Heikki Tahvanainen

OPC UA performance evaluation

School of Electrical Engineering

Thesis submitted for examination for the degree of Master of

Science in Technology.
Espoo 14.3.2016

Thesis supervisor:

D.Sc. Ilkka Seilonen

Thesis advisor:

M.Sc. (Tech.) Jouni Aro

aalto university abstract of the
school of electrical engineering masters thesis

Author: Heikki Tahvanainen

Title: OPC UA performance evaluation

Date: 14.3.2016 Language: English Number of pages: 7+56

Department of Automation and Systems Technology
Professorship: Information and computer systems in automation
Supervisor: D.Sc. Ilkka Seilonen
Advisor: M.Sc. (Tech.) Jouni Aro
OPC Unified Architecture is an industrial communication specification, employing
a client-server communication model. In this thesis, performance testing of OPC
UA applications is explored.
First, performance testing of software applications is discussed on a general level.
After this, details of OPC UA and previous studies of OPC UA performance are
examined. Based on this information, example test cases are designed. The testing
procedure and results of the tests are presented in detail. As an important part
of the work, the design and implementation of a generic test client for read and
subscription testing is presented.
The results provide information about overhead of OPC UA security modes in two
different hardware platforms and comparison of two server applications in terms
of read and subscription functionality. Substantial differences can be observed in
terms of maximum number of requests as well as in the resulting resource usage.
When overloaded, the servers again exhibit differences in their behaviour. The test
results should be thought of as examples of features, limits and differences that
can be observed by performance testing OPC UA applications. The results of this
thesis are used in the future development of the Prosys OPC UA Java SDK.

Keywords: OPC UA, performance testing

aalto-yliopisto diplomityn
shktekniikan korkeakoulu tiivistelm

Tekij: Heikki Tahvanainen

Tyn nimi: OPC UA suorituskyvyn arviointi

Pivmr: 14.3.2016 Kieli: Englanti Sivumr: 7+56

Automaatio- ja systeemitekniikan laitos

Professuuri: Automaation tietotekniikka

Tyn valvoja: D.Sc. Ilkka Seilonen
Tyn ohjaaja: TkT Ilkka Seilonen
OPC Unified Architecture on erityisesti teollisiin ympristihin suunniteltu tiedon-
siirtomrittely, joka kytt asiakas-palvelin kommunikointimallia. Tss tyss
tutkitaan OPC UA -sovellusten suorituskykytestausta.
Ensin muodostetaan yleisell tasolla ksitys tavoista mitata ohjelmistojen suoritus-
kyky. Tmn jlkeen OPC UA -mrittelyn yksityiskohtia kydn lpi thn
tyhn liittyvin osin. Aiemmin aihepiirist julkaistujen tutkimusten avulla muo-
dostetaan kuva siit, mit ja miten oikeastaan kannattaa tutkia. Suoritettavat
testit dokumentoidaan mahdollisimman hyvin ja testien tulokset raportoidaan
yksityiskohtaisesti. Trken osana tyt esitelln yleiskyttisen testisovelluksen
suunnittelu ja toteutus.
Tutkimuksen perusteella esitetn tietoa OPC UA -viestien tietoturvalliseen k-
sittelyyn kuluvasta resurssien kulutuksesta kahdella eri laitteistolla ja vertaillaan
kahden OPC UA -palvelintoteutuksen ominaisuuksia. Toteutuksissa voidaan havai-
ta selkeit eroja ksiteltyjen pyyntjen maksimimrss sek palvelinohjelmistojen
resurssienkytss. Kun palvelimiin tehdn enemmn pyyntj kuin ne ehtivt
ksitell, on tuloksena taas selkeit eroja eri implementaatioiden vlill. Esiteltvt
testitulokset ptevt tss tietyss konfiguraatiossa ja ne tulee ksitt esimerk-
kein siit minklaisia ominaisuuksia, rajoituksia ja eroja OPC UA -sovelluksissa
on havaittavissa. Tyn tuloksia tullaan kyttmn Prosys OPC UA Java SDK:n
jatkokehityksess.

Avainsanat: OPC UA, suorituskykytestaus

 iv

Preface
I would like to thank my instructor Jouni Aro for providing the thesis topic and for
his support during the whole process. Another thanks goes to Ilkka Seilonen who
operated as my supervisor. All the comments and improvements were appreciated.
This thesis has been made at Prosys PMS Ltd, Espoo, Finland. Everyone at Prosys
deserves a thanks for providing a good atmosphere to work in. Special thanks goes
to my sister Hanna, who helped with the valuable job of proofreading.

Otaniemi, 3.3.2016

Heikki Tahvanainen
 v

Contents
Abstract ii

Abstract (in Finnish) iii

Preface iv

Contents v

Abbreviations vii

1 Introduction 1
1.1 Background and motivation . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives and scope . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Research methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Structure of the work . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Performance evaluation of software systems 3

2.1 Measuring application performance . . . . . . . . . . . . . . . . . . . 3
2.2 Usefulness of performance testing . . . . . . . . . . . . . . . . . . . . 5
2.3 Examples of other performance testing frameworks . . . . . . . . . . 7
2.3.1 httperf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Faban . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 OPC UA 8
3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 Subscription model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.5 Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 Previous studies of OPC UA performance 16

4.1 Speed indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Utilization indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Protocol comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Test requirements 24
5.1 Elapsed time (batch) measurements . . . . . . . . . . . . . . . . . . . 24
5.1.1 Impact of signing and encrypting . . . . . . . . . . . . . . . . 25
5.2 Response time measurements . . . . . . . . . . . . . . . . . . . . . . 25
5.2.1 Read mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2.2 Subscription mechanism . . . . . . . . . . . . . . . . . . . . . 26
 vi

6 Research material and methods 27

6.1 Design of response time test client . . . . . . . . . . . . . . . . . . . . 27
6.1.1 Read testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6.1.2 Subscription testing . . . . . . . . . . . . . . . . . . . . . . . . 29

7 Results 33
7.1 Elapsed time results . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.2 Response time results . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2.1 Test setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2.2 Read mechanism . . . . . . . . . . . . . . . . . . . . . . . . . 37
7.2.3 Subscription mechanism . . . . . . . . . . . . . . . . . . . . . 44

8 Conclusions 51

References 53
 vii

Abbreviations
Abbreviations
OPC OLE for Process Control
OPC UA OPC Unified Architecture
SCADA Supervisory Control And Data Acquisition
SDK Software Development Kit
GUI Graphical User Interface
LAN Local Area Network
XML Extensible Markup Language
PLC Programmable Logic Controller
CA Certificate Authority
QA Quality Assurance
JVM Java Virtual Machine
JDK Java Development Kit
1 Introduction
1.1 Background and motivation
Information transfer is an essential factor in todays automation systems. Currently,
industrial networks are often characterized with automation pyramid approach. [1]
OPC Unified Architecture (abbreviated OPC UA) is a software specification which
creates a common infrastructure for sharing information between different systems.
The most common use case is an OPC UA server acting as a gateway exposing
data from underlying data source to OPC UA clients residing on higher level of
this conceptual automation pyramid. OPC Classic, UAs predecessor, enabled data
from automation instruments of different manufacturers to be accessed using one
protocol. OPC UA is an evolution of this previous standard and an attempt to
create an extensible specification to expose and access data regardless of the purpose
or meaning of the data, enabling interoperability between any number of systems
of different purposes [2]. OPC UA is an IEC standard IEC 62541 [3]. OPC UA
specification addresses data communication in a client-server architecture and also
defines information modeling in the form of an address space model, server discovery
services, security modes and subscription mechanism [4]. Among other things, the
actual data communication is an important feature of OPC UA specification.
OPC UA is already a mature specification used in industrial environments.
Things like scalability and performance are growing in importance as more use
cases are explored. When developing a product or service utilizing OPC UA data
communications it is a matter of importance for the developer and user to know
the possible performance constraints. The importance of performance studies can
be seen by looking at recent studies where, for example, OPC UA performance has
been studied in the context of an agricultural system [5] and in the context of smart
electric grids [6]. It seems that performance studies are carried out at the same time
when other study goals are the main objective. Thus, there is a need for a study
focusing on OPC UA performance.
The evaluation of performance aspects provides valuable information for a system
designer or for end-users. Mostly, end-users are not interested in detailed information.
They want to know if a given service may be used for their application scenario,
and if so, which parameters seem to have the biggest impact for communication.
Using results from pre-tested platforms, system designers can quickly gauge the
level of hardware performance they require and possibly eliminate the need to test
systems themselves. [7] With historical benchmarking data, it is also possible to show
for example an improvement in design and implementation when developing OPC
UA products. Organizations that are currently migrating their existing production
systems to OPC UA are interested in performance figures. [8] One aspect regarding
this migration is that the new system must preserve or increase the quality of
communication performance. For this reason some organizations must run their
own performance tests when considering migration. These organizations themselves
declare that they would benefit from performance tests that would help them choose
between protocols.
 2

1.2 Objectives and scope

This thesis explores performance testing of OPC UA applications considering es-
pecially server application testing. The work will be done considering specially a
proprietary SDK developed for Java by Prosys PMS Ltd [9]. The objective of this
thesis is to answer some of the performance related questions that system designers
using OPC UA will face. To meet the goals of answering OPC UA performance
evaluation questions, the thesis should provide information on how tests were devel-
oped as well as providing generic benchmarking results where possible. In the case of
OPC UA, only a few ready made performance testing tools exist so it is necessary to
device one during this thesis. Actual implementation of a publicly available generic
performance testing tool is not in the scope of this thesis. However, the presented
information can be used as a base knowledge for future development. The main
research questions of this thesis are: 1) what kind of features and limits can be
observed in OPC UA applications by performance testing and 2) what kind of issues
surround performance testing of OPC UA applications.

1.3 Research methods

The main sources for information in this work have been previous studies of OPC
UA performance and existing performance testing tools of OPC UA and other
information systems. Relatively many publications exist on the topic of OPC UA
performance. Previous studies provide information on OPC UA performance as well
as help to understand which areas will be most fruitful for further studying. This
thesis concentrates on experimental testing. Ready made performance testing tools
exist for many information systems (for example web servers [10]). The existing
performance testing tools are examined as examples.

1.4 Structure of the work

This thesis is organized into multiple sections. First, a look at performance testing
of software systems in general is presented. Next, a brief introduction into OPC
UA specification is provided. After this, a look at previous studies of OPC UA
performance evaluation is made. Areas that need further studying in this topic
are selected based on previous research. After the interesting topics are selected,
practical aspects of developing a test client are documented in the chapter research
material and methods. Then the results chapter documents the obtained results. At
the end, conclusions chapter presents the most important results in a compact form
and answers the research questions decided here in the beginning.
 3

2 Performance evaluation of software systems

2.1 Measuring application performance
Performance Testing refers to testing done to analyze and improve the performance
of an application. The focus here is on the optimization of resource consumption
by analyzing data collected during testing. Performance Testing to a certain extent
should be done by developers, but more elaborate, large scale testing may be con-
ducted by a separate performance team. In some organizations, the performance
team is a part of the QA function. Sometimes performance testing that is focused
on understanding how an application scales is called Scalability Testing. As such
scalability testing can be viewed as a kind of performance testing where the goal is
to understand at what point the application stops scaling and identify the reasons
for this. Scalability can be described in many ways, ranging from loose definitions
referring to softwares ability to extract the most performance out of the available
resources, to more strict definitions like the ability to improve throughput or capacity
when additional computing resources are added. Load Testing refers to the kind of
testing usually done by QA organizations to ensure that the application can handle
a certain load level. Criteria are set to ensure that releases of a product meet certain
conditions, like the number of users they can support while delivering a certain
response time. The best term for tests presented in this thesis is performance testing.
Performance measurement can be defined as the process of quantifying the
effectiveness and efficiency of an action. Effectiveness is a measure of the level of
performance and efficiency is a measure of how economically resources are utilized
when providing a given level of performance. For example, if a server application
handles 1000 requests per seconds and consumes 50% of the host machine CPU
time, then the 1000 requests per second is a measurement of the effectiveness of the
application and CPU usage is a measure of the efficiency. Performance is related
to both time (speed of communication) and space (load and resources). These
classifications are sometimes called speed and utilization indices [11]. The distinction
between the two classes of measures is that the speed indices provide an absolute
measure of the performance and utilization indices provide a basis to understand
where the bottlenecks are. For example by measuring network bandwidth utilization
it can be shown that bottleneck is not the network, but internal architecture of
the server and utilization of other hardware resources [12]. Commonly measured
effectiveness metrics are for example system throughput (number of requests handled
per time period) and response time for a request. In addition to these, performance
study might also report efficiency metrics, including measures such as CPU utilization
and communication line utilization. Evaluated utilization metrics can also be memory
or disk space, message size, queue lengths or even the energy consumed by different
services [13]. This classification is illustrated in table 1.
Table 2 specifies a way to categorize performance tests based on their level of
abstraction. Application software tests compare the performance or functionality
of applications or application bundles. System software tests provide the basis for
evaluation and comparison of software that serves applications and is not necessarily
 4

Table 1: Classification of benchmark results [11].

speed indices (effectiveness) utilization indices (efficiency)
System throughput CPU usage
Response time for a request Communication line usage
Time needed to activate a session Length of systems queues
Memory usage
Message size
Energy consumed

directly used by an end user. OPC UA communication SDKs, which are the basis for
this thesis, belong to this system software category. Micro-benchmarks can be used
to compare for example two different algorithms that accomplish the same thing.
Same kind of classification is presented in the context of hardware testing.

