Upgrading and Performance Analysis of Thin Clients in Server Based Scientific Computing

Institutionen för Systemteknik

Department of Electrical Engineering

Examensarbete

Upgrading and Performance Analysis of

Thin Clients

in Server Based Scientific Computing

Master Thesis in ISY Communication System

By

Rizwan Azhar

LiTH-ISY-EX - - 11/4388 - - SE

Linköping 2011

Department of Electrical Engineering Linköpings Tekniska Högskola

Linköpings universitet Linköpings universitet

Upgrading and Performance Analysis of

Thin Clients

in Server Based Scientific Computing

Master Thesis in ISY Communication System

at Linköping Institute of Technology

By

Rizwan Azhar

LiTH-ISY-EX - - 11/4388 - - SE

Examiner: Dr. Lasse Alfredsson Advisor: Dr. Alexandr Malusek Supervisor: Dr. Peter Lundberg

Presentation Date

04-02-2011

Publishing Date (Electronic version)

Department and Division

Department of Electrical Engineering

URL, Electronic Version

http://www.ep.liu.se

Publication Title

Upgrading and Performance Analysis of Thin Clients in Server Based Scientific Computing

Author

Rizwan Azhar

Abstract

Server Based Computing (SBC) technology allows applications to be deployed, managed, supported and executed on the server and not on the client; only the screen information is transmitted between the server and client. This architecture solves many fundamental problems with application deployment, technical support, data storage, hardware and software upgrades. This thesis is targeted at upgrading and evaluating performance of thin clients in scientific Server Based Computing (SBC). Performance of Linux based SBC was assessed via methods of both quantitative and qualitative research. Quantitative method used benchmarks that measured typical-load performance with SAR and graphics performance with X11perf, Xbench and SPECviewperf. Structured interview, a qualitative research method, was adopted in which the number of open-ended questions in specific order was presented to users in order to estimate user-perceived performance.

The first performance bottleneck identified was the CPU speed. The second performance bottleneck, with respect to graphics intensive applications, includes the network latency of the X11 protocol and the subsequent performance of old thin clients. An upgrade of both the computational server and thin clients was suggested.

The evaluation after the upgrade involved performance analysis via quantitative and qualitative methods. The results showed that the new configuration had improved the performance.

Keywords

SBC, Performance analysis, Xbench, X11perf, SPECviewperf,

Language

X English

Other (specify below)

55 Number of Pages Type of Publication Licentiate thesis X Degree thesis Thesis C-level Thesis D-level Report

Other (specify below)

ISBN (Licentiate thesis) ISRN:

LiTH-ISY-EX - - 11/4388 - - SE

Title of series (Licentiate thesis) Series number/ISSN (Licentiate thesis)

Upphovsrätt

Detta dokument hålls tillgängligt på Internet – eller dess framtida ersättare – under 25 år från publiceringsdatum under förutsättning att inga extraordinära omständigheter uppstår.

Tillgång till dokumentet innebär tillstånd för var och en att läsa, ladda ner, skriva ut enstaka kopior för enskilt bruk och att använda det oförändrat för ickekommersiell forskning och för undervisning. Överföring av upphovsrätten vid en senare tidpunkt kan inte upphäva detta tillstånd. All annan användning av dokumentet kräver upphovsmannens medgivande. För att garantera äktheten, säkerheten och tillgängligheten finns lösningar av teknisk och administrativ art.

Upphovsmannens ideella rätt innefattar rätt att bli nämnd som upphovsman i den omfattning som god sed kräver vid användning av dokumentet på ovan beskrivna sätt samt skydd mot att dokumentet ändras eller presenteras i sådan form eller i sådant sammanhang som är kränkande för upphovsmannens litterära eller konstnärliga anseende eller egenart.

För ytterligare information om Linköping University Electronic Press se förlagets hemsida

http://www.ep.liu.se/.

Copyright

The publishers will keep this document online on the Internet – or its possible replacement – for a period of 25 years starting from the date of publication barring exceptional circumstances.

The online availability of the document implies permanent permission for anyone to read, to download, or to print out single copies for his/her own use and to use it unchanged for non-commercial research and educational purposes. Subsequent transfers of copyright cannot revoke this permission. All other uses of the document are conditional upon the consent of the copyright owner. The publisher has taken technical and administrative measures to assure authenticity, security and accessibility.

According to intellectual property law the author has the right to be mentioned when his/her work is accessed as described above and to be protected against infringement.

For additional information about Linköping University Electronic Press and its procedures for publication and for assurance of document integrity, please refer to its www home page:

http://www.ep.liu.se/.

