&LWUL[#,&$

)
6HUYHU#%DVHG#&RPSXWLQJ=
(OHPHQWV#RI#3HUIRUPDQFH

April 1999

A Quick Review
The Citrix Independent Computing Architecture (ICA) is a general-purpose distributed presentation
services architecture. ICA is a Citrix-invented technology that shifts the application processing from
client to server. A core technology of Citrixs WinFrame and new MetaFrame server-based
computing software, ICA helps I/S organizations reduce total cost of ownership by delivering
business-critical applications over heterogeneous computing environments, while safeguarding
application performance, data security and administrative control.
With over two million ports in use worldwide, Citrixs ICA technology has become a de facto industry
standard for server-based computing and is being broadly adopted by leading vendors for building
thin-client computing capability into a broad new range of devices.
Citrix ICA server-based computing technology includes the following equally important components:

Server Software Component - On the server, ICA separates the application logic from the user
interface and transports this to the client via standard network protocolsIPX, SPX, NetBEUI,
NetBIOS, TCP/IP & PPP. The server also provides important user administration tools,
application management and security management.

Client Software Component - On the client, users see and work with the applications interface,
but 100% of the application logic executes on the server. As a result, applications consume just a
fraction of the network bandwidth usually required.

Protocol Component - The ICA protocol transports keystrokes, mouse clicks and screen
updates to and from the client consuming typically 10 kilobits-per-second of network bandwidth.
The ICA protocol is comprised of many components at the heart of it all Thinwire. Thinwire
combined with things like Virtual Channels all work together to provide a complete user
experience and improved system administration.

The flexibility of ICA brings manageable, high-performance deployment of Server-based applications

to mobile professionals, telecommuters, branch office workers and extended enterprise users over
any kind of network connectiondial-up, LAN, WAN and Internetto any type of client or device.
This distributed architecture allows applications to perform at very high speed over low and high
bandwidth connections.

Key ICA Differentiators

Bandwidth Adaptability ICA adjusts dynamically to conditions on the network, client and
server that affect performance. Intelligence is a basic component of ICA that takes advantage of
performance conditions when bandwidth or system resources are available and economizes
when they are not without compromising the users experience and most importantly without the
user ever knowing this is taking place. We can make this happen by caching images, reducing or
eliminating unnecessary drawing instructions, and through advanced compression mechanisms.

Efficient Use of Client Resources - ICA functions on devices requiring as little as an Intel 286
processor.

Thinness on the Wire - ICA consumes, on average, 10kb of network bandwidth and is optimized
for connections as low as 14.4Kbps. This allows it to operate consistently, even over dial-up and
ISDN connections, regardless of bandwidth or application robustness.

Platform Independence. ICA is inherently platform-independent. Its modularity makes it easily

adaptable to a variety of operating systems and devices. By IDC estimates, ICA accommodates
95% of the desktop operating systems in use today. Citrix also supports various other real-time
operating systems.

Universal Client. ICA works with any Win16 or Win32 application, allowing applications to be
developed with off-the-shelf Windows tools and deployed with only one piece of ICA-based client
software.

Protocol Independence. ICA is designed to run over industry standard network protocols
including: TCP/IP, NetBEUI, NetBIOS and IPX/SPX; and over industry standard communications
such as PPP, ISDN, Frame Relay and ATM.

Citrix ICA Clients

Citrix ICA clients support the following operating platforms. These clients are available for download
from the Citrix website, located at http://www.citrix.com.

DOS

OS/2

Windows 3.1 & 3.11 WFWG

Windows 95 & 98

Windows NT Workstation

Windows CE (HPC, Jupiter & WBT)

Java

Microsoft Internet Explorer ActiveX

Control

Netscape Navigator Plug-in

UNIX: Linux, Sun Solaris (SPARC &

x86), HP-UX, IBM AIX, SunOS, Digital,
SGI IRIX & SCO

Macintosh (Motorola and PowerPC)

Additionally, Citrixs Independent Computing Architecture has been ported to RISC OS and QNX OS
embedded operating systems.

