfully interconnect
communications network. This permits what can be termed DQRAP networks and, using ATM cells to transport data,
heterogeneous networks, e.g., the Internet, to be supported in a enable DQSA to satisfy most communications requirements.
more efficient manner, i.e., a packet originating in Boston does There is a discussion of how CDQ could be deployed in Sec-
not have to travel via Kansas City to reach New York. tion 4.3.

The thinking behind CDQ was that the network would consist
instead of a series of “satellite” DQRAP networks that are cas-
4 DQSA and the Real World
caded (thus the named CDQ). The idea was that when a
request reached a hub in one segment, the request would be Let’s now look at how DQSA might fare in real world applica-
forwarded immediately with the hope that before the data tions. Potential applications range from parallel bus arbitra-
reached the first hub, a position on the TQ would have been tion, e.g., PCI, to LANs, to MANs, to WANs, satellite
assigned thus permitting almost immediate onward transmis- networks, wireless networks, PCS networks, etc. We shall
sion. The station at the hub would double as the end station for confine the discussion to LANs, WANs, a discussion of CDQ
the next segment. Needless to say this approach did not work as applied to WANs, and a few other odds and ends.
but once again a student, Chen-Hung Chang, worked out
details that turned out to be far more complicated than origi- 4.1 DQSA in LANs and MANs
nally envisaged.
Figure 7 illustrates a DQSA LAN. At first glance it could be a
The CDQ member of DQSA lands us with the grandaddy of all typical Ethernet LAN but there are startling differences. Let’s
network problems, e.g., the potential for congestion at the enumerate:
nodes joining the segments. However, having the networks
distributed across the segment rather than feeding them 4.1.1 Distance: A connecting link can range up to several
directly to a switch/router opens up new approaches to flow Km if fiber is used. There is in fact no upper limit on the dis-
control -- wonderful things can be accomplished when total tance of a link, the distance is limited only by the power bud-
traffic consists only of “distant” traffic, i.e., from the next seg- get, meaning that the configuration shown could, with the use
ment arriving via a single port plus “local” traffic that can be of repeaters, represent a MAN as well as a LAN.
throttled very quickly.
4.1.2 Speed: The Ethernet-like simplicity of the physical
The first stage of the research is complete but not yet pub- access method means that at whatever speed a bit can be gen-
lished but here are some interim results. Extensive simulations erated for reading from/writing to a medium it is likely that the
were carried out of a network similar to that shown in Figure 9 same chip logic can be used to implement the four-state FSM
excepting that there were twenty segments. Each segment had and binary counters that constitute XDQRAP. Thus a gigabit
DQSA LAN with a radius of several kilometers could serve an
“a” = 100- thus the network corresponded approximately to a
entire campus.
network carrying ATM cells, plus overhead, from New York
to San Francisco operating at something less than DS3 speed. 4.1.3 QoS: A typical DQSA LAN will achieve over 90%
Two types of simulations were carried out, one assumed a mix utilization of the available bandwidth, i.e., the minislot over-
of traffic that added up to 90% average Poisson type load (at head will typically be less than 10%. As described in previous
the origin) on a segment made up of 30% through traffic from sections, the three classes of service are easily implemented,
N.Y. to S.F., 30% local to each segment, and 30% passing i.e., best effort, controlled load, and guaranteed bandwidth.
through multiple segments. The results showed no cells lost, The diagram shows both a phone and a TV camera connected.
delay of approximately two slots at each hub, maximum queue In both cases the data can be compressed and packetized and
length of approximately 70 at each hub, and an overall transit then transmitted using either controlled load or guaranteed
time of approximately 1.3 times the actual propagation delay, bandwidth. Both types of signals can also be transmitted in a
e.g., some 35 ms for a cross-country trip. Other than propaga- basic, non-encapsulated format utilizing TDM-like slots. The
tion delay the major delay is access to the initial segment latter case shows the versatility of DQSA in that bandwidth
which as pointed out earlier is approximately 1.6 times the can be dynamically allocated so that a DQSA LAN could
round-trip time for the segment at 90% loading. function as a conventional switch, e.g., in a 100 Mbps system a
DS3 channel plus multiple DS1 channels could be supported