Abstract

Server Based Computing (SBC) technology allows applications to be deployed, managed, supported and executed on the server and not on the client; only the screen information is transmitted between the server and client. This architecture solves many fundamental problems with application deployment, technical support, data storage, hardware and software upgrades.

This thesis is targeted at upgrading and evaluating performance of thin clients in scientific Server Based Computing (SBC). Performance of Linux based SBC was assessed via methods of both quantitative and qualitative research. Quantitative method used benchmarks that measured typical-load performance with SAR and graphics performance with X11perf, Xbench and SPECviewperf. Structured interview, a qualitative research method, was adopted in which the number of open-ended questions in specific order was presented to users in order to estimate user-perceived performance.

The first performance bottleneck identified was the CPU speed. The second performance bottleneck, with respect to graphics intensive applications, includes the network latency of the X11 protocol and the subsequent performance of old thin clients. An upgrade of both the computational server and thin clients was suggested.

The evaluation after the upgrade involved performance analysis via quantitative and qualitative methods. The results showed that the new configuration had improved the performance.

Acknowledgement

In the beginning, unlimited thank to Almighty ALLAH, THE most Merciful and Beneficent, without Whom I would not be able to complete this thesis.

I would like to thank my advisor Dr. Alexandr Malusek, for his guidance and magnificent technical support. Many thanks for his availability even on weekends and for his effort to make this thesis more interesting.

In addition, I also want to thank Dr. Peter Lundberg for providing this opportunity in a very professional and competitive environment.

Also, I would like to thank my examiner Dr. Lasse Alfredsson, who explained to me the skills of technical writing. I also like to thank Mr. Shehryar Khan for proof reading of the thesis report.

Last but not the least, tremendous gratitude to my parents for the support, love and prayers. I dedicate this thesis to my parents, especially the daily motivation that I have received via Skype, without which I would not be able to complete this thesis.

Acronyms

API Application Programmable Interface DDI Device Driver Interface

DNS Domain Name System

DRI Direct Rendering Infrastructure DXPC Differential X Protocol Compressor GDI Graphics Device Interface

GUI Graphical User Interface

GNOME GNU Network Object Model Environment ICA Independent Computing Architecture IceWM Ice Window Manager

IOSTAT Input Output Statistics IPC Inter Process Communication ISAG Interactive System Activity Grapher KDE K Desktop Environment

NFS Network File System

NIS Network Information Service

NTC-O New Thin Client with original software NTC-T New Thin Client with ThinLinc Software OpenGL Open Graphics Library

OTC-O Old Thin Client with original software RDP Remote Desktop Protocol

RFB Request Frame Buffer SAR System Activity Report SBC Server based Computing

SPEC Standard Performance Evaluation Corporation SSH Secure Shell

TCC Thin Client Computing

TCP Transport Communication Protocol TLCOS ThinLinc Client Operating System VNC Virtual Network Computing VMSTAT Virtual Memory Statistics XCB X Protocol C Language Binding XDMCP X Display Manager Control Protocol

Table of contents

Chapter1: Introduction 1.1 Introduction…….………..1 1.2 Aims………..2 1.3 Layout………...3 Chapter2: Background 2.1 Introduction..………...5 2.2 Advantages of SCB………...6 Manageability………...6 Security………...7 Scalability………...7 Availability………...7 Cost reduction...…………...7 2.3 SBC Implementation……….7 2.3.1 Hardware Implementation………..7 Thin client..………7 Workstation ………...7 2.3.2 Software Implementation………...8 RDP………...8 ICA………9 NX………...9 VNC……….10 X Window system………...11

X client libraries and extensions………...12

Mesa3D………….……...12 DRI………...12 XSync………...12 XFT2………...12 XCB………...12 ThinLinc………...13

Chapter3: Performance analysis tools and benchmarks 3.1 Introduction………...15

3.2 Performance analysis tools………...15

IOSTAT………...15 VMSTAT………...16 SAR………...17 ISAG……….18 SPEC CPU………...18 3.3 Graphics Benchmarks………...18 X11perf………...18 Xbench……….18 SPECviewperf………...19

Chapter4: Typical-load performance analysis

4.1Introduction………..………21

Old computing environment……….………...22

New computing environment……….…….………...24

4.2 Methods………...24 4.3 Results………...24 CPU Utilization………...25 Paging statistics………...26 IO transfer rate………..………...27 4.4 Discussion………...28 4.5 Conclusion………..29