Table 2: Categories for computer system performance testing [11].

Main category Sub category Description
Software Application software Compare performance and func-
tionality of specific applications
or application bundles, e.g. mail
servers
System software Compare performance and func-
tionality of software that serves
applications, e.g. database sys-
tems
Micro-benchmark Measure the performance of a very
small and specific piece of software
Hardware System Assessment of a system in its en-
tirety
Component Assessment of specific parts of
a computer system (for example
graphical processing unit)
Micro-benchmark Measure the performance of a very
small and specific piece of hard-
ware

System software testing, the most suitable classification for tests done in this
thesis, can further be divided into multiple categories. Some tests measure "how fast"
a given unit of work can be processed or acknowledged; others measure "how much"
work can be performed with a given quantity of computing resources [14]. A usual way
to classify performance tests is to categorize them to Elapsed time, Throughput
or Response time tests [15]. The simplest way of measuring performance is to select
a certain task and see how long it takes to accomplish. This is called elapsed time
or batch measurement. These measurements measure the amount of time (or other
resources) that must be used to accomplish certain workload. Conversely, throughput
 5

measurements measure the amount of work that can be accomplished in a certain

period of time. The most usual throughput test in client-server applications is that
a client sends a request to the server. When it receives a response, it immediately
sends a new request. That process continues; at the end of the test, the client reports
the total number of operations that it achieved. [15, pp. 25] One defining factor is
that in throughput tests, the workload is not fixed.
Contrary to throughput measurements, in response time tests, the effectiveness
of the server is based on how quickly it responds to a fixed load. The most important
factor to measure is the amount of time that elapses between the sending of a request
from a client and the receipt of the response. The difference between a response time
test and a throughput test is that clients in a response time test sleep for some period
of time between operations. This is referred to as think time [15]. When think time
is introduced into a test, throughput becomes fixed. At that point, the important
measurement is the response time for the request. The effectiveness of the server is
based on how quickly it responds to the fixed load. These different categorizations
are illustrated in table 3.
Usually response time is reported as an average value or as a percentile. For
example, if 90% of responses are less than 1.5 seconds and 10% of responses are
greater than 1.5 seconds, then 1.5 seconds is the 90th percentile response time. If
data contains huge outliers, then percentile values may give a more accurate view
than average value. Huge outliers are rare in general, but smaller outliers may occur
because of multiple reasons. In Java applications garbage collection introduces pause
times which may easily cause outliers in measurements.

Table 3: Test categorization

Elapsed time Throughput Response time
Workload Constant Variable Constant
Throughput Variable Variable Constant

2.2 Usefulness of performance testing

Best performance results would of course be obtained by profiling production ap-
plications, however this may be impossible in most cases. The other possibilities
are empirical approach (making test application) or theoretical (using simulation to
obtain results). There is a tradeoff here: differences observed in tests run with actual
applications show results that are results of different implementations of certain
protocol but may be platform dependent. With theoretical or simulated results
there is no real guarantee of applicability of results. Also, obtaining reliable results
by profiling an application is by no means trivial. Performance testing is a usual
practice. However, there is always a risk of wrong performance testing results. For
this reason, it is important to concentrate on getting useful results and to be careful
in order to avoid potential errors. The following list contains aspects which can be
used to assess the usefulness of performance testing [16].
 6

Relevancy The results reflect important issues of the domain.

Repeatability The test is able to run multiple times yielding the same result.

Fairness All the compared systems are able to participate equally.

Verifiable There is confidence that the documented result is real.

Economical Users afford to run the test.

Not all of the criteria have to be fulfilled in perfection, but most successful
benchmarks are strong in one or two of these aspects [16]. Often, in order to satisfy
the last four of these items, a test developer must choose to give up on some of the
first. This means that creating economical, verifiable, fair and repeatable tests often
lowers its amount of relevancy in practice. Relevant tests have among other things
understandable metric and a target audience that wants the information. Perhaps
the most important requirement is appropriate use of software. The software features
should be used in the way that a typical customer application would. Appropriate
representation is an important factor also. A performance test result is not very useful
if there is not a high degree of confidence that it represents the actual performance
of the system under test.
Repeatability can become an issue when measuring performance [17]. For example,
the amount of logging usually has a substantial effect on performance (this is also the
reason why applications usually have different log levels). Managed runtime systems,
such as Java, have their own quirks. Just-in-time compilation of Java causes the
identical code to perform more effectively over time and this means that certain
warm-up period must be taken into account in order to get reliable results. Java
applications also build up garbage in the Java heap that must be cleaned out, which
then introduces regular garbage collection pauses. Performance testing should try to
ensure repeatability and consistency. There is a trade-off between repeatability and
reality. One of these trade-offs is the creation of a steady-state period within the
benchmark. Real applications are hardly steady in the way that they generate work
on the system, but a benchmark where results will be compared requires either that
the application and associated performance does not change over a period of time or
that the exact same work flow runs for each iteration of the benchmark.
Testing client-server applications is cumbersome because it is not necessarily clear
whether the client side or server side is the bottleneck. This can lead to results that
are hard to verify. All client-server tests run the risk that the client cannot send
data quickly enough to the server [15]. This may occur because there are not enough
CPU cycles on the client machine to run the desired number of threads, or because
the client has to spend a lot of time processing the request before it can send a new
request. In those cases, the test is effectively measuring the client performance rather
than the server performance, which is usually not the goal. The risk depends on the
amount of work that each client thread performs. A throughput-oriented test is more
likely to encounter this situation, since each client thread is performing a lot of work.
The relevancy of results needs to be considered especially in the case of multiple
measurements. It is common for tests that measure throughput also to report the
 7

average response time of its requests. Changes in that number are not indicative of a
performance problem unless the reported throughput is the same. A server that can
sustain 500 operations per second with a 0.5 second response time is performing better
than a server that reports a 0.3 second response time but only 400 operations per
second. [15] Throughput measurements are usually taken after a suitable warm-up
period, particularly since what is being measured is not a fixed set of work.

2.3 Examples of other performance testing frameworks

2.3.1 httperf
When considering tools for performance measurement, examples could be taken from
other information systems. For example database systems, computer networks, email
servers and http servers all have performance measurement tools which are developed
to measure values specific for the information system at hand. One example of
long-lived web server testing software is httperf [10]. Httperf has been originally
implemented in 1998 so as of writing this it has existed for 17 years. Considering
that the web is a highly dynamic system and subject to relatively frequent and
fundamental changes it is a major accomplishment for a tool to stay topical for this
long. Surely there are also more modern tools that are more capable, but nevertheless
httperf is still widely in use and can thus be used as an example design. To put it
simply, httperf overloads a server beyond its normal capacity. Using the tool, you
saturate the server with TCP connections. Each httperf test will consist of a session
(or sessions) made up of calls spaced out at certain intervals.
This same kind of server performance testing tool could be useful in the case
of OPC UA servers. One of the premises in the design of httperf has been the
separation of performing HTTP calls from issues such as what kind of workload and
measurements should be used [10]. In the same sense, it should be beneficial to have
a common tool which could be used to test multiple different UA servers.

2.3.2 Faban
Faban is free and open-source, Java-based load generator. It can be used to create
and execute workloads. [18] Faban comes with a fhb program, that can be used to
measure the performance of a URL [15].
fhb W 1000 r 300/300/60 c 25 http : / / h o s t : p o r t / t e s t u r l
This example measures 25 clients making requests to the server. Each request has a
1-second cycle time. The benchmark has a 5-minute (300-second) warm-up period, 5
minute measurement period and a 1-minute ramp-down period. Following the test,
fhb reports the OPS and response times for the test. Because there is think time
in this example, the response times will be the important metric and operations
per second will be constant (unless the server is overloaded by this workload). In
addition to the fhb program, Faban has a framework for defining benchmark load
generators in Java.
 8

3 OPC UA
3.1 Overview
OPC Unified Architecture is already a mature industrial communication specification
and many good introductions to the subject exist [2] [19]. Another brief introduction
is provided here in the context of this thesis. OPC UA is essentially an application-
layer protocol which is designed to be used in automation and industrial use cases.
The primary purpose of an application-layer protocol is to define how processes
running on different end systems pass messages to each other. An application-layer
protocol defines the types of messages (requests and responses), the syntax of the
messages (fields), the meaning of information in these fields and rules determining
when and how a process sends messages and responds to messages [20]. Applications
operating in the control layer of the industrial computer systems are designed not
only to perform real-time data exchange between each other, but also to communicate
information to higher levels (SCADA, MES, ERP). OPC Unified Architecture is
an example of a standard that is designed to handle this type of communication.
The fundamental parts of OPC UA are data transport and information modeling [2,
pp.19]. It leaves it up to other organizations to define what data is described in their
information models.
OPC UA protocol is a classical example of the client-server communication model.
All communication between OPC UA applications is based on the exchange of
messages initiated by the client application. [4] Usually the OPC UA server resides
on an automation device. One of the areas where OPC UA has really excelled is PLC
interfaces. An OPC UA server encapsulates the source of process information like a
PLC and makes the information available via its interface. An OPC UA client, for
example SCADA system, connects to the OPC UA server and can access and consume
the offered data. Applications consuming and providing data can be both client and
server at the same time. In its current form, OPC UA applications mostly use binary
encoding and TCP transport protocol called UA TCP to access automation data.
Also, OPC UA communication is inherently tightly coupled and session-oriented. In
many applications this connection oriented approach is advantageous, for example
timely detection of communication failure and rapid recovery with no loss of data can
be achieved with this approach. However keeping track of sessions must in theory
introduce some boundaries on scalability of OPC UA applications.
OPC UA is an evolution of the previous OPC standard (currently referred to as
OPC Classic). [2] To understand OPC UA, it is best to become familiar with reasons
leading to the creation of the original OPC Classic protocol. The motivation for the
original OPC standard was to provide access to intelligent field devices provided
by multiple vendors and thus enable connecting industrial devices to control and
supervision applications. The development of OPC standard began in 1995 as an
answer to this connectivity problem. The standard became successful and in fact has
become a de facto standard in factory automation for integrating different production
automation related software in multivendor environments. This original version was
denoted OPC Data Access (DA). Later on OPC Classic developed to contain several
 9

different protocol specifications, for example Alarms & Event (AE) and Historical
Data Access (HDA) among others. These specifications broadened the scope of OPC
Classic.
OPC classic was based on proprietary COM/DCOM technology which restricted
its use practically to PC computers in LAN networks because of scalability and
security issues. Also the robustness of OPC was a problem and the segmentation
to different parts was considered negative. The motivation for the development of
OPC UA was to offer modern equivalent to OPC without loss of functionality or
performance. [2] The name Unified Architecture literally means unifying different
OPC specifications under one umbrella. OPC UA was designed to be better than OPC
in terms of information security, platform independence and enhanced information
modeling. OPC Foundation, which is the organization developing OPC, introduced
this new version in 2009.

3.2 Parts
OPC UA standard consists of 13 parts, which are listed in table 4. As of writing
this, the latest OPC UA specification version 1.03 was released by OPC Foundation
in 2015. The specification describes UA internal mechanisms, which get handled
through the communication stack and are mostly only of interest for those that port
a stack to a specific target or those that want to implement their own UA stack.
The OPC UA application developers usually code using some commercial SDK and
therefore mainly use API documentation. However, parts 3, 4, and 5 may also be of
interest for application developers. These parts describe the address space model,
services and the default information model of OPC UA.
All data in OPC UA communication is accessed via objects in address space.
Specification part 3 describes the address space and its objects. [21] Part 3 is
the OPC UA meta model on which OPC UA information models are based. The
primary objective of the OPC UA address space is to provide a standard way for
servers to represent objects to clients. Objects and their components are represented
in the address space as a set of nodes described by attributes and interconnected
by references. Specification part 5 defines the standardised nodes of the address
space of an empty OPC UA Server. These nodes are standardised types as well as
standardised instances used for diagnostics or as entry points to server-specific nodes.
The default information model contains about 5000 predefined nodes [22]. However,
it is not expected that all servers will provide all of these nodes [23].
OPC UA services are defined in specification part 4 and organized into nine
service sets as illustrated in table 5. Out of these, Discovery, Secure channel and
Session service sets are related to the underlying connection and the rest are used
to view, modify and use the data that is available from the information model. [24]
Each OPC UA Service described in Part 4 has a request and response message.
The defined Services are considered abstract because no particular mechanism for
implementation is defined in part 4. Part 6 specifies concrete mappings supported
for implementation.
 10

Table 4: OPC UA specification parts.