9
along with random traffic. Also the shared medium makes it concentrate our efforts on describing how a DQSA WAN can
trivially simple to multicast or broadcast the TV signal in any be implemented.
format to multiple receivers.
A DQSA WAN can be implemented as a distributed Ether-
In section 1.1 we agreed with the statement that the Ethernet
switch, distributed IP switch, distributed Frame Relay switch
frame would be around for a long time but argued that its use
or a distributed ATM switch. The difference lies in the seg-
would continue to be in the shared medium -- the DQSA envi-
mentation process as illustrated in Figures 2 and 3. Figure 8
ronment. Our statement also applies to any enhancements to
illustrates a DQSA WAN consisting of three “spokes” emanat-
the standard that aim at implementing some type of QoS in
ing from a “hub” in Chicago. Each spoke contains subspokes
Ethernet, an almost impossible task for shared Ethernet but
that are joined to the main spoke via a simple QTap. The
possible for switched Ethernet. An example is the recently
spokes are circuits that can range in speed from DS0 to DS3
adopted IEEE 802.3ac extending the length of the standard
and higher speeds.
IEEE 802.3 frame to accommodate amongst other features a
priority capability. As described above DQSA will be able to
A branch office is connected via a local loop to the spoke via
utilize the extension in a much better fashion than can Ether-
the same “QTap” as described above and shown in the inset
net.
detail in Figure 8. The QTap fulfills the same function as the
QHub in a DQSA LAN. It basically consists of an “and” gate
4.1.4 Network Management: The shared nature of the
to merge two digital signal paths -- what complexity there is in
medium means that the simplicity of management that was the
a QTap is devoted to the necessity of synchronizing the under-
hallmark of early Ethernet systems, is now possible even with
lying DSx frames before merging. It is still probably more
say a gigabit DQSA LAN connecting hundreds of stations
than an order of magnitude simpler than a router.
spread over kilometers. CDQ backbones can be utilized where
it is desired to physically localize packets.
The inset detail in Figure 8 illustrates amongst other things
4.1.5 Economics: The complexity and projected cost of that a single NIC is used to interface to the network. This NIC,
both the DQSA NIC and the QHub will be comparable to effectively as simple and thus as inexpensive as a DQSA LAN
equivalent Ethernet components. The same physical media NIC or Ethernet LAN NIC, provides in conjunction with the
used to support Ethernet will be used in a DQSA LAN. The DQSA NICs at all the other sites all network control. Nothing
standard IEEE 802.3/Ethernet frame will be used and so little else is required. Practicality dictates that there will be a net-
or no change is required to upper layer protocols. Any changes work manager residing at HQ to monitor the allocation of
required will be desirable in that they will reflect support for a guaranteed bandwidth, gather traffic measurements, etc. The
QoS not now possible with Ethernet. entire network, spokes and sub-spokes, can be a single speed,
e.g., DS1, or the main spokes could be say DS3 while the local
4.1.6 Summing Up: DQSA achieves 100% utilization of loops or some of the sub-spokes can be lower speed, e.g., DS1.
available data slots, typically over 90% of total bandwidth.
However, if great bandwidth utilization was important then Implementing DQSA over STM circuits actually offers a great
when Token Ring arrived on the scene it would have immedi- advantage with respect to utilization -- the synchronization
ately displaced Ethernet. The marketplace suggests that sim- available in STM at the bit level means that two bits suffice for
plicity and the low-cost that simplicity usually brings to the a minislot, e.g., a distributed ATM switch using DQRAP
table are the key factors. A rudimentary QoS, where required, requires only a single byte (8 bits) to carry three minislots plus
is achieved with Ethernet, much to the chagrin of Token Ring priority. This means a utilization of 424/(424+8) = 98.1%. The
advocates, by throwing bandwidth at the problem. But as only disadvantage to implementing DQSA over STM, e.g.,
pointed out above, a DQSA LAN provides simplicity along DS1, is that whereas a single and-gate is all that is required to
with its accompanying economic benefits as well as a QoS that merge two DS1 circuits, synchronization of the underlying
even Token Ring cannot provide. And great utilization and DS1 frames is required before merging. Implementing DQSA
distance insensitivity comes with the rations. on “dark” circuits increases the bandwidth used by minislots
and dataslots in that guardbands, preamble, etc., are required,
A proof-of-concept 10 Mbps DQSA LAN is in operation and
but the “joining” of two circuits is now accomplished with
confirms both theory and simulation results.
only an “or” operation. Thus a network implemented on dark
fiber could be built using only passive taps and splitters, if the
4.2 DQSA in Wide Area Networks power budget is satisfied. This suggests that all the silicon
devoted to creating and maintaining synchronous timing plus
DQSA in WANs introduces a paradigm shift in communica- the DSx and OCx framing could be exchanged for bandwidth.
tions since the use of shared access has not even been contem-
plated since the retirement in the 1980s of multi-point WANs SONET (synchronous optical network) is an attempt to move
utilizing 2.4 Kbps - 9.6 Kbps analog modems. We could now away from conventional switches towards the economy and
enumerate the benefits of a DQSA WAN in a manner similar simplicity of shared access, i.e., fiber in a ring topology.
to those presented in Section 4.1 about DQSA LANs. Instead SONET interfaces to the STM (synchronous transfer method)
we simply refer the reader to that section because a DQSA world employing OCx whose 125 microsecond frames result
WAN is simply a DQSA LAN writ large. In the main we will in a granularity of thousands of bytes, not conducive to effi-