Chapter5: Graphics performance analysis 5.1 Introduction……….31

5.2 Methods………...32

5.3 Results………...34

5.3.1Xbench……….………34

Effect of desktop environment……....…..………...37

Effect of remote display………...38

5.3.2 X11perf………...39

5.3.3 SPECviewperf………....42

5.4 Discussion………...43

5.5 Conclusion………..43

Chapter6: User perceived performance analysis 6.1 Introduction……….45

6.2 Methods...………....45

6.3 Results……….46

User interviewing before upgrade.…...………...46

User interviewing after upgrade..……..………..……...47

6.4 Discussion………...47 6.5 Conclusion ……….47 Chapter7: Conclusion 7.1 Conclusion………..49 Appendix A………...50 Appendix B………...52 Appendix C………...53 Bibliography………..55

List of Figures

Figure1: Thin client computing model………....2

Figure2: SBC Environment………...6

Figure3: RDP Architecture………...8

Figure4: NX Architecture………...10

Figure5: VNC………...………...10

Figure6: X Architecture………...11

Figure7: IOSTAT command output………...16

Figure8: VMSTAT command output………16

Figure9: SAR command output………....17

Figure10: Schematic view of network configuration………23

Figure11: CPU utilization graph on a) nestor and b) tintin………...25

Figure12: Paging statistics graph on a) nestor and b) tintin………..26

Figure13: IO transfer rate graph on a) nestor and b) tintin………...27

Figure14: Xbench performance on a) KDE b) Gnome and c) IceWM………...35

Figure15: Xbench performance ratio on a) KDE b) Gnome and c) IceWM……….36

Figure16: Desktop environment performance on NTC-O………37

Figure17: Remote display performance………38

Figure18: X11perf performance on a) KDE b) Gnome and c) IceWM………....40

Figure19: X11perf performance ratio on a) KDE b) Gnome and c) IceWM………41

List of Tables

Table1: IOSTAT parameters……….16

Table2: VMSTAT parameters………..16

Table3: Old Server and workstations configuration……….………....22

Table4: Old Thin clients configuration……….22

Table5: New server configuration……….24

Table6: New thin client configuration………...24

Table7: SPECint results…….………...28

Table8: SPECfp results...………...28

Table9: Experimental design for X11 and Xbench………...32

Table10: Experimental design for remote display………32

Table11: Desktop environment versions………...33

Table12: Experimental design for SPECviewperf………33

Table13: Ratio of average performance of Desktop environment………37

Table14: SPECviewperf results on Desktop environment………42

Table15: Performance evaluation based on user experience before upgrade…...………46

Table16: Performance evaluation based on user experience after upgrade…...………...47

Table17: Xbench tests………...51

Table18: Thin client technical specification comparison………..53

Table19: Considered alternatives for computational server...………...53

1

1

Introduction

Concepts presented in this section come from an article by Yu et al [1]. Computing technology can be divided in to the three distinct phases of evolution:

1. Mainframe Computing: Mainframe computing involves processing power to be condensed in a centralized system with all functionalities and resources. This type of architecture was popular during the 1960s up to the 1970s. Drawbacks: Slow link speed, simple terminal and high cost-performance ratio.

2. Standalone Computing: Standalone computing consisted of PCs and workstations in

a client-server configuration through a PC LAN interface. These systems were equipped to handle a wide range of applications as more processing power and independent storage units were invested in the client. This reduced the load on the server which was used for additional file and print services. This type of architecture was popular during the 1980s. Drawbacks: Maintenance cost has increased, as applications were installed and executed on the client.

3. Network-centric Computing: Network centric computing is a decoupled

client-server configuration that enabled machines to access applications and data on the servers. This type of architecture was developed in the 1990s. Dominant form of a network centric computing is an Internet computing, which has changed the way applications were developed. Advantage: No applications have to be installed on the client. Disadvantages: Applications have to be reconfigured or redeveloped for the network architecture.

2

As applications have to be reconfigured for an Internet computing, the thin client computing models were developed to resolve the compatibility issues of data intensive applications with the widely popular internet computing architecture.

Client/Server Model Web-based Model SBC Model Figure 1: Thin client computing models

Figure 1 illustrates the Client server model, the Internet computing model and the Server Based Computing (SBC) model. The overall architecture of the three models is similar as all the application code is executed on the server side and the client devices are used to display screen updates. However the models use different protocols and data compression techniques for the communication between the client and the server. The models differ with respect to the functionality partitioning where the effort has been made to make the client layer thin. This impacts the networking protocol and data compression demands to reduce the network traffic and the client side processing. The SBC model is the thinnest model as it is only used to display screen updates.