Citrix ICA Licensees

ICA is being broadly adopted by leading industry vendors. ICA licensees include Microsoft, IBM, Sun
Microsystems, HP, Motorola, Sharp, Acorn, Wyse Technology, Boundless Technologies, and a
rapidly growing number of other companies for inclusion in new and future hardware and software
products that extend the reach of enterprise applications into new markets. These and future
relationships are establishing ICA as the de facto standard for server-based computing.
Windows-based Terminal Manufacturers
Acer
Addonics (ART 2000)
Boundless Technologies (Viewpoint TC & Capio)
Bryant Computers (Sunnix 2000)
Motorola Semiconductor Products Sector (WinCept 100/100 WBT reference designs)
Neoware Systems, Inc. (NeoStation and netOS)
Netier Technologies, Inc. (NetXpress SL and XL)
Plexcom, Inc. (QuantumNet)
TECO Information Systems (Relysis TR-3300)
UMAX (NC-200 and NC-320)
VXL (Winlinx)
Wyse Technology (Winterm)
Network Computer Manufacturers
Aranex (Internet Client Station Series 2000)
IBM (Network Station)
MTX, Inc. (MTX 1683 Network Computer Terminal)
Sun Microsystems (Java Virtual Machine)
Handheld PC/Wireless Manufacturers
Compaq (Series C)
Hewlett-Packard (HP 620LX & 360LX Palmtop PCs)
Motorola (Telephones)
Sharp Electronics Corporation (Mobilon line of handheld PCs)
Symbol Technologies (PPT 4300 portable pen terminals)
Telxon Corporation (wireless, ruggedized pen-based computers)
Information Appliance Manufacturers
Acorn Group plc (DeskLite reference design)
Boca Research, Inc. (BocaVision set-top box)
Clientec (Computer keyboards)
Key Tronic Corporation (Computer keyboards)
Websonic (universal communicators)
Operating System Suppliers
Caldera
NCI
QNX
Red Hat
This list constantly growing as more device and software manufacturers join the opportunities
created by Citrix server-based computing.

Performance Measurement
Measuring performance is not quite the same as what might typically be considered under other
legacy methods of computing. In the case of ICA, application activity does not occur on the users
desktop, but this is where the user experience takes place. Measuring this computing mechanism
brings with it numerous challenges because it involves all of the components described on the first
page of this document; the server, network, and the client desktop. Measuring any one of these
items may actually be easier than the very interactive and dynamic combination of all of these
components.
To help you understand the ICA performance spectrum, Citrix has created the material below to
outline what needs to be considered. We begin by trying to outline what users experience, then
defining what performance actually is using ICA or any other remote presentation protocol, like
RDP. We conclude with some helpful information to improve the accuracy of measurements.

Whats important to the Customer?

Citrix customers all share the following items in common relative to their expectations from our
MetaFrame and WinFrame server-based computing solutions and the client-side ICA interaction with
this software:
1. Usability
Responsiveness and perceived speed of operation when performing typical tasks and
features providing a PC-like experience.
2. Efficient use of the Network
Minimizing the amount of data and number of packets generated.
3. Client-side Resource Utilization
Minimizing client-side CPU and memory utilization.
4. Server-side Resource Utilization
Minimizing server-side CPU and memory utilization. Gathering data to indicate server
resources required and maximum supported users (scalability).

Does the Customer Environment Affect Performance?

YES. Remember that all of these items interact to provide the user with their performance
experience. During measurement of performance, all points above need to be measured with
different:

Network bandwidths,
Numbers of users,
Applications,
Interactive operations with other servers or services (e.g., back-end DB servers, the Internet,
etc.)
Target client types,
Server configurations.

Elements of Performance
Expanding on the previous page, the following items elaborate on the elements of performance. Each
and every one of these items contributes to the user experience. Depending on who you are (e.g., the
person using the application, the IT manager, the network manager, etc.), the focus may be on different
components. But the composition of all of these items is important to the collective business enterprise.
For example, if perceived user performance being equal between a Citrix solution and an alternative
occurs using a fraction of the network resources, then performance is not equal. Understandably, the
same result using fewer resources does so more efficiently and therefore with greater performance.
Speed

Measured as time to visibly complete a given set of actions and benchmark

reported figures. Time to visibly complete the operation is used because the
server can be fooled into thinking that a set of display operations has completed
but the client still has to display most of the data. This renders server-side
measurement tools inaccurate.

Network
efficiency

Network efficiency for a given set of operations are measured as:

Total KB transmitted to and from the server. This determines when the network
will saturate.
Number of packets used to transmit that data. High numbers of packets will
cause collisions and network saturation at a lower bandwidth and reduce the
available bandwidth.