Part Description
Part 1 Overview and concepts.
Part 2 Describes the security model.
Part 3 Address Space Model. Describes the OPC UA meta model on which
OPC UA information models are based.
Part 4 Services. Defines collection of abstract Remote Procedure Calls
(RPC) that are implemented by OPC UA Servers and called by
OPC UA Clients.
Part 5 Information Model. Describes standardised Nodes of a Servers
AddressSpace.
Part 6 Mappings. Specifies relation between parts 2, 4 and 5 and the
physical network protocols.
Part 7 Profiles. Describes profiles which are used to segregate features of
OPC UA products.
Part 8 Data Access. Defines how clients may read, write or monitor
DataItems.
Part 9 Alarms and Conditions. Specifies the representation of Alarms and
Conditions.
Part 10 Programs. Defines the information model and associated behaviour
for programs.
Part 11 Historical Access. Defines the information model and associated
behaviour for Historical Access (HA).
Part 12 Discovery. Specifies how Clients and Servers interact with Discov-
eryServers.
Part 13 Aggregates. Defines the information model associated with Aggre-
gates

3.3 Mappings
OPC UA standard is defined in such a way that same abstract features can be
implemented with different protocols. This is done to increase the flexibility of
the standard. [2] Building actual OPC UA applications still requires that the
implementation details have been agreed upon. Specification part 6 describes mapping
between the security model described in Part 2, the abstract service definitions,
described in Part 4, the data structures defined in Part 5 and the physical network
protocols that can be used to implement the OPC UA specification [25]. Mapping
specifies how to implement an OPC UA feature with a specific technology. For
example, the OPC UA binary encoding is a mapping that specifies how to serialize
OPC UA data structures as sequences of bytes. Mappings can be organized into
three groups: data encodings, security protocols and transport protocol as illustrated
in table 6. Security Protocol ensures the integrity and privacy of UA Messages that
are exchanged between OPC UA applications. Data encoding is a way to serialize
OPC UA messages and data structures. Transport protocol represents a way to
 11

Table 5: OPC UA service sets.

Service set Use case
Discovery Discover servers and their security settings.
Secure channel Services related to the security model.
Session Maintain the session between a client and a server.
Node management Modify the address space.
View Browse through the address space.
Attribute Read and write attributes of nodes.
Method Call methods.
Monitored item Setup monitoring for attribute value changes or events.
Subscription Subscribe for attribute value changes or events.

exchange serialized OPC UA messages between OPC UA applications.

Table 6: OPC UA mappings.

Group Options
Data encodings UA Binary
UA XML
Security protocols UA-SC
Transport protocols UA-TCP
HTTPS

OPC UA currently specifies two encodings: OPC UA Binary and XML. OPC
UA Binary has been designed with performance and overhead on the wire in mind.
This is the most efficient way for data exchange in OPC UA. In addition to this, the
specification defines XML encoding. This structured format offers for example easier
debugging of messages. The cost of any structured encoding format such as XML is
the overhead introduced in the form of larger messages. Large majority of current
OPC UA applications use UA Binary encoding.
There is currently one security protocol defined for OPC UA in order to map
the abstract services defined in part 4 of the specification: UASecureConversation.
This security protocol is based on a certificate-based connection establishment.
UASecureConversation is a security protocol defined by the OPC UA working
group. It is a combination of techniques and mechanisms of the standards TLS4 and
WSSecureConversation [2, pp. 196]. Previously there was also a second optional
security protocol called WS-SecureConversation. However, in the version 1.03 of the
specification this is marked as deprecated because of not being widely adopted by
industry.
For transporting data, OPC UA currently has two technology mappings supported,
a TCP protocol based UA TCP and a HTTPS protocol. These protocols are used
for establishing a connection between an OPC UA client and server at network level.
UA TCP is by far the most widely used one of the transport protocols. UA TCP is a
protocol defined by OPC Foundation on top of TCP. UA TCP contains application
level configurable buffer sizes for sending and receiving data, the possibility to use
 12

Table 7: OPC UA transport facets.

Transport protocol Security protocol Serialization
UA-TCP UA-SC UA Binary
HTTPS SSL/TLS UA Binary
HTTPS SSL/TLS UA XML

same IP-address and port for different endpoints of OPC UA server and the possibility
to react on and recover from errors occurring at transport level [2, pp. 198]. HTTPS
refers to HTTP Messages exchanged over an SSL/TLS connection. In this protocol,
OPC UA messages are transferred in the body of HTTP messages. Previously
there was also a third optional transport protocol called SOAP-HTTP. However,
in the version 1.03 of the specification this is marked as deprecated because of not
being widely adopted by industry. This protocol was based on SOAP web service
technologies.
The combination of data encoding, security protocol and transport protocol is
referred to as transport facet by the OPC UA specification. The standard currently
defines three different transport facets (see table 7). In order for OPC UA commu-
nication to work between two applications, both must support at least one mutual
transport facet. In practice, binary UA-TCP is the most common transport facet in
use today.

3.4 Subscription model

A subscription is the context to exchange data changes and event notifications between
server and client. Clients send publish requests to servers and receive notifications
in the form of publish responses. The biggest benefit of the subscription model is
efficiency compared to polling values periodically. A client can subscribe for three
different types of information from an OPC UA server. These three types are variable
value changes, event notifiers and calculated aggregate values. A subscription is used
to group sources of information together. A monitored item is used to manage a source
of information. A piece of information is called a notification. A subscription can
contain all three different types of monitored items. Figure 1 illustrates association
of session, subscription and monitored item.
Figure 2 illustrates different subscription-related settings. The sampling interval
defines the rate the server checks variable values for changes or defines the time the
aggregate value gets calculated. The monitoring mode defines whether the monitored
item is active or inactive. The queue size defines how many notifications can be
queued for delivery. The default value for data changes is one and infinite for events
where the size of infinite depends on the resources available in the server. The
filter settings are different for data changes, events, and aggregate calculation. The
publish interval defines the interval when the server clears the queues and delivers
the notifications to the client. The publish enabled setting defines whether the
data gets delivered to the client. Services used to actually deliver the notifications
in a notification message to the client are the publish service for transferring the
 13

Monitored item
Service set Monitored Item
N
1
Subscription
Service set Subscription
N
1

Session
Figure 1: Subscriptions have a set of MonitoredItems assigned to them by the Client.
MonitoredItems generate Notifications that are to be reported to the Client by the
Subscription (modified from [2]).
Sampling interval Queue of data or events

Subscription
Monitored item for data
Variable
f
value
Publish enabled

Monitored item for events

Notification
Event notifier f message sent
to client

Monitored item for aggregates

Variable
f Publish interval
value

Filtering functionality Monitoring mode

Figure 2: Subscription and monitored item settings (modified from [2]).

notification messages and the republish service to get lost notification messages from
the server.
A subscription requires a session to transport the data to the client. The
subscription can be transferred to another session, for example to be used in a session
created by a redundant backup client if the client that created the subscription is no
longer available. Therefore the subscription lifetime is independent of the session
lifetime and a subscription has a timeout that gets reset every time data or keep-alive
messages get sent to the client. [2]

3.5 Security
OPC UA specifies security capabilities, which make use of standard public key
cryptography mechanisms. If HTTPS protocol is used, then TLS security is used
to encrypt the traffic already in the Transport Layer. This section will cover the
 14

security policies employed with the UA-TCP transport protocol.

Figure 3 illustrates the concepts of the transport layer, secure channel and session.
Each OPC UA application has an Application Instance Certificate, which is a
standard X.509v3 certificate with some extra fields for additional OPC UA validation.
This enables authentication of the applications which may communicate with each
other. The respective RSA public and private keys are used to perform a secure
hand-shake, when applications create the Secure Channel between them. Both
applications will perform the authentication of the other party in the hand-shake,
which in practice is an OPC UA OpenSecureChannel service message. During the
hand-shake the applications also exchange a symmetric encryption key, which is
then used to secure all forthcoming messages through the Secure Channel. When
a secure communication policy is selected, Secure Channel ensures the integrity or
confidentiality and integrity of the messages that are sent. Once the Secure Channel
is in place, the client application will create a Session in the server, over which all
other service messages are sent and validated. Session is used to authenticate and
authorize users. The session is always bound to a secure channel, which is also
renewed frequently. [2]

OPC UA Client OPC UA Server

Application layer Session Application layer
User authentication User authentication

Communication layer Secure channel Communication layer

Application authentication Application authentication

Transport layer

Figure 3: Transport Layer, Secure Channel and Session (modified from[26, pp. 12])

OPC UA enables a flexible selection of the used security policy and security mode
between each connection. Specification defines several alternative Security Policies:
None, Basic128Rsa15, Basic256 and Basic256Sha256 as of writing this. New policies
can be defined and old ones made obsolete as security requirements increase in future.
These define the suite of algorithms used to sign or encrypt messages. For example,
if the security policy Basic256Sha256 is used, then asymmetric signature algorithm is
Rsa_Sha256, symmetric signature algorithm is Hmac_Sha256, asymmetric encryp-
tion algorithm Rsa_Oaep and symmetric encryption algorithm Aes256_CBC. Three
Message Security Modes are available: None, Sign and SignAndEncrypt. These define
the level of security applied to each message. If "None" is selected, then the messages
will not be secured. Of course also no overhead is imposed because of signing or
encrypting. When "Sign" is chosen then the message is signed. Signing messages
 15

allows detecting whether a received message has been manipulated by an untrusted

third party. If "SignAndEncrypt" is used, then the message is additionally encrypted.
Encrypting messages prevents or at least makes it very difficult for untrusted third
parties to read the content of messages. The server administrator may configure
which policies and modes are available. Of course, both the client and server need
to support the same Security Policy and Mode in order to communicate with each
other.
User authentication is performed on the session level. OPC UA defines alternative
authentication methods: Anonymous, User name and Password combination, X.509
certificates and also ways to use external user authentication systems, such as
Kerberos, via External Tokens. [2]
 16

4 Previous studies of OPC UA performance

Previous studies could be classified in multiple different ways. The first factor is
of course what OPC UA service set the test is measuring. Most previous studies
concentrate on attribute and subscription service set (see table 5) which is reasonable
as these are the services which are used all the time during an OPC UA session.
The second factor that might be used to classify previous studies is which transport
mappings or security modes (none,sign,sign and encrypt) they use. Some studies
start by defining the needed performance levels and then performing measurements
[5] [27] as others are just interested in absolute performance. One separating factor
is whether the test measures speed indices or utilization indices (see table 1) and
which indices are considered. For example speed indices might be either throughput
or response time measurements. In this chapter the previous tests are divided into
speed indices (effectiveness) and utilization indices (efficiency) as that provides a
logical separation of test cases.

4.1 Speed indices

Industrial applications generally feature two different kinds of data exchange. The
first is asynchronous, and occurs when a client (e.g. SCADA) requires the read or
write access to one or more variables at unforeseeable time instants. The second kind
of data exchange is cyclic or periodic and occurs when a client accesses values of
variables according to a periodic signals; an example is a client that needs to access
a temperature value produced by a sampling algorithm. In the case of asynchronous
data exchange, round-trip delay seems a suitable parameter for the performance
measurements; considering the read service issued by a client, round-trip delay may
be defined as the total response time between the instant at which a request to read
one or a set of variables is issued by a client and the instant at which the relevant
values are delivered to the client. Considering the periodic data exchange and taking
into account a particular variable whose values are periodically produced on the
server-side, the delay between the instant at which each value has been produced
and the instant at which the client receives that value, seems to be of interest during
performance evaluation. In the following, this time interval will be simply called delay.
The average delay for each variable gives a measurement of the efficiency of the data
exchange. [28] In the subscription case, also the overall throughput is interesting.
This is because it is much more important that all values are communicated reliably
than it is that values are communicated quickly but some values would be dropped.
Multiple studies of OPC UA communication performance have concentrated on
elapsed time tests and in particular the read service. Measured round-trip times
without network effect for small message sizes vary between 0.2 milliseconds [2]
and 50 milliseconds [29]. When message size grows and real network is used, then
round-trip times grow. Tests have been carried out with both scalar values and data
blocks. Roundtrip times for scalar values depend on the number of variables per call
and the network speed, while data type, such as Boolean, integer or double, is not a
significant factor. Roundtrip times for a single variable call depends on the block
 17

size and the network speed. For example reading 2.5 megabyte data block from an
OPC UA server takes about 200 milliseconds [7]. The network that is used has an
effect on the absolute performance. When read calls on localhost are in the range
of average 1.4 milliseconds, same read calls can be almost as fast (1.5 milliseconds)
when executed in fast LAN network. However same read calls over rural 3G network
takes on average 96 milliseconds. Ping time to some common servers in this same
network is in the range of 30 - 40 milliseconds. [5]
The read service call uses request-response pattern, and delays are present in both
steps. In contrast, notification messages are transferred without a specific request,
making them more responsive. For this reason, the latency of subscriptions can
be assumed to be smaller than periodic reading of variable values discussed above.
Sometimes it is more important to achieve the biggest throughput possible rather
than concentrate on fast responses [30]. However, previous studies do not seem to
present comprehensive throughput results. Fojcik and Folkert [12] present result of
250 kB/s throughput. However, that result is not the maximum possible throughput
as the test case is defined to test what is the average network traffic generated with
certain use case.
Signing and encrypting data has been of interest in many studies. Signing and
encryption adds overhead to the communication [31]. The conventional wisdom is
that signing and encrypting data is a time demanding task and potentially unusable
in resource limited devices. The previous results show somewhat ambiguous results:
some studies report that encrypting messages seems to increase round-trip time
by a factor of ten, whereas signing seems to have little or no effect [32]. On the
other hand, some studies show that with small message sizes encryption doubles
the amount of round-trip delay and signing and encryption operations take almost
identical times [33]. These studies also differ so that the first one reads variable
with static size 102400 bytes whereas the second one shows results with multiple
variable sizes and concludes that the impact of signing and encrypting gets smaller
as message sizes increase. These two results are illustrated in figures 4 and 5. In
figure 5 the results are presented normalizing the values of the round-trip time to
those achieved considering scenario no security and UA TCP. This study presents
different results for security modes "no security" and "secure channel with security
mode none". It is not completely clear what the difference between these is. The
conclusion is however that there is a clear difference between these results.
In addition to the attribute and subscription service sets, the session service
set could also be evaluated as an important part during an OPC UA session. In
some use cases where the goal is to gather data from multiple sources, the connect
and disconnect times may become important. Results obtained with OMNeT++
simulator show that creating an OPC UA session takes from 0.054 seconds to 14
seconds depending on what kind of CA verification is used. [33]
In conclusion, OPC UA communication effectiveness depends on the way that
the data in the OPC UA server address space is updated, the network speed, number
of nodes per call and data block size. Here, updating the server address space is
considered to include possible subscription settings such as sampling interval and
publish interval in case of subscription. To some extent the selected security mode
 18

60

50

40
Milliseconds

none
30 128sign
256sign
20 128enc
256enc
10

0
0 20 40 60 80 100
Measurement

Figure 4: Signing and encrypting overhead according to Post, Seppl, Koivisto

(modified from [32]).