1.2 Aims of this thesis

Scientific calculations at the MR-unit were CPU intensive and required large amount of computer memory. Powerful computational servers were used for scientific calculations and Unix workstations were used for data processing. Up to three students worked in a small computer room at the MR-unit. Heat and noise produced by the workstations would have made working conditions in this room difficult. Therefore only monitors, keyboards and mice were placed in the computer room; the workstations were placed in a separate "machine" room and connected via long (about 18 m long) cables. This configuration worked well for Sun workstations with display resolutions up to 1280 x 1024 pixels. For higher resolutions, however, either very expensive DVI extenders or a solution based on thin clients were needed. The latter was implemented in 2007 using thin clients in an SBC configuration. This infrastructure worked well for several years. It became obsolete in 2010 as the server operating system was no longer supported by the manufacturer and the speed of the thin clients was not sufficient for new, graphics intensive applications. An update of the IT infrastructure was needed.

3

The purpose and scope of this thesis is to (i) analyze the current computer configuration at the MR-unit of the Department of Radiological Science (ii) propose an upgrade (iii) implement the proposed changes and (iv) analyze the new computer configuration at the MR-unit.

1.3 Layout

This section defines the layout and a brief introduction of the thesis. Chapter1: Introduction

This chapter describes the evolution phases of computing technology, the thin client computing models, and aims of the thesis.

Chapter 2: Background

This chapter introduces the SBC, its advantages, components and various protocols based on SBC.

Chapter 3: Performance analysis tools and benchmarks

This chapter defines the performance analysis in general, tools and benchmarks used for the performance analysis of the SBC environment.

Chapter 4: Typical-load performance analysis

This chapter briefly presents the load performance analysis, the methods and the results, discussion and conclusion.

Chapter 5: Graphics performance analysis

This chapter describes the graphics performance analysis of the implemented SBC, the methods and the results, discussion and conclusion.

Chapter 6: User perceived performance analysis

This chapter presents the results and discussion on the performance analysis based on user experience via qualitative research method.

Chapter 7: Conclusion

5

2

Background

2.1 Introduction

Concepts described in sections 2.1 and 2.2 are based on information provided in [2]. Server-based computing (SBC) is a technology whereby applications are deployed, managed, supported and executed on the server and not on the client. Only the screen information is transmitted between the server and the client.

SBC is also known as Thin Client Computing (TCC) and typically consist of the three components: a terminal server, one or more thin clients, and a communication protocol. The terminal server runs an operating system that supports a multi user environment. The thin client runs a stripped-down version of an operating system that is capable of running a program that connects the thin client to the server. The last component is the protocol which allows the communication between thin clients and the terminal server by sending keystrokes, mouse clicks and screen updates via network. In larger deployments, more than one terminal server may be used; load balancing software then distributes workload across the servers.

6

Figure 2: Server based computing environment

SBC implements a centralized computing model, in which users share software and hardware resources of the dedicated servers via thin clients. In this way SBC minimizes administration costs and also increases the utilization of resources, some advantages of SBC are discussed in section 2.2.

2.2 Advantages of SCB

Advantages of server based computing are manageability, security, availability, scalability and cost reduction which are discussed in this section.

Manageability

An SBC environment implements a centralized computing model in which the data deployment and the management is more convenient as most of the work is performed on the servers only.

7

Security

In SBC, all the data is kept on secure servers. Data security is covered from the perspectives of both the physical security and the technical security. The physical security in the sense that server normally resides in a server room, in which only authorized persons are allowed to enter. The technical security is provided by the use of protocols to keep the data integrity and confidentiality.

Scalability

In an SBC environment, new thin clients can be added to the system to support more users without changing the architectural design of the system. Similarly new servers can be added to balance the load of the system without changing the architecture of the system.

Availability

Servers can support fault tolerance. For instance features like redundant disk drives, redundant power supplies etc. Moreover they can run special applications to manage the load and monitor the performance. In case of a server failure, the user is redirected to a different server.

Cost Reduction

Server based computing reduces costs on account of hardware, software, maintenance, administrator staffing and training etc.

2.3 SBC Implementation

There are two aspects of an SBC implementation: SBC Hardware implementation and SBC Software implementation.

2.3.1 SBC Hardware Implementation

There are two ways of SBC hardware implementation namely (i) thin client and (ii) workstation or personal computer (PC).

Thin Client

A thin client is a computing device which totally relies on a server for computing. All application logic is executed on the server. A thin client comprises of a display screen, a mouse and a keyboard combine with the adequate processing capabilities and memory for the graphical rendering and the network communication with a server. It has no long-term user state and requires no disk.

Workstation

An ordinary workstation or Personal Computer (PC) can also be used for an SBC. The workstation has to be installed with an SBC supportive operating system like the TLCOS [27]. Alternatively, if a full OS is not feasible at the workstation then a client side application

8

of the SBC supporting software being used at the server, has to be installed on the workstation.

2.3.2 SBC Software Implementation

The most important component of the SBC is a software protocol which is used to facilitate the communication between a thin client and a server. There are several protocols for SBC some of them are discussed in this section.