10
Single Chip
Qnodes

Fiber Optic
Add-Drop

Metropolitan/Regional
DQSA Networks

Figure 9. CDQ Cross-Country Distributed ATM Network

ciency with respect to shared access. In contrast a fiber-optic single building, a campus, or continent-wide as shown in Fig-
ring, possibly called ASONET (Asynchronous Optical Net- ure 9 where conventional DQSA networks are connected as
work), based on DQSA could switch individual ATM cells via sub-spokes to the main spoke connecting two Qnodes. The
add-drop optical taps in and out of the ring at gigabit speeds. number of networks/attachments is only limited by the capac-
Chang has simulated CDQ in a ring topology [22]. ity of the segment. The connecting points could be as simple
as fiber-optic add-drop units if there is no speed change.
DQSA WANs can be implemented by the same corporations/ Access to the network via lower-speed networks would utilize
organizations that utilize existing WANs. The implementa- small buffers. Egress could present the classic problem of too
tions could be private networks using leased circuits or public much arriving at high speed for a lower speed sink but even
networks offering ATM or Frame Relay service. There is an this problem will be taken care of by the CDQ flow control
ongoing move away from private networks to the utilization of mechanism. A CDQ network will operate as a distributed
commercial Frame Relay/ATM networks or the Internet, in ATM switch with performance superior to conventional ATM
both cases possibly using VPNs (virtual private networks). networks in that upwards of 90% utilization can be expected
The illustration of Figure 8 shows a network that could be without dropping cells excepting due to the BER of the line.
established using private leased circuits, or it could also repre- Aside from a very good “best effort” performance and even
sent a network established by one of the carriers, e.g., AT&T, better “controlled load” performance using PDQ priorities,
using circuits dynamically assigned to this particular cus- guaranteed bandwidth channels can be established across the
tomer. country that will support isochronous requirements.

A DQSA WAN can provide a level of service with respect to A CDQ backbone could operate in the gigabit/s range and the
delay, throughput, and guaranteed bandwidth not possible with segments can be formed into a ring. As with any of the DQSA
existing networks. A carrier could actually provide a customer protocols, if the power budget permits then an entire fiber net-
with a network consisting only of the carrier’s circuits, i.e., no work, excepting for the Qnodes, could be passive.
routers, and possibly charge a premium because of the level of
the service guarantee. Guarantees with respect to uptime can CDQ does make DQSA an architecture for all seasons.
be based solely on the reliability of the underlying circuits.