3.5

3
Relative round-trip time

2.5
UA TCP none
2 UA TCP sign
UA TCP encrypt
1.5

1
4 1024 2048 4096 8192 16384 32768
Size of variable in bytes

Figure 5: Signing and encrypting overhead according to Cavalieri et al. (modified

from [33]).
 19

(whether or not to use signing and encrypting) has an effect. The programming
language and operating system may have an impact on performance, but the impact
is probably much smaller than with the before mentioned features. Different data
types do not influence read and write performance.

4.2 Utilization indices

Performance does not only mean the speed of communication, but also the load and
resource requirements imposed on the target system by the application. In theory,
server load should depend on the number of clients and amount of data to process.
Processor load in an application that produces constant workload has been
investigated in a case study from year 2009 [7]. The hardware used in this specific
test is Intel Atom Z530 and two different profiles were used. The embedded UA
server profile consisted of 500 variables across 5 subscriptions. The standard UA
server profile in turn included 37500 variables across 75 subscriptions. When the
embedded profile was sampled at 100 milliseconds sampling rate, the processor load
was 2%. When the standard server profile and sampling rate of 1000 milliseconds
were used, the processor load rose to 20%. When sampling was done with the same
profile but with sampling rate of 500 ms, the processor load was 40%. According
to this, sampling had a huge linear effect on the processor load, as doubling the
amount of sampling also doubled the processor load. The amount of subscriptions
and handled variables of course also has an effect on the processor load.
It is possible that a large amount of monitored items can inhibit OPC UA server
from working correctly [34]. A study by Fojcik and Folker from year 2012 suggests
that even less than 1000 monitored items may prevent acceptable performance level
[12]. Figure 6 illustrates this result. It is interesting that the reported level is so
low as the test uses PC-class computers with quad-code processors and 4 gigabytes
RAM memory. This finding is contrary to previous where 37500 monitored items
caused CPU load of 20%. Of course, there is difference in the OPC UA software
stack used and a different hardware than the previously mentioned. The same study
also illustrates how server application architecture can make a huge difference. Two
versions of server applications were tested. In the first case the server was updating
tags values directly after acquiring new data from data source and in the second
case there were two separate threads: one for gathering data from the data source in
dedicated queue and second only for updating associated monitored items. In the
first case, where tags were updated without buffering, the CPU load for 20 clients
was on the level of 60% and the CPU was not able to handle all of the processes. In
the second case the CPU load for 50 clients was on the level of 15%.
In conclusion, resource requirements imposed on the target system by an OPC UA
application are more application specific than communication effectiveness discussed
in previous section. There are less publications regarding hardware utilization
than there are publications measuring communication speed. A typical thing is
that round-trip values are presented in the case of read service tests and resource
usage is reported in the case of subscription testing. This is logical because in a
normal application scenario the important thing about subscriptions is that all value
 20

Figure 6: CPU load as the function of monitored items according to Fojcik and
Folkert (modified from [12]).

changes are reported reliably and the speed of actual data communication is not that
important. The resource usage of other services besides subscriptions was not found
in previous studies.

4.3 Protocol comparison

OPC UA does not exist in a vacuum. Knowing how OPC UA implementations
compare against competitors is valuable information. The older OPC standard is still
widely deployed in legacy applications and in some sense could be considered as a
competitor to OPC UA. Thus, the first comparison is naturally between OPC classic
and OPC UA [2]. As said before, OPC UA supports multiple transport protocols.
Differences between these transport protocols inside OPC UA specification are of
course interesting. Lastly, differences between completely different protocols like
OPC UA and web service technologies would be interesting.
Historically, the COM and DCOM technologies used in OPC Classic were high
performance binary protocols. One of the requirements for OPC UA was to maintain
or even enhance the performance of Classic OPC [2]. Tests comparing these two
specifications confirm the hypothesis that OPC DA and OPC UA data transferring
capabilities are roughly the same [2]. A study made at Cern reports update times as
function of subscribed elements [35]. The time that it takes for the whole process is
rather similar with a low number of items, but as the number of items increases in
the servers address space, the OPC UA Client/Server remains having a very low
update time (similar as having a low number of items), while OPC DA Client/Server
has an increasingly higher update time. For 5000 elements the update times are 1000
milliseconds for UA application and 2000 milliseconds for OPC DA application. For
30000 elements, UA application still updates at a rate of 1000 milliseconds whereas
OPC DA application is taking almost ten times as long (10000 milliseconds). Thus,
 21

it appears that systems based on the OPC UA scale better in terms of address space
size than analogous systems based on OPC-DA. Using UA TCP and binary encoding,
OPC UA is able to keep the performance of Classic OPC, while at the same time
providing new features, such as security of the communication [36, pp. 179].
Historically XML encoding has been seen as resource intensive [2, pp. 308].
The OPC UA book presents results of measuring round-trip times with different
transport protocols and data encodings. The numbers indicate that there is only
a small overhead for using SOAP/HTTP protocol with binary encoding instead of
the UA TCP protocol. There is a much bigger overhead when using SOAP/HTTP
protocol with XML encoding instead of the UA TCP protocol. SOAP/HTTP protocol
with XML encoding is 1.8 times slower for small messages and 18 times slower for
large messages [2]. Some publications even deem the XML format prohibiting high
performance real-time systems [37]. However, XML or other structured data encodings
have also positive sides such as human-understandable format and flexibility. Usually
using structured data such as XML messages results in much bigger message sizes,
16 to 25 times larger than a conventional binary message [13]. This illustrates the
overhead that is associated with using a structured versus a custom data format.
However, Efficient XML Interchange (EXI) data encoding of XML messages may
increase the efficiency of structured data formats. In an example case, message size
is 10 kB for UA Binary, 270 kB for XML and 15 kB for EXI encoded XML message
[38]. According to this example, the message size for UA Binary coding (3.75% of
the original XML message size) is nearly the same as for the default EXI coding
(5.15% of the original XML message size). However, this study does not report
results of round-trip times or resource usage when using EXI encoded XML-messages
and thus it is not clear that the smaller message size leads to better communication
performance automatically.
Some protocol comparisons are more of a qualitative kind. One interesting study
from year 2011 compares multiple IP-based protocols in terms of suitability for
automation gateway role. SNMP, LDAP, SQL, Web servers, OPC XML-DA, OPC
UA and a proprietary protocol are compared. [1] One interesting conclusion is that
web servers presenting HTML are well suited for presenting fieldbus and gateway
data to a human operator but introduces overheads with periodic data exchange.
Using SQL as gateway is also an interesting premise. The study criticizes support
for generic clients and support for different data types when using SQL.
In conclusion, it can be stated that binary data encoding should be more efficient
than XML or other structured formatting. OPC UA binary and classic OPC should
exhibit same kind of performance. OPC UA standard has been developed to use
cases in automation gateway role and contains features needed in that application
area.

4.4 Tools
Unified Automations UaExpert software is shipped with a Performance View feature.
This performance view can be used to measure the performance of certain UA service
calls to a UA server. It can be configured to call a service for a defined number
 22

of times or to call a service as often as possible in a specified time span. Figure 7

illustrates the user interface of this performance view feature. Current version of
UaExpert as of writing this is 1.4.1. The supported services of the tool are read, write,
read registered, write registered, create monitored items and delete monitored items.
The tool can be used to measure round trip times of these service calls or the amount
of service calls made in a specified period of time. In round-trip mode, minimum,
average and maximum round-trip times are reported and in addition to these, all
individual round-trip times are available. In duration mode, the total number of
calls and the average time per call are reported. The amount of service calls per time
period is done with synchronous calls from client application. This means that for
example new read call is made as soon as previous read call returns. This approach
is different from the approach taken for example in the httperf application presented
before where the aim is to saturate server with asynchronous service calls.
In duration mode, tests produced by this tool are perfect examples of the through-
put test as defined in the chapter 2.1. In the round-trip mode the definition of these
tests is a bit more unclear. In principle, the test case sounds like a response time
test, because what is measured is a response time. However, there is no concept
of think time in these tests, because next request is sent when previous is gotten
from the server. In other words; throughput demanded by the client is not fixed.
Still, the amount of work is fixed, because the amount of service requests is decided
before starting the test. This leads to classifying this test case as batch measurement.
Actually, the amount of service calls can be thought of as a "batch", which explains
the naming quite well. This tool is not designed to be used with third type of tests,
response time tests. In response time tests, some think time between requests would
be introduced.
 23

Figure 7: UaExpert Performance View

 24

5 Test requirements
It would not be feasible to discuss all the possible aspects of OPC UA performance
more deeply in this thesis. In this section OPC UA functionality is assessed based on
previous studies. Usually users are in the first place interested in data transferring
capabilities of OPC UA applications [39]. Performance and possible limitations of
OPC UA subscription model are interesting, especially because the subscription
model has previously been the limiting factor of performance in some cases [34].
When analyzing the data transferring capabilities of an OPC UA application it
would be obvious to expect that the network imposes restrictions on data throughput.
However, previous research has concluded that indeed this is not the case with fast
LAN network but actually the internal architecture of the server and utilization of
hardware resources can be a bottleneck [12]. OPC UA supports multiple technology
mappings. Using HTTPS as the transport or XML as the encoding is not that
popular. In the following tests only UA TCP and UA binary encoding are considered.
Based on previous studies, three main research topics can be formulated:

Effect of signing and encryption

Overloading the server with asynchronous reads

Subscription performance testing

To provide some logical separation, the selected test cases are logically divided
into elapsed time (batch) measurements, throughput measurements and response
time measurements. Based on previous studies, executing response time tests is the
most interesting part, since that kind of studies have not been done previously with
OPC UA, whereas with other information systems response time testing is usual. In
response time tests, workload and throughput are constant.
For the elapsed time tests, the server application needed to contain variables
of selected size. In these tests, client and server application resided on the same
host. Round-trip times without network effect were measured as well as relative
effect of signing and encrypting. Synchronic (blocking) read calls were used. Only
the round-trip time was measured and in this case the client functionality was
really straightforward. For the response time tests, a custom client application was
implemented and the idea was that the response time client could be used with any
OPC UA server. The client functionality as well as collected metrics were more
complicated. When running the response time tests, client and server applications
were always on different host machines. In response time tests the main idea was
to measure the response times until the server application becomes saturated with
more requests than it can handle.

5.1 Elapsed time (batch) measurements

In these tests, the workload is pre-determined and the interesting thing to measure
is how fast the given workload is executed.
 25

5.1.1 Impact of signing and encrypting

As previously mentioned, signing and encrypting data has been studied many times
previously but with somewhat ambiguous results. This is one aspect of interest where
more results and documentation is needed. Besides the test results documenting
round-trip times and resource usage, it is important to provide detailed information
of the software and hardware platform used in this test.

5.2 Response time measurements

Based on previous publications, read and subscription services are considered the
most interesting ones for response time measurement testing. It is important to
ensure that the client side application is not the limiting factor during these tests. If
however the client side is the limiting factor, then documenting the resulting resource
usage in the given hardware and software platform is essential. In response time
tests, a suitable think time should be introduced between requests and this leads the
throughput to be fixed unless the server application is overloaded with more requests
than it can handle.

5.2.1 Read mechanism

The read services request and response format are pretty straightforward. One thing
to note is that every read request in OPC UA can contain a list of nodes and their
attributes to read [24, pp. 48]. This means that single read service call may request
attribute values from multiple nodes and one read response may contain multiple
attribute values for multiple nodes. This list size is referred to as "batch size" in the
following text. In order to determine how OPC UA servers react to growing number
of read requests, the following measurements should be made.

1) Read one value with growing frequency

2) Grow batch size, while keeping the amount of sessions and frequency constant.
The session amount should be one and frequency should be decided based on
previous test case.

3) Grow the amount of sessions while keeping other parameters constant. Batch
size and frequency of reads should be decided based on results of previous test
case.

Response times (arithmetic mean, maximum, 90th percentile time) as well as

amount of responses per second are recorded in these test cases in order to determine
the achieved response times and throughput. These tests can be classified as response
time tests measuring speed indices. Resource usage on server and client shall be
monitored manually. All read requests in this test should be made asynchronously
to allow client to keep sending messages even if the server has not responded to all
previous requests. This approach makes it possible to overload the server.
 26

5.2.2 Subscription mechanism

The purpose of this test case is to grow the number of sessions, subscriptions and
monitored items and measure the delays as well as achieved subscription throughput.
This test should be run with enough client machines to ensure that the client side
application is not the limiting factor. The intended usage of subscriptions is that
one subscription contains multiple monitored items. Usually, one session has 1-10
subscriptions, each of which may contain even thousands of monitored items. One
important detail is that the monitored items should be added in one complete set
instead of repetitive requests [24, pp. 62]. In order to determine the available
subscription mechanism performance, the following measurements should be made.