RDP

Remote Desktop Protocol (RDP) is a protocol developed by Microsoft which provides remote access services.At the server side, RDP utilizes a virtual display driver to render the display information. After creating the local display information, this information is pushed in network packets with the help of RDP protocol and then forwarded across the network to the client side. At the client side, RDP receives the packet via network and transforms these network packets into appropriate Graphics Device Interface (GDI) Application Programmable Interface (API) calls. In this way, GUI is displayed on the client side. Upon displaying the GUI the user response is activated on the input path. User responses consisting of mouse clicks and keyboard events are then forwarded to the server. On the server side, RDP takes an advantage of its own keyboard and mouse driver to collect the screen updates. RDP uses sixty four thousand virtual channels for the data communication with multipoint transmission [28].

Key strokes, Mouse ops Bitmaps, Colour tables,

Ordering ops, Drawing Ops, etc

Client Server Figure 3: RDP Architecture

Performance of the RDP is affected by display encoding, encoding compression, display update policy and client caching. For the display encoding, RDP uses graphics primitives just like Windows DDI video driver interface. The encoding mechanism uses higher-level semantics and potentially differentiates between various graphic primitives including boxes, glyphs, fills etc. The higher-level semantics saves bandwidth but requires an additional processing at the client side. RDP uses the combination of run-length encoding and other compression schemes for the compression of display encoding. For the display update policy, buffers are used at the server side. On the client side, contents are flushed at a varying rate which depends on the input from the client side and the graphical output produced at the server side. RDP also uses client side caching to accelerate the display of graphical objects. RDP 4.0 and 5.0 clients uses 1.5 MB of RAM for graphical objects including glyphs, bitmaps etc. RDP 5.0 also uses 10 MB persistent disk cache for caching just like paging system [5].

Cache

9

The RDP uses RC4 encryption algorithm to secure the data communication which provides three different levels of security namely low, medium and high. The RC4 low level security provides unidirectional encryption with the help of a 56 or 40 bit key encryption. RC4 medium level security uses bidirectional encryption with the same key length used in low level security. RC4 uses high level security and provides bidirectional encryption with 128 bit key encryption [6].

ICA

Independent Computing Architecture (ICA) is a protocol developed by Citrix to provide remote access service. On the server side, ICA uses a local device driver to render graphics. For encoding the graphical information, it uses a graphic primitive which is similar to a Windows video driver interface and RDP. The encoded ICA packet is then compressed with a combination of run length coding and other compression techniques. The TCP protocol encapsulates the compressed ICA packet and forwards it to the network layer, for creating the session between the client and the server side. On the client side, the ICA client software intercepts the server message via the TCP protocol and fetches the screen from the server. As a result, the session is created and the ICA client software in return sends the screen updates to the server. On the server side, ICA uses a screen update policy with the help of buffers. The contents of the buffer flushed with changing rate according to the user input in the form of key strokes, mouse events and the amount of display output being generated. To observe buffered display an evaluator named Speed-screen is used. On the client side, it uses a queuing mechanism for sending and receiving updates form the server. The data transmission rate can be changed in the client software. The ICA client uses a caching mechanism before downloading updates from the server, which increases the transport performance. The cache is first checked for updated information and if the update is true it downloads from cache instead of downloading from server, across the network [5].

Just like RDP, the ICA performance also depends on display encoding, encoding compression, display update policy and client caching. The display encoding, encoding compression and display update policy characteristics are similar to RDP. However, for client caching, ICA uses 3.0 MB of RAM for graphical objects and it also uses a disk for persistent caching [5].

NX

G.F Pinzari of NoMachine originally developed an NX which uses the SSH protocol to securely connect and display the remote access services. The NX was derived from the Differential X Protocol Compressor (DXPC). It uses a two way proxy system to provide the data communication between the local system and the remote system with the help of the NX protocol based on the X window system. The remote proxy communicates with the remote X11 applications and uses the X11 protocol. The remote proxy incorporates a kind of X server, named nxagent, that becomes the X server of all the remote X clients. The nxagent translates the X11 protocol into the NX protocol, receiving X11 requests as drawing commands and sending them to the local proxy. Thus the remote NX proxy poses itself to the X clients as if it is the X server. All the roundtrips take place at this point and since they happen on the same host machine, they are quickly resolved through the UNIX domain sockets. The local NX proxy communicates with the local X server and interprets the NX protocol back to X11. Both the local and the remote proxy keep their own cache of

10

transferred data. These two sets of cache are synchronized to be useful for saving all transmission of pixmaps, icons etc between the local and the remote proxies [9].