A proof-of-concept T1 DQSA system is operating and meets 4.4 DQSA and the Wireless World
all expectations with respect to performance. All the good stuff said about DQSA applies to the wireless
world, but we concentrate on just one facet. The availability of
4.3 DQSA as a Backbone an efficient shared access mechanism opens up the possibility
for true convergence of data and voice, i.e., packet telephony.
CDQ lends itself to a backbone operation, whether it be for a An effective integrated system requires that voice be carried in

11
packets, but in most current voice systems the lack of an effec- deployment of any cable access indicates that neither is opti-
tive MAC has meant that some type of TDM (time division mal. The author did discover that one should not attend a Stan-
multiplexing) must be used thus making it very difficult to dards dance unless one has a partner with clout.
integrate voice and data. Implementing ATM in the wireless
world is an obvious solution, but again the existing access 4.5.4 DQSA and xDSL: One of the more exciting appli-
methods are so inefficient that ATM cells must be “bundled” cations for DQSA is that it will permit clusters of DSL (digital
to achieve any sort of efficiency thus obviating much of the subscriber loop) lines to be treated as a single network. DSL
advantage of ATM. A packet telephony system based on allows up to several megabits/s to be transmitted over the ordi-
DQSA is the solution. nary telephone wires serving homes and businesses. The
power of DQSA means that true integrated services with
We point out that a DQSA wireless system utilizing DQRAP/ packet telephony could be offered over the ubiquitous copper
ATM or XDQRAP/IP provides all the goodies described in pairs. The DQSA hub could in turn be connected to a DQSA
Section 4.1. Yes, there is the overhead problem for the minis- backbone and thence to other neighbourhoods and communi-
lots, a non-trivial problem in the wireless world with pream- ties thus establishing a network for all communications ser-
ble, sync, and other requirements but this overhead will vices that bypasses the conventional switched system.
consume less than 20% of the bandwidth, yielding at least
twice the throughput of any other system when the traffic is 4.5.5 DQSA and RFID: The efficiency of DQSA in
bursty. Lin demonstrates the performance of DQSA when sub- resolving contention means that it is ideal for RFID (radio fre-
jected to voice traffic in [12]. quency identification) systems. We present an example that
also shows the efficiency of the QCR (queueing contention
resolution) subsystem.
4.5 Odds-and-Ends Assume there are 10,000 items in a warehouse, each with its
An efficient shared access method quite literally opens up the RFID implementing DQRAP. Assume that a relatively low-
entire world of communications but we must cut it off some- speed, low cost, radio system utilizes a time slot of 60 milli-
where. We confine ourself to listing a few more potential seconds which includes outbound beacon and feedback plus
applications along with a brief note on each. the inbound minislots and dataslot. When inventory is to be
taken, the master unit issues a beacon to start the process and
4.5.1 DQSA and the Parallel Bus: All the bus arbitration then repeats it every 60 ms. All 10,000 items respond to the
schemes, e.g., ISA, EISA, PCI, etc., are shared access net- first beacon. But after the master controller sends out the feed-
works writ small. A short study of applying DQSA to a paral- back along with the second beacon the DQRAP contention
lel bus arbitration scheme carried out by Wu indicates that it resolution process kicks in and the number of responses drops
could have better performance characteristics than any system to approximately 3333 items in the second round. In succeed-
extant [21]. The main benefit is that it is truly distributed, e.g., ing rounds the number of responses will drop by approxi-
a master arbitrator or controller is not required. One hundred mately two-thirds each time: 1111, 370, 123, 41, 14, 5, and
percent of the bus bandwidth is available either being accessed then 2. Thus after nine cycles, approximately 540 ms, the first
randomly in single cycles or in blocks. As with all DQSA both couple of items will be successful in the minislots and thence
priorities and reserved channels are available. move to the data transmission queue. The remaining 9,998
items will then resolve their contention in parallel with the
4.5.2 DQSA and VOD Mass Storage: VOD (video-on- ongoing data transmission and join the distributed data trans-
demand) has been looming on the horizon for some time and mission queue, The entire inventory will be complete in little
perhaps one of these years it will be arrive. A DQSA network more than 10,000 x 60 ms = 600 seconds.
will be the ideal way to deliver a movie-on-demand whether it
be in a single building or a metro area. However our interest O’Connell investigated the contention resolution capability of
here is in the storage mechanism itself. Low-cost storage DQRAP by simulation of situations where up to 100,000 arriv-
devices, e.g., DVD, make feasible massive terastores with als contend in a single arrival period [23].
hundreds of DVD players where each player has a separate
micro-controller that interfaces to a DQSA bus. Each control- 4.5.6 Wrap-up: We will stop listing potential applica-
ler will accept and respond to requests for some 30 - 40 hours tions at this point since by now we hope that the message is
of compressed video. The viewer after making a selection will clear: once shared-access is viewed as a viable alternative then
interact directly with a controller and be permitted to fast-for- virtually all communications applications are subject to
ward, freeze, pause, etc., everything that can be done on a rethinking with respect to how best they can be implemented.
VCR. DQSA is the only shared access method that would sup-
port a distributed system such as is described here.
5 Conclusions
4.5.3 DQSA and Cable TV: A natural for DQSA and yet
it is on the sidelines except for the previously mentioned appli- The rise of the Internet has been dramatic even by the stan-
cation of the concept of three minislots to the IEEE 802.14 dards of an industry that over the past thirty years has become
standard. However the fact that two “standards”, MCNS and accustomed to individual segments literally exploding. The
IEEE 802.14, exist before there has been any significant growth will not slow down since more and more services are