1) Test response times and throughput with one session, one subscription and
variable amount of monitored items.

2) Test response times and throughput with one session, variable amount of
subscriptions and constant amount of monitored items. The constant amount
is decided based on the previous test.

3) Test response times and throughput with variable amount of sessions and
constant amount of subscriptions and monitored items. The constant amount
is again decided based on the previous test.
 27

6 Research material and methods

Because of the client-server architecture, where the server is passive and only responds
to requests made by a client, it is possible to make a generic test client application
which is usable with any server application.

6.1 Design of response time test client

In order to stress the server side application, a custom client application was developed.
This makes it easier to reproduce tests and potentially enables people to run tests
with their own applications if the test client is published. Sample client application
provided with the Prosys OPC UA Java SDK (from now on referred to also as the
SDK) was used as a base for this test application. The most important aspects
were the read and subscription testing and test functionality for these services was
developed with highest priority. As the output of this process, a command line
application named UaPerfClient was made. An example of command line options to
start the application is shown here:
n " ns =2; s=MyLevel " c l i e n t s 1 b a t c h s i z e 100 r a t e 100
r e q u e s t s 9000 d u r a t i o n 300 p e r c e n t i l e 0 . 9 opc . tc p
: / / 1 0 . 5 0 . 1 0 0 . 1 5 6 : 5 2 5 2 0 /OPCUA/ S a m p l e C o n s o l e S e r v e r
This example connects to the specified OPC UA server residing on a ma-
chine with ip address 10.50.100.156 and starts reading the value attribute of node
"ns=2;s=MyLevel" with a rate of 100 read cycles per second. In this example, one
read cycle consists of one session (-clients 1) sending one read request to the server.
The batchsize parameter defines the size of the nodesToRead list in a read request.
In this example 100 values are read in every request. Parameter requests defines
the total amount of read requests to make. With these settings, the test will run
for 90 seconds when everything goes normally, because 100 requests per second *
90 seconds = 9000 requests. If the server cannot keep up with the requests that
the client supplies, the test run will take a longer time. After duration amount of
seconds the client execution is stopped in any case. After the test is finished, the
UaPerfClient will print statistical information about the run. In addition to the
parameters introduced here, table 8 lists all the command line parameters of the
UaPerfClient application.

6.1.1 Read testing

In OPC UA, the read requests are always performed asynchronously. Ordinarily,
the OPC UA SDKs offer both asynchronous and synchronous read functions. The
synchronous reads (or blocking reads) explicitly wait until a response arrives from
the server. This provides convenience for the client application developer. The
Prosys OPC UA Java SDK also supports reading with synchronous or asynchronous
methods. As stated before, the client side application should be able to overload the
server side application with more requests than it can handle. For this purpose, the
obvious choice was to use the asynchronous read methods.
 28

Table 8: Command line parameters of UaPerfClient application.

Parameter Declaration
clients number of client sessions to establish.
batchsize defines the number of ReadValueIds in nodesToRead list in
batch mode and number of read requests per client connection
in single mode.
duration the maximum time to wait for asynchronous responses or the
duration of subscription test.
subscriptions number of subscriptions to establish per client connection.
items number of monitored items established per client connection.
single only request one value per read request. Default operation is
not to use this mode.
rate define the rate (requests/second) at which read requests are
issued during read test.
requests amount of read requests. This defines the duration of read
test.
percentile Which percentile value of round-trip or delay times is calculated.
Default 0.9

Periodic reads from the client application was developed by defining a Runnable
object which handled the asynchronous read calls and scheduling this Runnable with
ScheduledExecutorService. ScheduledExecutorService is a fixed-size thread pool
that supports delayed and periodic task execution, similar to Timer [14, pp. 120].
Scheduled thread pools allow multiple threads for executing periodic tasks. Sched-
uledThreadPoolExecutor also deals properly with ill-behaved tasks and continues
the execution even if some individual tasks ended up in error. At the end of the
executions, statistical information including information about round-trip times per
single request and total throughput statistics will be printed. Parameter percentile
can be used to print certain percentile value of all arrived requests. The default
option is to print 90th percentile response time of all request-response round-trip
times. An example of statistics printed out by the program is shown below.
Number o f read r e q u e s t s : 9000
Number o f read r e s p o n s e s : 9000
T ot al d u r a t i o n : 8 9 9 9 8 . 9 7 9 1 4 6 9 9 9 9 9 m i l l i s e c o n d s
Average r e s p o n s e time : 1 . 7 4 5 7 1 5
Max r e s p o n s e time : 1 0 . 3 2 3 4 3 5
0 . 9 p e r c e n t i l e r e s p o n s e time : 1 . 8 4 5 7 5 4
Number o f r e s u l t s : 900000
R e s u l t s p er second : 1 0 0 0 0 . 1 1 3 4 2 9 3 9 7 7 2
In the default mode, the Prosys OPC UA Java SDK contains functionality to monitor
server status. This server status monitoring is done by reading the server status as
any other data. With large amounts of read requests, the client may falsely think that
the server is not responding because of long wait times before the server responds to
 29

status reads. Clients can set an interval for this status check as well as a timeout
value. By adjusting the timeout, the user can configure for how long the client
continues to operate without knowing what the status of the server is. In the client
application, this status check timeout is set in the following manner.
c l i e n t . setStatusCheckTimeout (< s u i t a b l e value >) ;
However, during testing there is really no point to monitor the status at all.
Either the client stays connected to the server or, if not, the test run can be repeated.
Also, the server status reading actually interferes with the test run because it also
involves issuing read requests. Because of this, it is better to disable the status check
altogether by setting the status check interval to zero.
c l i e n t . setStatusCheckInterval (0) ;

6.1.2 Subscription testing

The main architecture of UaPerfClient with subscriptions is that sessions, subscrip-
tions and monitored items are initialized. After that,a client side wait is introduced
before measurement phase begins. During measurement, the number of publish re-
quests, publish responses, monitored items, test duration, delays and subscriptionIDs
in publish responses are tracked. After the time specified by the parameter duration
has passed, logging of publish requests, responses and monitored items is stopped
and subscriptions are unsubscribed. The intention is to ensure that only the time
when all subscriptions are fully initialized is taken into account in the measurement.
An example of statistics printed out by the program is shown below.
Number Of P u b l i s h Requests : 90
Number Of P u b l i s h Responses : 90
Number o f Monitored Items : 9000
Duration : 1 8 . 0 1 8 4 7 9 1 3 7 s e c o n d s
Monitored i t e m s p e r second : 4 9 9 . 4 8 7 2 1 7 0 7 1 4 4 3 8
Average d e l a y time (ms) : 7
Max d e l a y time (ms) : 15
0 . 9 p e r c e n t i l e d e l a y time (ms) : 13
S u b s c r i p t i o n min f r e q : 18
S u b s c r i p t i o n max f r e q : 18
Command line options to obtain this result were
n " ns =2; s=MyLevel " c l i e n t s 1 s u b s c r i p t i o n s 5 i t e m s 100
d u r a t i o n 18 opc . t cp : / / 1 0 . 5 0 . 1 0 0 . 1 5 6 : 5 2 5 2 0 /OPCUA/
SampleConsoleServer
In the case of read requests, round-trip time was a parameter of interest. In the
case of subscriptions, there is no similar concept of round-trip time. Instead, the
time when new value is obtained by the server and the time when client application
gets notified can be monitored. This is called just delay in this text. As in read
requests, in delay measurement, the average, maximum and percentile values were
 30

obtained. The delay could be measured through number of ways: 1) based on the
timestamp field of response header in publish responses, 2) from the publish time
field of notification message (there is one notification message per publish response)
or 3) from individual data values server timestamp. In practice, the difference in all
these approaches is probably not that big but the most correct way would be to use
the timestamp of each individual data value. The specification states that the server
timestamp is used to reflect the time that the server received a variable value or knew
it to be accurate [24, pp. 121]. During testing, it was however noticed that calculating
the time difference from every received value update consumes surprisingly much
resources and client side easily becomes a bottleneck because of this. The next best
thing in delay measurement is the publish time field of notification message, which
OPC UA specification defines as the time that the message was sent to the client
[24, pp. 136]. This was deemed very suitable for these measurements as there is
considerably less notification messages than individual data values.
The tests were made with server and client applications residing on different hosts.
In read service testing the round-trip time is measured with the client system clock.
In subscription delay measurement however, server machine sets the publish time,
and client application calculates delay from this time. Differences in these clocks
are reflected in the measured delay. The first delay time measurements pointed out
that internal clocks actually drift a considerable amount if no clock synchronization
protocol is used. Adjusting clocks before every test would be daunting and would also
introduce huge differences in measured delay values. Because of this, it is important
to use clock synchronization protocol. Despite the use of clock synchronization
protocol, some difference of clocks was visible between different client hosts. Two
client hosts were used, and immediately after synchronizing clocks, the other one
reported average delay time of 7 milliseconds. In similar test run on the other host,
the delay was in the range of 60 milliseconds. In absolute values, this may seem like
a small difference, but it is still an increase of almost ten times. The delay presented
in the subscription tests should be anyway thought of in relative terms. This means
that the most interesting thing in practice is to see how the delay time grows when
more data notifications are handled.
With subscriptions, an important aspect of testing was to try to evaluate whether
or not all notification messages actually arrive at the client application. This is very
important in the case of subscription mechanism, where one missing data notification
could mean that the client never gets the newest value of some variable. Missing
some value update altogether might potentially be very harmful. Delivering all
notifications reliably is more important than delay of notifications. In OPC UA, one
publish response always contains data notifications for only one subscription. In
UaPerfClient, incoming data notifications are counted and in addition to this, the
publish responses per subscription are monitored. This is by no means a perfect
way of telling if all notification messages arrived as they should have but it does
not consume much resources and it still provides some information about how fairly
subscriptions have been treated. Measuring all individual notification messages and
analysing that they arrived in order would be better, but when designing the system
it was deemed too resource intensive because it is possible that high amount of
 31

sessions, subscriptions and monitored items may be measured. As in the case of

delay measurement, the client side easily becomes the bottleneck and thus resource
usage in the client side was kept to a minimum.
The intention is to ensure that only the time when all subscriptions are fully
initialized is taken into account in the measurement. To make sure that all sessions,
subscriptions and monitored items have been initialized correctly, a wait was intro-
duced before measurement phase begins. This wait was decided to be 60 seconds as
that seemed long enough without making test cases unnecessarily long.
In subscription testing, it is of course necessary that the tested server application
has some node with deterministic value changes. In the Prosys OPC UA Java SDK
SampleConsoleServer the sample node "ns=2;s=MyLevel" provides this service as
its value is changing once every second. Correspondingly, in the UA Demo Server
node "ns=4;s=Counter1" changes at intervals. This value changing could perhaps
also be done by another client program continuously writing values to a selected
server node. This approach would offer more flexibility as basically any server could
be tested with any data change interval desired. However, it would also mean that
not only subscription service performance would be tested because the server would
be handling both the write calls and subscriptions at the same time. This is why
this approach was not studied further during this thesis.
In Prosys OPC UA Java SDK, the UaClient class represents a client connection
interface to an OPC UA server. The most straightforward way to add custom func-
tionality to the UaClient is to define a custom UaClientListener. SDK sample imple-
mentation MyUaClientListener was extended with a class called UaPerfClientListener.
The class UaPerfClientListener keeps track of publish requests, publish responses and
data change notifications. One instance of this class is shared by multiple UaClients.
The variables were defined as AtomicIntegers and updated via incrementAndGet()
methods to avoid multithreading issues.
Compared to reading and writing variables, the subscription mechanism is more
complicated. There are multiple parameters that affect the communication. The
most important ones of these communication parameters are listed here:
Client side subscription settings:

- NotificationBufferSize defines the maximum number of Notification-

Data packets to keep in the client internal buffer. The buffer is used to
handle the incoming notification packets in the client. Setting the value
too low causes the client to ask for republish frequently.
- MaxNotificationsPerPublish is the maximum number of notifications
that the client wishes to receive in a single publish response.
- MaxMonitoredItemsPerCall defines the maximum number of items
in one service call.

Client settable server side MonitoredItem parameters

- SamplingInterval is the interval that defines the fastest rate at which

the MonitoredItem(s) should be accessed and evaluated. This interval is
 32

defined in milliseconds.
- QueueSize is the requested size of the MonitoredItem queue.

Server side subscription settings:

RetransmissionQueueSize defines the amount of messages to keep

stored for retransmission. This effects how easily subscription starts to
report "Message not available" errors. Clients cannot set this variable.
The default value is 10. Specification part 4 states that the session shall
maintain a retransmission queue size of at least two times the number of
publish requests per session the server supports [24, pp. 69].