Figure 4: NX Architecture

NX also has a potential to run over slow network connections with the help of compression mechanism which reduces the round trip time and increases the drawing speed of the X window system. When NX receives a new X request, it evaluates the checksum and tries to find the request in the message-store. The Message-store is a cache for holding the lasts X messages sent to the main memory. After searching, if the message exists in the message store, the NX sends the status information combined with the position information of message store and disparity encoding of all the fields except the checksum. In this way NX achieves the compression boost [9].

VNC

Virtual Networking Computing (VNC) is used to provide the remote desktop sharing with the help of Request Frame Buffer (RFB) protocol. It mainly comprises of the three components: a VNC viewer, a VNC server and a RFB protocol. A VNC viewer resides on the user side which transmits the keyboard strokes and the mouse events on a network. The VNC server sends a remote display as the rectangular bitmap images to the VNC viewer. VNC uses different encoding methods to control the bandwidth utilization. The most undemanding method is a raw encoding, in which the VNC server sends the data in left-to-right scanline arrangement, until the full-screen transmission of the display, and after that only transfer those pixels that varies on the keyboard and the mouse events. This encoding scheme is effective for a small portion of screen changes on every updates and suffers a lot for large variation in screen positions. RFB protocol works with the TCP protocol on port 5900+N; where N is the number which is used for every remote display attached. The protocol is platform independent, as it runs on frame-buffer level, which allows it to executing on different operating systems for example Unix, Linux, MAC and Windows.

11

X Window System

The X window system is a network-transparent window system, which was developed at the MIT in 1984. It provides a Graphical User Interface (GUI) for both the local users as well as the network based users. The X window system is based on the client server model. The reference view in an X window system is the application and not the end user and this change in reference causes an inversion of the client server model with respect to the traditional sense. The user's terminal now acts as the server and the application acts as the client, as applications are now end users of services therefore behave as the clients.

There are two scenarios; if the client and the server reside on the same system they interact with the Inter Process Communication (IPC) but if the client and the server reside on the different system a network connection is used. Simultaneous multiple connections are supported for two cases, a single client can open several connections to a server and several clients can open several connections to various servers at the same time. The X-server manages the display and the clients; it multiplexes requests from the clients to the display and supports the reverse by de-multiplexing the keyboard and the mouse inputs back to the appropriate clients. Typically, the server is implemented as a single sequential process, using round robin scheduling among the clients [10].

Figure 6: X windows architecture

The X server communicates with different X client programs by calling Xlib. The Xlib is an X window system protocol client library written in the C programming language. The Xlib forwards a request from the client to the server by accumulating the request in a buffer known as the Request buffer. The buffer is flushed when all the requests are forwarded to the server. The Xlib not only stores the send requests in a buffer but also accumulates received events in a queue. This queue is used by the client application to examine and retrieve events. The client application may also flush the request buffer incase of the queue blocking [11].

12 X Client Libraries and Extensions

The X window system consists of various libraries and extensions. Libraries and extensions related to graphics performance are discussed in this section.

MESA3D

Mesa is an open source implementation of OpenGL for rendering the interactive 3D graphics. The rendering can be performed purely in the software but if the graphic card supports hardware acceleration of 3D graphics then the Mesa can utilize these functions too [14].

DRI

DRI is an acronym for Direct Rendering Infrastructure, which enables the hardware acceleration by accessing video hardware without requiring data to be passed through the X server [29].

XSYNC

X Synchronization is an extension which provides primitives that allows the synchronization between the clients to take place entirely within the X server [30]. This extension eliminates network errors and provides the real time synchronization for the multimedia applications.

XFT2

Xft2 is a freetype2 library which is used for rendering fonts. It uses render extension for the accelerated text drawing. However if the render extensions are not available then it uses the core protocol to draw the client-side glyphs [14].

XCB

XCB is the C language library for an X window system. It is an alternative to Xlib, designed to solve the issues of the Xlib by providing latency hiding, threaded applications compatibility and small platform compatibility. The latency hiding issue is solved by redesigning the network layer protocol from synchronous to an asynchronous interface. The small platform compatibility is solved by having an XCB as a binding instead of trying to reduce the large Xlib code. The XCB is designed keeping threaded applications in mind; Xlib API was error-prone. In addition, the XCB provides a basis for toolkit implementation, direct protocol-level programming, and lightweight emulation of commonly used portions of the Xlib API. Toolkits provide user interface elements, examples are GTK+, Qt, Motif, Intrinsics, etc [13]. GTK+ is a rich toolkit in features, which is initially designed for the GUI of X window system. The toolkit is written in C language which supports various languages like Python, C++, etc. It uses an Xlib to display graphics on the screen. GTK+ implementation comprises of the three libraries Glib, Pango and ATK. Glib provides a low level portability layer and common functions which is used by GTK+ and Gnome, the Pango provides an internationalized text layout and the ATK deals with the accessibility. Versions of GTK+ are available as GTK+1.1, GTK+1.2, GTK+2.0, etc and the latest stable release is GTK+ 2.20 [12].