12
moving to the Internet, the most significant of which are voice ber sufficient to suggest that shared access can play a
and video. Naysayers opine that the Internet will never have major role in all aspects of communications.
the capacity nor capability with respect to quality to make
We conclude by saying that, contrary to reports, shared access
major inroads on conventional voice traffic or to provide reli-
is alive and well and will be marching briskly into the new
able video, but the process is under way. The current and
millennium.
expected growth prompted one eminent personage to state
publicly what many felt in private, that the Internet would suf-
fer a meltdown. This has not happened as yet in the main References
because silicon and bandwidth are being thrown at the Internet [1] W.Stallings, Data and Computer Communications, Fifth
in sufficient quantities to keep ahead of the curve. However Edition, Prentice Hall, 1997.
these are brute force methods and will not be able to keep pace
with demand forever -- the time frame of the Internet. [2] IEEE 802 Web Site. www.stdsbbs.ieee.org/groups/802/
overview.html
The author argues that the cost of bandwidth is dropping faster
than the cost of silicon switching, thus the optimum strategy [3] Specifications available at www.cablemodem.com
for staying ahead of the demand curve is to, where possible,
substitute bandwidth for silicon. The shared access approach [4] W. Stallings, Local and Metropolitan Area Networks, Fifth
to communications is effectively just that, trading bandwidth Edition, Prentice Hall, 1997.
for silicon. We also argue that the major benefit of this
approach will accrue not from less silicon but from the lesser [5] W. Xu and G. Campbell "DQRAP - A Distributed Queue-
network management requirements. ing Random Access Protocol for a Broadcast Channel", pre-
sented at SIGCOMM '93, San Francisco, September 14, 1993.
The ability of DQSA to support isochronous traffic means that Computer Communication Review, Vol 23, No. 4, Oct 1993,
DQSA distributed switches could support both random access pp. 270-278.
and conventional DSx circuits during any transition period.
DQSA allows single ATM cells to be switched at any speed [6] C.T. Wu and G. Campbell, "Extended DQRAP
over any distance on either synchronous circuits that utilize (XDQRAP): A Cable TV Protocol Functioning as a Distrib-
DSx or OCx framing or “dark” circuits where each transmis- uted Switch", Proceedings of 1st International Workshop on
sion is self-contained in that preamble, sync, etc., are required. Community Networking, July 1994, San Francisco.
STM circuits simplify the implementation of DQSA because
the underlying synchronization allows minislots to be only a [7] C.T. Wu and G. Campbell, "DQLAN - A DQRAP Based
few bits long and permits back-to-back transmissions by dif- LAN Protocol”, Proceedings of the 1st Workshop on High-
ferent stations. The author estimates that implementing DQSA Speed Network Computing, 9th Int'l Parallel Processing Sym-
on “dark” circuits increases overhead by about 10% to satisfy posium, Santa Barbara, CA, April 1995.
guardband, preamble and sync requirements but, the payoff is
elimination of out-of-band STM overhead. The suggestion [8] J. L. Massey, “Collision resolution algorithm and random
was made in section 4.2 but here is the question: “Is it worth access communications,” in Multiuser Communication Sys-
giving up say 10% of the bandwidth if it means that DSx and tems, G. Longo Ed. N.Y.: Springer-Verlag, 1981. pp. 73-137.
OCx framing could be eliminated?” The two main existing
communications mechanisms, STM switched-circuits and [9] C.T. Wu and G. Campbell, "Interleaved DQRAP With
routed/switched packets, are exemplified by the products of Global TQ", DQRAP Research Group Report 94-4, 1994.
Lucent and Cisco. These two spend much time eyeing each
other warily and attempting, as voice becomes interchangeable [10] H. J. Lin and G. Campbell, "PDQRAP - Prioritized Dis-
with data, to poach in each other’s territory. Perhaps each tributed Queueing Random Access Protocol", Proc. of 19th
should instead be looking over their respective shoulders. Conference on Local Computer Networks, Oct. 1994, pp 82 -
91.
The performance of contention-based shared access methods,
until now, has been poor but there is no law of nature that [11] C. T. Wu and G. Campbell "CBR Channels on a DQRAP-
states that performance which approaches that of an ideal sys- based HFC Network", SPIE '95 (Photonics East), Philadel-
tem is not possible. Summing up: phia, PA Oct. 1995.