A couple of the most influential parameters to subscription performance were

detected when instantiating a large amount of subscriptions. The first case was that
about 10000 monitored items can be added in one request with normal settings.
This happens because the message just grows to be such a big one. When trying to
subscribe to more monitored items the following error is shown.
com.prosysopc.ua.ServiceException: Bad_EncodingLimitsExceeded
com.prosysopc.ua.ServiceException: ServiceFault: Bad_TooManyOperations
Monitored items must be thus added in smaller batches. The Prosys OPC UA
Java SDK contains functionality of OperationLimits to work with this situation.
When creating subscription, the subscription can be limited to add a maximum of
10000 monitored items per time.
s u b s c r i p t i o n . setMaxMonitoredItemsPerCall ( 1 0 0 0 0 ) ;
Another case was that with about 100000 monitored items, the client started
reporting missing notification messages from republish.
com.prosysopc.ua.client.Subscription - Requesting missing NotificationMessage
(SequenceNumber=261) from the server using Republish: Target
SequenceNumber=265
com.prosysopc.ua.client.Subscription - Message not available
This message was shown because the client side notification buffer was too small
for this particular test case. The client had received the mentioned message and
acknowledged it. After this, another thread is responsible for actually handling the
publish response. If notification buffer overflows, then messages are lost in the client
side application before they can be handled. Setting client side notification buffer
size to larger value should help in this situation.
subscription . s e t N o t i f i c a t i o n B u f f e r S i z e (5000) ;
 33

7 Results
7.1 Elapsed time results
Modern PC hardware usually contains hardware-accelerated encryption. However,
not all devices support such features. The purpose of this test case was to find out
what kind of performance could be expected from different hardware platforms. We
compared the following computers:

Dell laptop, OS Windows 8.1 64-bit, Intel Core i7 @2.70GHz, 8 GB RAM

memory and SSD drive.

Raspberry Pi, OS Raspbian Linux, 700 MHz single-core ARM, 512 MB RAM
memory and SD card as a storage.

The Dell laptop contains hardware-accelerated encryption whereas the Raspberry

Pi does not have this feature. The average results of profiling the time used to
encrypt a single OPC UA chunk at a given platform are shown in table 9. The
most interesting thing in practice is not the absolute time but the relative difference
between the devices. We see that on the Raspberry Pi platform the encryption takes
approximately 20 times more time than on normal modern PC hardware.

Table 9: Average time to encrypt a single request value.

Platform Milliseconds
Laptop 0.19
Raspberry Pi 3.87

There is considerable difference in these measurements which is also just what

was expected. The User of OPC UA SDK is however not interested in the time taken
to encrypt one value but more in the round-trip time of service calls. It is again
expected that adding signing and encryption will increase the round-trip time. The
interesting thing is how much overhead the security modes will induce. Figures 8
and 9 show the time required to make a complete read service call, including the
request and response messages. The read consists of a single variable of byte array
with 10 elements.
The measurements were made with both the server and the client on the same
machine using the sample applications included in Prosys OPC UA Java SDK. The
communication protocol was UA Binary and the security policy was Basic128RSA15.
The times were recorded by profiling synchronic read service calls and averaging it
for 1000 calls. As a conclusion we can see that encryption adds overhead to the
communication as expected. Round trips with encryption take 1.5 2 times as
long as without encryption. However, in most application areas this could still be
deemed negligible because the absolute time is in the range of 0.5 ms to 10 ms,
only. Figures 10 and 11 illustrate the same measurements but now with a byte
array variable containing 10 000 elements. It can be seen that when the amount of
transferred data grows, also the overhead of encryption seems to grow. In the case of
 34

Laptop round trip times

1.2

0.8
Milliseconds

0.6

0.4

0.2

0
none sign sign and encrypt
Security mode

Figure 8: Average round-trip time with laptop and byte array of length 10.

Raspberry Pi round trip times

35

30

25
Milliseconds

20

15

10

0
none sign sign and encrypt
Security mode

Figure 9: Average round-trip time with Raspberry Pi and byte array of length 10.
 35

Table 10: Relative average round-trip values compared to security mode none with
different security modes and the security policy Basic128RSA15.
Platform Variable size None Sign Sign and encrypt
Laptop 10 1 1.2 2.2
Laptop 10000 1 1.5 2.7
Raspberry Pi 10 1 1.1 1.6
Raspberry Pi 10000 1 1.3 3.8

the laptop computer, round trip times with signing take 1.5 times as long as the time
without security. Adding encryption adds more overhead and round trip times with
encryption take 2.7 times as long as without security. In the case of the Raspberry
Pi, the corresponding values are 1.3 and 3.8. These values are illustrated in table
10. Of course in absolute measurements the laptop is always faster in round trip
times. Looking at the relative numbers however shows that with small message sizes,
adding signing or encryption actually adds relatively more overhead to the laptop
version. When the message size is increased this relative difference is no longer true.
Signing messages adds much less relative overhead than encrypting messages.
 36

Laptop round trip times, larger message size

1.6
1.4
1.2
Milliseconds

1
0.8
0.6
0.4
0.2
0
none sign sign and encrypt
Security Mode

Figure 10: Average round-trip time with laptop and byte array of length 10 000.

Raspberry Pi round trip times, larger message size

90
80
70
60
Milliseconds

50
40
30
20
10
0
none sign sign and encrypt
Security Mode

Figure 11: Average round-trip time with Raspberry Pi and byte array of length 10
000.
 37

7.2 Response time results

7.2.1 Test setup
It is always advisable to have a separate performance testing environment resembling
the production environment as much as possible. The following test setup was created
for response time tests:

Server machine:

OS: Windows 7 Enterprise 64-bit

CPU: Intel Core 2 Quad Q6600 @ 2.40GHz
RAM: 4,00GB Dual-Channel DDR2
Storage: SATA-II 3.0Gb/s Hard drive

Client1:

OS: Windows 7 Enterprise 64-bit

CPU: Intel Core 2 Quad Q9300 @ 2.50GHz
RAM: 4,00GB Dual-Channel DDR2
Storage: SATA-II 3.0Gb/s Hard drive

Client2:

OS: Windows 7 Enterprise 64-bit

CPU: Intel Xeon W3520 @ 2.67GHz
RAM: 4,00GB Triple-Channel DDR3
Storage: SATA-II 3.0Gb/s Hard drive

The default server application was SampleConsoleServer application which is

shipped in connection with Prosys OPC UA Java SDK as a sample application.
The version of SDK used in these tests was 2.2.0. An important part of OPC
UA communication is the software library called stack, which is shipped by OPC
Foundation. The version of OPC UA Java stack used in these tests was 1.02.337.5.
Oracle Java SE Development Kit was used, specifically JDK 8u05. Additionally,
in some tests UaDemo server was used. The version shipped in conjunction with
Unified Automation UA SDK C++ Bundle 1.3.3 (Evaluation Edition) was used.
The client side application was the before mentioned UaPerfClient. The computers
were connected together by a fast LAN network.

7.2.2 Read mechanism

The first read test was run with one client application reading one value attribute
from the server and the rate of read requests was varied. The maximum rate of reads
was decided to be 1000 reads per second (one read per millisecond). The minimum
 38

Table 11: One client reading one value from server

rate (reads Average (ms) Max (ms) 90th percentile (ms) Throughput
per second) (responses
per second)
100 10.3 203.1 16.0 99
200 10.0 218.5 16.0 199.7
300 11.6 218.7 16.1 299.6
400 11.9 218.6 16.2 399.5
500 11.9 203.0 16.3 499.4
... ... ... ... ...
1000 12.9 218.6 16.9 998.7

amount worth of testing was 100 requests per second. The duration of the test was
180 seconds.
Table 11 shows that on average it takes 25 percent more time to read one value
when the rate increases to 10 times the original. Considering the 90th percentile
value, the difference is even smaller. It is noticeable that there is a good reason why
the percentile values are widely used. The 90th percentile value is growing logically
with more reads and not fluctuating. Average values are also pretty reliable. Some
random variations can be seen in the maximum values from run to run, which is
expected. When interpreting these results, it also has to be considered that the client
side handling of read requests and responses is mostly counted in this time also. To
test whether or not the response times change when two clients connect to the server
at the same time, a test with a rate 1000 was run also with the other test client.
After this, the measurement was done with both test clients started at the same time
with rate being 1000. If the server is the bottleneck, connecting two simultaneous
client applications should add further delay. The tests were started manually and
it was made sure that they were started and finished during the same second time
interval. Based on table 12, the conclusion can be drawn that the OPC UA server is
not a limiting factor when client reads one attribute value per millisecond.

Table 12: Measurement with rate 1000

Client Concurrent Average (ms) Max (ms) 90th per- Throughput
centile (responses
(ms) per second)
1 no 12.9 218.6 16.9 998.6
2 no 14.0 203.1 16.6 998.7
1 yes 13.0 203.9 16.9 998.8
2 yes 14.0 219.4 16.6 998.7

The next test is to increase the amount of batch size per read request and keep
the session count still at one. In practical terms, the read rate of previous tests
(one value per millisecond) is really high. Therefore a rate of 100 reads per second
was considered. This is still an unrealistically high value considering any real world
 39

application. It is used here because this way the test can overload the server and
show rise in round-trip times. In previous test, the duration of 180 seconds was
used. This was deemed needlessly long and following read tests were executed with
the duration of 90 seconds as this seemed to provide good balance between the test
duration and reliability and consistency of results.
In the following tests, SampleConsoleServer and UaPerfClient applications were
run with explicitly defined JVM options in order to make sure that the applications
would not run out of memory. These command line options were:
Xmx2048m Xms2048m
In table 13 it can be seen that 3000 value requests 100 times per second is
clearly out of reach for the tested server. The total duration should be 90 but
it is prolonged to 120 seconds. More clearly, the round-trip values have become
unsustainable averaging at over one minute. At the level of 2500 value requests 100
times per second, the server is already pretty stressed and round-trip times have
grown noticeably. Still at this level of batchsize 2500, the throughput remains good.

Table 13: Measurement with client 1 and variable amount of nodes to read
nodesToRead Total dura- Average Max (ms) 90th per- Throughput
tion (s) (ms) centile (responses
(ms) per second)
100 90 1.8 36.4 1.9 9998.3
500 90 6.6 143.1 6.5 49988.5
1000 90 16.4 813.4 13.4 99982.7
1500 90 28.2 1108.3 17.0 149971.8
2000 90 53.8 2511.7 26.6 199928.9
2500 90 103.8 3060.6 62.5 249922.7
3000 120 64030.1 112318.4 92214.8 223641.6

Many OPC UA servers have a setting that can be used to limit the number of
simultaneous sessions. There must be some background to this setting, so testing
the possible number of sessions is an interesting part of the read functionality. In
order to reliably overload the server and be able to observe differences between OPC
UA server applications, suitable values for batchsize and rate were selected. Based
on previous tests, the number of batchsize 100 and rate 100 were deemed suitable.
Amount of sessions was varied while the batchsize and rate were kept at these values.
The test was run with SampleConsoleServer on the server machine and UaPerf-
Client application on the Client1 machine. Server application was restarted after
each test run. Results are shown in table 14. It is visible that the response times
start to degrade rapidly with rising amount of sessions. However, the user might not
notice this because only at the level of 25 sessions does the total duration start to
rise and even at this level the throughput remains pretty much at the desired level.
After 30 sessions, it is clear that the server application is not working correctly. Here
the drop in throughput is clearly visible.
 40

Table 14: Measurement with variable amount of sessions.

Sessions Total Average (ms) Max (ms) 90th percentile Throughput
dura- (ms) (responses per
tion second)
(s)
1 90 1.8 36.4 1.9 9998.3
5 90 2.8 36.5 3.7 49994.5
10 90 6.6 282.3 6.9 99984.4
15 90 11.0 512.1 10.2 149972.5
20 90 109.7 2336.5 16.6 199980.2
25 92 4164.3 7408.9 4797.1 244199.9
30 124 30204.8 54770.2 49640.6 216922.8

As in client-server application testing in general, also in this case it is important to

assess whether we are actually measuring the server performance or client performance.
If the 30 session test case from one physical machine seems to be the limit for
throughput, then the naive solution would be to try the same test with two physical
machines and 15 sessions each. Table 15 illustrates these results. It is clear that this
tested server setup can deliver about 300,000 read responses to client applications
per second. When cases with 15 and 16 clients are compared, the 16 client case
experiences actually no additional throughput, but instead faces prolonged duration
and greatly prolonged round-trip times of requests.
Based on these results it is also clear that two physical machines both hosting 15
sessions is not the same as one single application hosting 30 sessions. A part of this
difference can be attributed to the fact that the read requests are distributed more
evenly on really fine-grained level. Another reason is that the two client applications
were started manually and thus there is a small difference in their starting times.
However, there is also the chance that actually the client side application is the
bottleneck in the 30-sessions test case. Table 16 lists a result when UA Demo Server
was tested on the same hardware and using the same kind of configuration. Based
on this result it can be reasoned that the client application is not the limiting factor
in the situation depicted in table 14. Also, this result shows substantial performance
differences between OPC UA products. For UA Demo Server, the response times are
on a pretty much normal level for the 30 sessions test case.

Table 15: Measurement with two client machines and 15 sessions each
Client Sessions Duration Average Max (ms) 90th per- Throughput
(s) (ms) centile (responses
(ms) per second)
1 15 90 13.3 377.5 14.9 149967.7
2 15 90 64.3 2265.5 24.8 149986.8
1 16 95 2531.5 13275.5 8959.0 151527.3
2 16 99 2192.5 13337.5 8413.6 145111.6
 41

Table 16: Measurement with UA Demo Server.