NOTE: It is important to realize that the amount of data and packets sent/received
will change depending on the other performance elements. There are frame or
graphics dropping techniques that will reduce data transmission but sometimes at
the expense of jerky graphics presentation. The recommendation is that a varying
number of configurations be tested before making statements on bandwidth
usage. The number of packets transmitted could be lower but possibly at the
expense of responsiveness
Server load

Server load includes memory usage and CPU usage. This should change as the
number of active sessions increases. For instance, it may be fine for a single
session to utilise 50% of the server CPU, when it is readily available, but not okay
when many sessions are active. Dont assume linear usage of server resources
based on single session data. Especially due to multi threading on multi CPUs
and the memory caches. Also, low CPU usage may not be efficient if the time it
takes to complete a task consequently increases.
Both WinFrame and MetaFrame leverage the economy of scale achieved by
using shared memory resources. By benefiting from existing memory contents,
duplicate images are reduced. Memory usage rarely scales linearly with the
number of users.

Client load

Client load includes memory & CPU usage. These can be effectively measured on
some devices; for instance Windows NT & UNIX but not all.

Responsiveness

This refers to how quickly the results of a given user action produce the resultant
change in displayed graphics (latency). A good example is after clicking on a
menu button how long it takes for the entire menu to be displayed. A bad
example is how long it takes a file to be loaded and displayed since the loading
time is not greatly affected by the client/server solution. Some aspects of
responsiveness are objective and not easily measured. For instance the way in
which a large bitmap is displayed may be pleasing or displeasing and lead the
user to feel that one client is more responsive than another even though they
both continue on to display the next graphical element after the same elapsed
delay. As mentioned above under Network Efficiency, some techniques involve
throwing away data in order to meet low bandwidth requirements but these lead to
jerkiness in interactive operation although producing good benchmark figures.

Total user
experience

The feel experienced when using applications software should be similar to, or
better than, that experienced when running the same application on the server
console. To what extent is the entire server console environment reproduced for
the client user? This includes device support, port support, audio and integration
features such as cut/copy/paste.

Effect of
Network

How do all of the above factors vary between high bandwidth LANs and low
bandwidth WANs? Also, what affect do various settings have using a client that
can be tuned for the bandwidth utilized?

Common Performance Measurement Pitfalls

IMPORTANT!
Many of the typical benchmarking programs available today were originally
designed to test the performance of a single user PC system (e.g., WinTach, WinBench, Winstone,
Sysmarks, etc.). However, In a server-based computing environment where client, network, and
server are all elements being measured, these benchmarking applications are not very well suited for
measurement of remote display protocols. Instead, these may actually be a better measure of server
performance (which is quite variable) rather than client or protocol performance. At this time, there
are no real server-based computing benchmarking applications available to help overcome this issue.

All figures tend to suffer from noise. This is partly due to the fact that operating systems such as NT
will use resources differently dependent on what was run on the system previously or even the exact
point in time after a reboot that a given operation is performed, mostly due to caching. It is therefore
important to take an average of multiple runs. It is also important to reboot and perform the same
series of operations for each run. When enough multiple runs are taken you can then use statistics to
determine if the difference seen in the different setup is due to chance or not. Please note that even
minor differences in configurations can make a dramatic difference in the results due to caching
algorithms.

In real user environments users often do perform the same operations time after time. It is therefore
also important that repetitive tasks are performed to see if the system learns from experience.

Short benchmarks are very easy to fool with large buffering systems. A minimum test of at least 5
minutes is recommended.

Benchmarks that use normal applications often have a fixed overhead of time during which files are
being loaded or calculations being performed. These are not affected by the client or protocol used
and can therefore make 2 clients appear closer in performance than they actually are.

Performance can come at some cost to other parameters. For example, a client that performs twice
as fast as another might consume a disproportionate amount of additional server CPU in doing so.
One would consider a system superior that gives an improvement in performance without a
proportionate increase in resources usage.