• A set of ten criteria is presented that describes the [12] H.J. Lin and G. Campbell "Using DQRAP (Distributed
ideal shared access system. Queueing Random Access Protocol) for Local Wireless Com-
• A set of shared access protocols that constitute the munications." Proceedings of Wireless '93, July 14, 1993, pp.
DQSA family is presented. It is further shown that the 625-635.
DQSA protocols satisfy or come close to satisfying the
ten ideal criteria. [13] M. Zhang and G. Campbell, "A Performance Analysis of
• It is shown how the DQSA protocols can be deployed Distributed Queueing Random Access Protocol", DQRAP
in a number of communication applications, a num-

13
Research Group Report 93-1, 1993. Arbitrator”, Unpublished, IIT, 1994.

[14] M. Miramica and G. Campbell, “Robustness Analysis of [22] C.H. Chang, “A Cascaded Distributed Queue Network”,
the DQRAP Protocol”, DQRAP Research Group Report 93-6, Ph.D. Dissertation, IIT, 1998 (Not Published)
1993.
[23] W. O'Connell, "Can Shared Access Networks Adequately
[15] Paul Desmond, “Happy Birthday Ethernet”, Network Support High-Performance Computing? Re-Visiting MAC
World, May 18, 1998, p 40. Protocols", Proceedings of the 1st Workshop on High-Speed
Network Computing, 9th Int'l Parallel Processing Symposium,
[16] C. Bisdikian, B. McNeil, R. Norman, R. Zeisz, “MLAP: Santa Barbara, CA., April 1995.
A MAC Level Access Protocol for the HFC 802.14 Network”,
IEEE Communications Magazine, pp. 114-121, March 1996 [24] Microsoft Professional Developers Conference, San
Diego, Sep 22-26, 1997.
[17] C. T. Wu and G. Campbell, "The Extended DQRAP
(XDQRAP) Algorithm", DQRAP Research Group Report 94- [25] J. I. Capetanakis, “Tree Algorithm for Packet Broadcast
3, 1994. Channels”, IEEE Transactions on Information Theory, Vol.
IT-25, no. 5, pp. 505-515, Sep. 1979.
[18] M. McPheters, “Approaching Reality with a Finite Net-
work Node Model Analysis of DQRAP (Distributed Queueing [26] F. Koperda and B. Lin, “Providing CBR Service over
Random Access Protocol)”, Ph.D. Dissertation, Illinois Insti- XDQRAP”, IEEE 802.14/95-147, Scientific Atlanta, 1995.
tute of Technology, 1993.
[27] Information available at www.homeRF.org.
[19] R. Khasawneh, “A DQRAP Based Residential-Metropoli-
tan Area Network Utilizing Cable TV Facilities”, Ph.D. Dis- [28] L. Kleinrock, “On Queueing Problems in Random-Access
sertation, Illinois Institute of Technology, 1993. Communications”, IEEE Trans. Inf. Theory, Vol IT-31, No. 2,
March 1985, pp 166-175
[20] Byung. Lee, “Interleaving and DQRAP (Distributed
Queueing Random Access Protocol) Based Networks Without NB: All DQRAP Research Group Reports plus simulation
Immediate Feedback”, Ph.D. Dissertation, IIT, 1996. programs, etc., and this paper are available at www.iit.edu/
~dqrap.
[21] C.T. Wu and G. Campbell, “DQRAP as a Parallel Bus

14

You might also like