Sessions Total Average (ms) Max (ms) 90th percentile Throughput
dura- (ms) (responses per
tion second)
(s)
30 90 9.2 259.2 15.7 299951.3

Next logical test is to test when the UA Demo Server reaches its throughput
limit. To assess this, the amount of sessions was grown from the 30. Results are
shown in table 17. Up to the level of 35 sessions from one client, the UA Demo
Server works quite correctly. At this level, again, total duration and round-trip times
start to grow. Also, the throughput level gets lower, even though only a little. Only
one session more, and on the level of 36 sessions, each reading 100 values 100 times
a second, this particular server setup will not be able to handle all requests. It is
interesting that the drop in server performance happens so fast in this situation. It
seems that about 350,000 reads per second can be handled by the server in this case
and the maximum is something below 360,000 reads per second.

Table 17: Measurement with variable amount of sessions and Ua Demo Server.
Sessions Total Average (ms) Max (ms) 90th percentile Throughput
dura- (ms) (responses per
tion second)
(s)
30 90 9.2 259.2 15.7 299951.3
31 90 9.8 325.4 16.7 309933.8
32 90 10.5 212.0 17.4 319945.6
33 90 63.5 337.4 124.9 329602.8
34 90 14.1 170.2 22.8 339949.6
35 92 181.8 401.1 238.5 348981.7
36 100 351.0 2789.0 634.4 22341.0

To test this hypothesis that about 360,000 values can be read from this particular
server per second, two physical client machines were used. Here, the two test
applications were started at the same time. Results are shown in table 18. In the
case of two clients both initiating 17 sessions, the server is not able to respond to
the load. This test shows that server throughput is something between 320,000 and
360,000 reads per second. It is noticeable that in the case of SampleConsoleServer,
two physical machines noticed larger throughput than one physical machine only. In
the case of UA Demo Server, the inverse is true. This is something that seems to be
implementation specific.
During these measurements, the CPU and memory consumption of the server
and client hosts was monitored manually. An interesting feature was noticed. The
two OPC UA servers had clearly a different kind of implementation which was
visible in the way they used CPU on the server machine. Figure 12 shows how the
 42

Table 18: Measurement with variable amount of sessions and UA Demo Server. Client
1 and 2 were used the same time.
Sessions Total Average (ms) Max (ms) 90th percentile Throughput
dura- (ms) (responses per
tion second)
(s)
13 90 8.3 97.7 13.3 129978.9
13 90 6.3 28.9 12.1 129986.8
14 90 5.3 66.3 8.6 139983.6
14 90 5.4 25.5 9.5 139986.1
15 90 8.7 61.6 15.9 149963.3
15 90 9.2 48.2 14.7 149981.2
16 90 184.2 255.1 236.0 159971.4
16 90 182.2 254.7 234.8 159613.7
17 100 273.1 2977.6 424.8 7894.8
17 100 410.8 3552.3 1863.0 13787.6

SampleConsoleServer used CPU. This shows that all CPU cores are working quite
symmetrically. On the other hand, figure 13 shows that the UA Demo Server was
working in a different fashion. One of the CPU cores is fully utilized (first from the
left in the picture), other is about 1/3 occupied (rightmost) and two cores are almost
not used at all (two in the middle). It is also evident that the memory usage differs
significantly between the servers. In figure 13 the memory usage is practically fixed
whereas in figure 12 the memory grows rapidly. These kinds of reproducible test
cases can be used to empirically quantify the benefit achieved from changes to the
server side application. One important continuation is to look at why the memory is
consumed and whether the working of the server can be changed. This however is
outside the scope of this thesis.
 43

Figure 12: SampleConsoleServer CPU usage with 300000 value attribute reads per
second (30 session reading 100 values 100 times a second)

Figure 13: Ua Demo Server CPU usage with same settings as in figure 12
 44

7.2.3 Subscription mechanism

In these subscription tests, the aim was to make as little changes as possible to
the server side. This would allow the tests to be easily run with many different
servers. However, usually there are some settings that directly prevent testing the
server. Two of these most usual settings are the maximum number for sessions and
subscriptions. In these tests, SampleConsoleServer and UA Demo Server were used.
In SampleConsoleServer, these settings are configured with the following methods:
s e r v e r . g e t S u b s c r i p t i o n M a n a g e r ( ) . s etMaxS ubscr iptio nCoun t ( 5 0 ) ;
s e r v e r . g e t S e s s i o n M a n a g e r ( ) . setMaxSessionCount ( 5 0 0 ) ;
The UA Demo Server is configured with an XML file called ServerConfig.xml. In
ServerConfig.xml, settings MaxSessionCount and MaxSubscriptionCount are used
to set these limits. The comment on MaxSessionCount says "maximum number
of sessions the server allows to create". The MaxSubscriptionCount is respectively
"maximum number of subscriptions the server allows to create". In both options,
value 0 is considered unlimited.
Subscribing to items and keeping track of subscriptions can be resource demanding
on the client side. To make sure not to run out of memory, JVM options Xmx and
Xms were used. Xmx option sets the maximum Java heap size and Xms option sets
the initial and minimum Java heap size. Setting these to the same value causes the
JVM to only use that one value for the whole test run. In all tests presented here,
the client side was run with explicit JVM memory options.
Xmx2048m Xms2048m
In the read tests previously, the original test duration was 180 seconds. This 3
minute time interval seemed appropriate also with the subscription tests. In all these
tests, if nothing different is stated, server application resides always on the "server"
machine and client application on "client1" machine. In some tests the "client2"
machine was used in addition to these. See chapter 7.2.1 for more information on
the physical machines. Because the delay time measurement depends on server and
client clocks, then the absolute value is not interesting but rather the changes when
the amount of monitored items grows. Server and client applications were restarted
between measurements.
The first test is to test response times with one session and subscription. Example
command line options used are presented below.
n " ns =2; s=MyLevel " c l i e n t s 1 s u b s c r i p t i o n s 1 i t e m s 1
d u r a t i o n 180 opc . t cp : / / 1 0 . 5 0 . 1 0 0 . 1 5 6 : 5 2 5 2 0 /OPCUA/
SampleConsoleServer
Based on table 19 it can be seen that one client initiating 100000 monitored
items is still normally handled by the server. The delay time grows somewhat, but
no other sign of lesser throughput is visible. The amount of publish requests and
responses grows, and this is probably because all data notifications do not fit into
one publish response and notifications need to be divided into multiple responses.
Publish requests are sent based on the amount of publish responses which explain the
 45

Table 19: Measurement with variable amount of monitored items. Number of sessions
and subscriptions is always one.

0.9 percentile (ms)

Publish responses

Items per second

Monitored Items

Publish requests

Avg. delay (ms)

Max delay (ms)

Max. updates
Min. updates
Data changes

Duration (s)
1 180 180 180 180 0.99 16 25 23 180 180
1,000 180 180 180000 180 999 16 24 22 180 180
10,000 181 181 1810000 180 10051 22 32 29 181 181
100,000 361 361 18050000 180 100248 38 124 44 361 361

Table 20: Measurement with two client machines and 100,000 monitored items.

0.9 percentile (ms)

Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Concurrent test

Max. updates
Min. updates
Data changes

Duration (s)
Client

No 1 360 360 17949909 180 99718 48 162 56 360 360

No 2 366 366 18000000 180 99949 100 142 108 366 366
Yes 1 460 460 18000000 180 99962 52 134 66 460 460
Yes 2 540 540 18000000 180 99962 113 166 127 540 540

grown amount of publish requests. The "Min. updates" and "Max. updates" fields
indicate the minimum and maximum amount of publish responses per subscription.
In table 19 it is clear that the delay is the only thing that really grows as the
number of monitored items grows. Based on table 20, it can be seen that throughput
remains the same even when two clients are subscribed to 100000 items at the same
time. The delay grows, but throughput is on the same level as previously. Between
these measurements the clocks were synced, so the absolute delay times are different.
What we are trying to find is whether or not the delay grows in the server side
application. The server side is not the bottleneck and it can be concluded that the
server is capable of handling 200000 monitored items simultaneously. The server
CPU usage level was 20% when 200000 monitored items were requested once per
second. The number of publish requests and publish responses is different based on
whether the tests are run concurrently. This seems weird and the reason for this
needs to be studied. However, examining this detail is outside of the scope of this
thesis.
 46

100000 monitored items each changing per second can be considered a large
amount and next we should consider the case where the number of subscriptions
is other than 1. Let us start with 1000 monitored items per subscription and start
growing the number of subscriptions. With growing number of subscriptions it quickly
becomes clear that actually creating subscriptions can be time consuming. When
creating 100 subscriptions takes 19 seconds, creating 500 subscriptions takes already
472 seconds in this setup.
Results in table 21 show that "items per second" value is not keeping up with
the actual amount of data changes when amount of subscriptions and monitored
items is growing. For example, in the case of 300 subscriptions, there should be
300000 items changing per second, and only 292314 changes are reported to the client.
From the average and 0.9 percentile delay not much can be stated other than they
do not seem to change much. From the max delay it can be stated that it clearly
gets bigger with more subscriptions. The minimum and maximum subscription ID
measurements come in handy here and show clear difference. In the case of 300
subscriptions, at least one subscription got publish response through to the client
only 147 times, although the number should be 180 for all subscriptions. The first
test case where the drop in subscription frequency is seen is the test case with 1
sessions, 20 subscriptions and 1000 monitored items. One thing that can be reasoned
from this table is that creating subscriptions requires more resources and is more
heavy than using monitored items. While the same server and client combination
handled 100,000 monitored items easily, the combination seems to be incapable of
reliably handling 100 subscriptions with 1000 monitored items each.

Table 21: Measurement with variable amount of subscriptions 0.9 percentile (ms)
Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Max. updates
Min. updates
Data changes
Subscriptions

Duration (s)

1 180 180 180000 180 999 44 53 51 180 180

10 1800 1800 1800000 180 9996 41 60 48 180 180
20 3565 3565 3569040 180 19820 52 75 60 173 181
30 5375 5374 5377778 180 29865 55 89 65 171 181
40 7059 7059 7102858 180 39445 62 102 69 171 181
50 8859 8858 8888344 180 49360 65 278 74 171 180
100 17586 17585 17710480 180 98344 44 77 51 172 180
200 35099 35099 35358603 180 196432 46 250 53 159 179
300 52404 52404 52632823 180 292314 54 955 59 147 180

Next the case where different amount of sessions is used is considered. In a normal
use case of OPC UA, the client would probably make many monitored items but
only small amount of subscriptions. Also, in previous test case it was confirmed that
 47

actually establishing subscriptions is more resource intensive than making monitored

items. For these reasons, 10 subscriptions and 1000 monitored items per sessions
were used.
Results of growing the amount of sessions are shown in table 22. The first thing
that is visible from the results is that the delay times grow pretty much logically as
the number of sessions is grown. The minimum number of updates for individual
subscription ID that client application logs are getting lower as session count grows.
With sessions 10, subscriptions 10 and monitored items 1000, the server still sends
about the correct amount of updates. When the sessions count still grows bigger,
some subscriptions just do not get all their updates through to the client. The exact
reason for this is not clear, but the symptom is clear. This seems weird and is
something that should be looked into.

Table 22: Measurement with variable amount of sessions. Amount of subscriptions is

always 10 and amount of monitored items 1000.

0.9 percentile (ms)

Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Max. updates
Min. updates
Data changes

Duration (s)
Sessions

1 1800 1800 1800998 180 10001 0 19 6 180 180

2 3601 3601 3601998 180 20003 3 22 10 179 181
3 5402 5402 5402000 180 29999 6 21 13 180 181
4 7204 7202 7201465 180 39993 8 40 15 179 181
5 8994 8991 8991000 180 49931 9 37 16 179 181
10 18004 18003 17998662 180 99935 9 54 16 178 181
20 35847 35845 35889120 180 199308 11 94 18 163 181
30 53213 53201 53315790 180 296088 14 237 22 143 186
40 68439 68431 69027462 180 383340 25 599 42 135 182
50 83273 83255 84257888 180 467908 44 744 82 142 178

To test whether or not the client side application was the bottleneck, a test case
with two physical machines hosting a test application with 50 sessions was made,
see table 23. If this test shows similar results as the test in table 22, then it can be
argued that the client side test application is the bottleneck. If on the other hand,
the throughput is lower, then server side is most probably the bottleneck. Results
show that the client side test application probably is not the bottleneck. From these
results, no definite limit for the subscription performance can be decided. At least the
test case depicted in table 23 is clearly out of the reach for the SampleConsoleServer
in the depicted hardware platform. Also, it could be argued that the test case for 20
sessions in table 22 is not sustainable as some subscriptions only get 163 responses
through whereas they should get 180 responses through to the client. In a normal
production system, the amount of data notification would probably fluctuate and
 48

it seems that the server could probably sustain momentary overload for some time
without crashing.

Table 23: Measurements with SampleConsoleServer and two physical machines

subscribing to 50 sessions each.