13

Qt is a toolkit for creating the GUI and the Non-GUI programs. The toolkit is widely used because it supports many platforms (X11, Windows, Mac OS X etc) and has APIs that are tailored for tasks such as file handling, process handling threading etc. The toolkit is written in C++ and also supports many different languages with the help of language bindings. Qt also provides an internationalized text and the unicode support. Versions of Qt are available as Qt1, Qt2 etc and the latest Qt4 has most accessibility support [12].

The X protocol does not use any compression mechanism for the transport of the graphical objects on the remote display. The important factor in the performance of X protocol is a transport latency which depends on the round trip time; from the X-server to the X-client and back to the X-server. The Xlib increases the transport latency as it mostly uses the synchronous operation; however, the design of XCB reduces the latency by introducing the asynchronous interface in the X protocol 14].

ThinLinc

ThinLinc is a software based implementation of an SBC, developed by Cendio AB. Thinlinc not only enhances the mechanism of an SBC but also provide various add-ons which include encryption, single sign-on, clustering, smartcard support, and integrates different systems and services [26].

On the server side graphic is generated with the local video device driver which transfers it on the network to the ThinLinc client by the graphical desktop sharing system known as Virtual Network Computing (VNC). VNC uses the Request Frame Buffer (RFB) protocol which transmits mouse clicks, keyboard events and graphics updates from the VNC server to the VNC viewer. Thinlinc provides both the software and the hardware acceleration for 3D graphics. Software based acceleration is provided by the rapid integration of the OpenGL with the 3D graphic library Mesa. Hardware based acceleration is provided with the combination of TightVNC and the virtualGL. TightVNC is a slightly modified version of VNC which uses Junior Picture Expert Group (JPEG) codec for the tight encoding; which gives intense graphics performance and extends streaming video encoding in the real time. VirtualGL allows remote access applications like tightVNC to execute the OpenGL commands with the full 3D hardware acceleration. Rendering of 3D data is performed on the application server with the 3D graphics accelerator. The rendered 3D images are sent on the network to the user side with the help of the TightVNC.

A user use thinlinc-client to log on to the thinlinc server; in this process the thinlinc-client creates a secure shell (SSH) tunnel to the thinlinc-server. The client then attempts to authenticate with the VNC session manager known as VSM server. If the authentication succeeds, the client then again creates a SSH tunnel to the VSM agent. After the connection establishment with the VSM agent the tunnel for VNC viewer is then created to start the VNC viewer [26].

15

3

Performance

analysis tools and

benchmarks

3.1 Introduction

Performance analysis is the process of evaluating and measuring the software/hardware behaviour, with the help of information gathered, when the specific program/hardware operates. Performance analysis not only evaluates the performance metric, but also indicates and predicts the future requirements of the system.

3.2 Performance analysis tool

There are various tools and techniques to analyze the performance of Linux based SBC. Some of the tools are discussed in this section.

IOSTAT

IOSTAT acronym means Input Output Statistics. IOSTAT is used to monitor the system performance with respect to the CPU and the disk utilization. The tool records the number of blocks read and written per second to each disk connected to the given system. It also calculates the time spent by the system in user, system, idle, steal, iowait, and nice mode [17]; see figure 7 and table 1.

16 rizwan.azhar@tintin:/> iostat

Linux 2.6.31.5-0.1-default (tintin) 09/13/10 _x86_64_ (8 CPU) avg-cpu: %user %nice %system %iowait %steal %idle

3.40 0.19 0.91 0.16 0.00 95.34

Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 1.00 7.03 21.52 25166094 77090737 sdb 0.23 4.35 78.12 15588060 279827257 scd0 0.00 0.26 0.00 936216 0

Figure 7: IOSTAT command output Table 1: Quantities reported by IOSTAT.

CPU

%user Percentage of CPU utilization that occurred while executing at the user level.

%nice Percentage of CPU utilization that occurred while executing at the user level with nice priorities.

%system Percentage of CPU utilization that occurred while executing at the kernel level.

%iowait Percentage of time that CPU was idle during which the system had an outstanding disk I/O request.