Never rely on benchmark statistics. The benchmarking software often thinks that operations have
finished as soon as the server based client driver tells it. It may well be some time before the client
user sees the end of the test. Check visually whether the test has finished on the client and use hand
timings to check correlation with reported benchmark numbers

Performance Testing Set-up Recommendations

The information below is intended to help conduct performance testing in a manner that provides
consistent, reliable and credible results.
Server Setup:
The server should be setup to be as independent as possible from other machines (i.e., it should be a
stand alone server-using local users not domain users, it should not require any files from a file server,
where possible all the default options should be taken in the setup). Each client connected should be their
own user. Another server component of performance affecting testing outcome is the presence of
applicable hotfixes and service packs the server should be as up to date as possible.
Benchmarks to use:
As stated earlier, the proper benchmarking tools do not yet exist from the major benchmarking labs like
Ziff-Davis or National Software Test Laboratories (NSTL, a division of CMP). So these recommendations
are made for utilizing the alternatives (e.g. WinTach, WinBench, ISAT, etc.).
Which networks to use:
We recommend using a variety of connection mechanisms as part of performing any tests because
results will vary for each of them. Connections should include at least the most common:
10Mbps LAN
Modem
WAN (if available, but please note the issues involved here as discussed in the first item above
relative to introducing noise.)
How to measure speed:
If automatic client and server side time stamping is possible for determining this would naturally be more
accurate. If not, then a stopwatch based on the client display actions, noting reaction times. Benchmarks
that are long enough mitigate most reaction time discrepancies by making any margin for error a relatively
small percentage.
How to set up the client:
Default options should be used were ever possible. If tests are applicable to low bandwidth
environments, then appropriate WAN mode settings should be applied.
How to ensure that you have the latest ICA client version:
The Citrix Download website contains the most recent versions of all of the ICA clients available.
Please use this to obtain any clients used for benchmarking analysis. Please include references to this
version number in any written material.
How to measure server load:
Server load can be measured on the server itself. This adds a small extra load but that can be measured
and subtracted if necessary. This can be better than trying to measure it remotely and generating extra
network traffic. The Citrix Resource Management Service product can provide an excellent source for
this kind of information particularly over long-term test periods.

The Server load data should not be recorded to disk as this puts a heavy load on the server. Ideally the
data should be collected automatically over the period of time for the benchmark. It should also be an
average for the whole benchmark. NTs performance monitor is not suitable to collect the data because
of the significant CPU overhead this tool adds on its own. Citrix uses a modified version that collects
individual values and keeps a running total and averages the collection period. Data should be collected
every second to capture short-term peaks in the average. Data should be collect on a per user basis.
This permits analysis separate of the load associated with a users action versus the load on the console
due to the data measurements, and the system processes load.
The items measured should be measured using a total system CPU load percentage measurement so
that a correct measurement is made even when there are multiple processors in a system.
How to measure client load:
This is best done in a similar manner to the server load measurement and therefore can be done best on
an NT based client.
How to measure network traffic without affecting the actual figures:
This is best done on a separate machine with the network card in promiscuous mode. This permits
receiving all network traffic on the network segment without interference. You should not try and save or
real-time decode the network data, instead just keep a running total of the amount of data/packets sent
during the benchmark. This is best achieved with something like Network Generals Basic Sniffer Network
Analyser.
How to run the benchmarks:
Ideally there should be a reboot of the server and client after every run, with the mouse pointer in the
client as it would be if a user were using the client. The mouse pointer should also be moved by the
benchmark and all user input should be generated from the client side. However, this makes it
difficult to synchronise the client with the server automatically (e.g., responses to items like yes/no, menu
pull-downs, etc.).
How to measure responsiveness:
Responsiveness should be measured by inputting user actions at the client and timing how long the
associated graphics change takes to happen. This should measure very basic types of user actions like
typing, menu operation, scrolling, maximizing, and minimizng within real applications.
How to measure non-graphics performance:
Disk I/O is easily measured with the Ziff Davis Winbench tool which has built-in disk tests that simulate
real applications that can be used on drive remotely mapped to the clients, network drives and the
servers local drives. Other types of non-graphical performance is difficult to measure as there are no
standard techniques defined and the performance varies by the type of printer and printer driver used.

Reporting Recommendations
A time will come to document any extensive work conducted as a result of this document. As indicated
earlier on, the variability of configurations makes every set-up unique. As a result, details on the
configurations should include the following:

Network configuration (protocol, speed, hardware components, servers attached, clients attached,
etc.)

Server configuration (CPU, memory, bus architecture, cache, processors, subsystems, etc.)

Client configuration (Hardware configuration, operating system, version, etc.)

ICA client type used, version number and build number