0.9 percentile (ms)

Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Client machine

Max. updates
Min. updates
Data changes

Duration (s)
Sessions

1 50 72584 72562 66950570 180 371901 679 2057 706 137 166
2 50 74013 73993 69228285 180 384436 303 4359 317 141 175

As in the case of read testing, a different OPC UA server was tested to see whether
or not some differences between OPC UA implementations exist. The same settings
were used as in previous testing with growing amount of sessions. The updates per
subscription show different values than previously but this is normal as the variable
in question might not be actually changing every second. These tests were started
for example with the following command line options:
n " ns =4; s=Counter1 " c l i e n t s 1 s u b s c r i p t i o n s 10 i t e m s
1000 d u r a t i o n 180 opc . t cp :// < ipa d d r e s s >:4841

Table 24: Measurements with UA Demo Server. See table 22 for comparison.
0.9 percentile (ms)
Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Max. updates
Min. updates
Data changes

Duration (s)
Sessions

1 1452 1452 1441000 180 8002 57 58 58 144 155

2 2871 2871 2871000 180 15943 59 76 61 143 145
3 4291 4291 4290000 180 23824 62 64 63 143 144
4 5711 5711 5705000 180 31682 49 66 50 142 145
5 7122 7121 7113000 180 39498 55 71 55 141 146
10 13855 13852 13843000 180 76896 65 97 65 137 141
20 20221 20200 20200000 180 112179 64 111 65 101 101
30 24976 24952 24952000 180 138617 70 163 70 83 84
40 ... ... ... ... ... ... ... ... ... ...
50 ... ... ... ... ... ... ... ... ... ...
 49

Figure 14: UA Demo Server error when 30 sessions make 10 subscriptions each
containing 1000 monitored items.

Results for UA Demo Server in the table 24 show similar results as in Sample-
ConsoleServer test up to the case of 10 sessions. The subscriptions show somewhat
different results but overall everything seems to go properly. After this, with test
cases 20 sessions and 30 sessions, the drop in throughput can be observed in the
minimum and maximum subscriptions IDs (min. updates and max. updates). Finally,
with test case of 30 sessions, the server stops working correctly. One interesting
thing is that the measured delay does not grow very much. Thus, it can be said
that the server is not able to handle all subscriptions correctly, but those publish
responses that it sends arrive at the client pretty quickly. Another difference between
SampleConsoleServer and UA Demo Server can be observed in the minimum and
maximum subscriptions IDs. Taking the 30 sessions test case as an example, the
SampleConsoleServer provides figures of minimum 143 and maximum of 186 publish
responses per subscriptions. In this same case, UA Demo Server provides correspond-
ing figures of 83 and 84. Even though UA Demo Server clearly is not sending all
publish responses, it is sending an equal amount of responses per subscription. In
SampleConsoleServer, some subscriptions send much less publish responses than
others.
Because of the error shown in figure 14, measurements with more sessions than 30
were not made. This however raises the question whether the crashing was somehow
caused by the client. To test whether or not this really was the limit of the server, a
test case was prepared where two physical client machines started 15 sessions each. In
this case the server application did not crash, which was a little surprising. This again
shows the importance of using actual physically different machines in performance
tests. Results of this test are shown in table 25. Based on this it seems that this
really was the limit that the server could handle. This is because the combined
items per second measurement shows similar value as in the test with one physical
 50

machine and 30 sessions. This is also true for the data changes measurement. Also,
minimum and maximum subscription ID frequencies are the same as in the test case
with one physical machine. This validates that the test client works and can overload
the server.

Table 25: Measurements with two physical machines subscribing to 15 sessions each.

0.9 percentile (ms)

Publish responses

Items per second

Publish requests

Avg. delay (ms)

Max delay (ms)

Client machine

Max. updates
Min. updates
Data changes

Duration (s)
Sessions

1 15 12598 12593 12593000 180 69934 156 234 157 83 84

2 15 12542 12532 12532000 180 69596 502 549 502 83 84
 51

8 Conclusions
OPC UA allows the user to configure security mode settings flexibly. Overhead of
security modes were assessed with elapsed time measurements. Adding signing or
encryption imposes overhead to OPC UA communication. With larger message sizes,
the overhead is proportionally larger. With less powerful hardware, the overhead
is larger in absolute measures but not necessarily in relative measures. General
statements in performance studies are dangerous. However it is fairly safe to say
that with desktop grade hardware the overhead of encryption is negligible. With
embedded hardware some consideration should be applied, but in a normal case
there should be no problem. Signing messages adds considerably less overhead than
encrypting messages.
With read and subscription testing, an example test client was presented. This
serves as a proof-of-concept for the idea of a generic response time test client for
OPC UA servers. The client was developed with the purpose that it could overload
server applications. Multiple differences between the tested servers were observed.
Read service testing offered an insight into how much read requests the servers are
able to handle. The tested servers proved to be able to handle hundreds of thousands
of value reads per second. In this particular setup, the Prosys OPC UA Java SDK
based SampleConsoleServer was able to handle 250,000 - 300,000 reads per second
and the Unified Automation C++ UA SDK based UA Demo Server was able to
handle 320,000 - 350,000 reads per second. The testing also showed that there are
differences between different server implementations in terms of resource usage.
In subscription testing, the limit of sessions, subscriptions and monitored items
of the two example server applications were studied. Both servers were able to
handle hundreds of thousands of monitored items changing once per second. For both
servers, the test case with 300,000 monitored items allocated to 300 subscriptions
and 30 sessions can be considered as an upper limit in this particular setup. For UA
Demo Server, this test case resulted in error which stopped the server application.
For SampleConsoleServer, the limit was observed in that some subscriptions did not
send publish responses with the desired frequency. Testing confirmed the fact that
normally multiple monitored items should be made instead of multiple subscriptions.
During subscription testing and in high load situations, the tested servers handled
subscriptions differently which was clearly visible in the minimum and maximum
number of publish responses per subscription. When an application is saturated
with more requests than it can handle, the outcome can be graceful degradation or
abrupt ending of all application functionality. The tested servers exhibited different
behaviour also in this area.
One of the research questions was to figure out what kind of issues surround
performance testing of OPC UA applications. Testing that all data notifications
arrive reliably from subscriptions under heavy load proved to be difficult. This
reliability is also perhaps the most central feature of subscriptions and thus, this
aspect should be figured out. However, with small and moderate load the tested
OPC UA servers handled all value updates reliably.
The test client developed during this thesis proved to be practical when analysing
 52

OPC UA server applications. Reproducible test cases that are easy to run proved
to be useful in product development work. During performance testing, parameters
with most effect on read and subscription performance were uncovered. Thus, it
can be argued that this thesis provides information that is valuable for product
development of OPC UA SDKs as well as to users of OPC UA applications. There
are many features that could be added to the test client. For example, support for
more services could be added. Write and browse services are the most obvious ones.
Also, more configuration options for existing services could be added. For example,
now the reads and subscriptions were always made to only one node. This thesis
concentrated on server side testing. A possible future agenda could be to broaden
the scope to client side testing.
 53

References
[1] T. Sauter and M. Lobashov, How to access factory floor information using
internet technologies and gateways, Industrial Informatics, IEEE Transactions
on, vol. 7, pp. 699712, Nov 2011.
[2] W. Mahnke, S.-H. Leitner, and M. Damm, OPC Unified Architecture. Springer,
2009.
[3] Iec webstore: Iec/tr 62541-1:2010, opc unified architecture - part 1. http:
//webstore.iec.ch/webstore/webstore.nsf/Artnum_PK/43769. Accessed:
2015-04-01.
[4] OPC Foundation. OPC Unified Architecture Specification, Part 1: Overview
and Concepts. Release 1.03. 2015.
[5] P. Piirainen, Opc ua based remote access to agricultural field machines, masters
thesis, Aalto University, 2014.
[6] A. Claassen, S. Rohjans, and S. Lehnhoff, Application of the opc ua for the
smart grid, in Innovative Smart Grid Technologies (ISGT Europe), 2011 2nd
IEEE PES International Conference and Exhibition on, pp. 18, IEEE, 2011.
[7] Intel, Ascolab, and Unified Automation, Reducing product
development effort for opc unified architecture. white paper:
http://www.ascolab.com/en/company-media/company-press-releases/150-
intel-case-study.html, 2009.
[8] Opc-ua end user considerations. https://wikis.web.cern.ch/wikis/
display/EN/OPC-UA+End+User+Considerations. Accessed: 2015-03-30.
[9] Prosys PMS Ltd. OPC UA Client Java SDK & OPC UA Server Java SDK.
2015. http://www.prosysopc.com/.
[10] D. Mosberger and T. Jin, httperfa tool for measuring web server performance,
ACM SIGMETRICS Performance Evaluation Review, vol. 26, no. 3, pp. 3137,
1998.
[11] A. Bog, Benchmarking transaction and analytical processing systems : the
creation of a mixed workload benchmark and its application. Springer, 2014.
[12] M. Fojcik and K. Folkert, Introduction to opc ua performance, in Computer
Networks (A. Kwiecie, P. Gaj, and P. Stera, eds.), vol. 291 of Communications
in Computer and Information Science, pp. 261270, Springer Berlin Heidelberg,
2012.
[13] C. Groba and S. Clarke, Web services on embedded systems - a performance
study, in Pervasive Computing and Communications Workshops (PERCOM
Workshops), 2010 8th IEEE International Conference on, pp. 726731, March
2010.
 54

[14] T. Peierls, B. Goetz, J. Bloch, J. Bowbeer, D. Lea, and D. Holmes, Java

Concurrency in Practice. Pearson Education, 2006.
[15] S. Oaks, Java Performance: The Definitive Guide. OReilly Media, 2014.
[16] K. Huppler, The art of building a good benchmark, in Performance Evaluation
and Benchmarking (R. Nambiar and M. Poess, eds.), vol. 5895 of Lecture Notes
in Computer Science, pp. 1830, Springer Berlin Heidelberg, 2009.
[17] A. Georges, D. Buytaert, and L. Eeckhout, Statistically rigorous java perfor-
mance evaluation, SIGPLAN Not., vol. 42, pp. 5776, Oct. 2007.
[18] Performance workload creation and execution framework faban. http://faban.
org/. Accessed: 2016-02-29.
[19] O. Palonen, Object-oriented implementation of opc ua information models in
java, masters thesis, Aalto University, 2010.
[20] J. Kurose and K. Ross, Computer Networking: A Top-down Approach. Pearson,
2013.
[21] OPC Foundation. OPC Unified Architecture Specification, Part 3: Address
Space Model. Release 1.03. 2015.
[22] Opc ua default information model nodes.. http://www.opcfoundation.org/
UA/schemas/1.02/NodeIds.csv. Accessed: 2015-04-20.
[23] OPC Foundation. OPC Unified Architecture Specification, Part 5: Information
Model. Release 1.03. 2015.
[24] OPC Foundation. OPC Unified Architecture Specification, Part 4: Services.
Release 1.03. 2015.
[25] OPC Foundation. OPC Unified Architecture Specification, Part 6: Mappings.
Release 1.03. 2015.
[26] OPC Foundation. OPC Unified Architecture Specification, Part 2: Security
Model. Release 1.03. 2015.
[27] K. Folkert, M. Fojcik, and R. Cupek, Efficiency of opc ua communication
in java-based implementations, in Computer Networks (A. Kwiecie, P. Gaj,
and P. Stera, eds.), vol. 160 of Communications in Computer and Information
Science, pp. 348357, Springer Berlin Heidelberg, 2011.
[28] S. Cavalieri and G. Cutuli, Performance evaluation of opc ua, in Emerging
Technologies and Factory Automation (ETFA), 2010 IEEE Conference on, pp. 1
8, IEEE, Sept 2010.
[29] M. Freund, C. Martin, A. Braune, and U. Steinkrauss, Jsuaan opc ua
javascript framework, in Emerging Technologies & Factory Automation (ETFA),
2013 IEEE 18th Conference on, pp. 14, IEEE, Sept 2013.
 55

[30] T. Burke, The performance and throughput of opc, tech. rep., Rockwell Soft-
ware Inc., 1998. White paper, available: http://www.ascolab.com/images/
stories/ascolab/doc/opc_performance.pdf last accessed 27.02.2015.

[31] T. Miettinen, Synchronized cooperative simulation: Opc ua based approach,

masters thesis, Aalto University, 2012.

[32] O. Post, J. Seppl, and H. Koivisto, The performance of opc-ua security model
at field device level., in ICINCO-RA, pp. 337341, 2009.

[33] S. Cavalieri, G. Cutuli, and S. Monteleone, Evaluating impact of security on

opc ua performance, in Human System Interactions (HSI), 2010 3rd Conference
on, pp. 687694, May 2010.

[34] Subscribing to a very large number of nodes in a prosys opc ua java

sdk based server. http://prosysopc.com/blog/subscribing-to-a-very-
large-number-of-nodes-in-a-prosys-opc-ua-java-sdk-based-server/.
Accessed: 2015-04-01.

[35] Opc ua vs opc da performance comparative. https://wikis.web.cern.

ch/wikis/display/EN/OPC+UA+vs+OPC+DA+performance+comparative. Ac-
cessed: 2015-03-30.

[36] S. Rohjans, Semantic Service Integration for Smart Grids, vol. 14 of Studies on
the Semantic Web. IOS Press, 2012.

[37] S. Chilingaryan and W. Eppler, High speed data exchange protocol for mod-
ern distributed data acquisition systems based on opc xml-da, in Real Time
Conference, 2005. 14th IEEE-NPSS, pp. 5 pp., June 2005.

[38] S. Sucic, Optimizing opc ua middleware performance for energy automation

applications, in Energy Conference (ENERGYCON), 2014 IEEE International,
pp. 15701575, IEEE, May 2014.

[39] Opc ua performance benchmarks forum discussion. http://forum.prosysopc.

com/forum/opc-ua-java-sdk/opc-ua-performance-benchmarks/. Ac-
cessed: 2015-05-14.