%steal Percentage of time spent in involuntary wait by the virtual CPU while the hypervisor serves another virtual processor

%idle Percentage of time spent by CPU in which do not have any process to execute

Disk I/O

Tps Input/Output transfers per second to the device blk_read/s Number of blocks read from the device per second blk_wrtn/s Number of blocks written to the device per second blk_read total number of blocks read from the device blk_wrtn total number of blocks written to the device

VMSTAT

VMSTAT acronym means Virtual Memory Statistics. VMSTAT is used to determine the resource utilization with respect to processes, virtual memory, paging activity, disk transfers and CPU activity. The tool uses kernel tables and counters to evaluate the resource utilization with the help of files like /proc/meminfo, /proc/stat, /proc/*/stat and then presents the result [18], see figure 8 and table 2 for more information.

rizwan.azhar@tintin:/> vmstat

procs ---memory--- ---swap-- ---io---- -system-- ---cpu--- r b swpd free buff cache si so bi bo in cs us sy id wa st 0 2 60124 525984 258968 38071688 0 0 1 6 0 1 4 1 95 0 0

17

Table 2: Quantities reported by VMSTAT

Process

R Number of processes waiting for run time B Number of processes in uninterruptable sleep

Memory

Swpd Virtual memory used (in kibibytes) Free Virtual memory idle (in kibibytes) Buff Memory (in kibibytes) used for buffering Cache Memory (in kibibytes) used for caching Paging

Si Amount of memory swapped in from disk (per second) So Amount of memory swapped to disk (per second) IO

Bi Blocks (per second) receive from block device Bo Blocks (per second) send to block device System

In Total number of interrupts (per seconds) including clock Cs Total number of context switches (per second)

CPU

Us Percentage of CPU utilization that occurred while executing at the user level

Sy Percentage of CPU utilization that occurred while executing at the kernel level.

Id Percentage of time spent by CPU in which do not have any process to execute

Wa Percentage of time that CPU was idle during which the system had an outstanding disk I/O request.

St Percentage of time spent in involuntary wait by the virtual CPU while the hypervisor serves another virtual processor

SAR

SAR acronym means System Activity Report. SAR collects data reports and records the system activity information. The activity information includes CPU utilization, IO statistics, memory statistics, paging statistics, system statistics, network statistics etc. For data collection, the system activity data collector (sadc) command is used which samples data from kernel tables and counters. The sa2 command is used to write a daily report activity generated from the sadc. As a result, the tool monitors major system resources see figure 9 for more information.

The tool writes to a standard output, the contents of selected cumulative activity counters in the operating system. The accounting system, based on the values in the count and interval parameters, writes information in the specified number of times spaced at the specified intervals in seconds [19].

rizwan.azhar@tintin:~> sar -u 2 5

Linux 2.6.31.5-0.1-default (tintin) 11/09/2010 _x86_64_ (8 CPU) 03:32:45 PM CPU %user %nice %system %iowait %steal %idle 03:32:47 PM all 0.92 0.00 0.41 0.00 0.00 98.67 03:32:49 PM all 0.06 0.00 0.22 0.00 0.00 99.72 03:32:51 PM all 0.09 0.03 0.09 0.06 0.00 99.72 03:32:53 PM all 0.06 0.00 0.13 0.00 0.00 99.81 03:32:55 PM all 0.06 0.00 0.28 0.06 0.00 99.60 Average: all 0.24 0.01 0.23 0.03 0.00 99.51

18

ISAG

ISAG acronym means Interactive System Grapher. ISAG does not evaluate the performance, however, it helps in the performance analysis, because ISAG displays the system activity information graphically. The data being generated from SAR are transformed in to tabular form and the tabular form is then plotted with the gnuplot [20].

SPEC CPU

The SPEC CPU is designed to provide performance measurements that can be used to compare the computing intensive workloads on different servers [25]. It includes two benchmark sets namely CINT and CFP. The CINT benchmark is used for measuring and comparing computing intensive integer performance, The CFP measures floating point performance.

3.3 Graphic benchmarks

There are various graphic benchmarks for the Linux based SBC. Some of the graphic benchmarks are discussed in this section.

X11perf

The X11perf is a graphic benchmark, which is used to analyze the performance of the X server. The benchmark executes various performance tests on the X server and reports the performance for executed tests. X11perf is a comprehensive benchmark which performs 384 tests to calculate the performance, for instance drawing of lines, circles, rectangles, copy-plane, scrolling, stipples, tiles etc. The benchmark not only measures the traditional graphic performance, but also measures the window management performance. X11perf tests measures the time to initialize an application, mapping of windows from already existence window to new windows, reorganizing the windows in different positions, mapping of bitmaps in to pixels etc. In addition, it also explores the particular strengths and weaknesses of servers, to analyze and improve a server performance. X11perf benchmark when executed, it firstly calculates the round trip time to the server, and parts this out of the final timing reported. It makes sure that the server has really executed the task requested by fetching a pixel back from the test window [21].

X11perfcomp is a utility of X11perf which helps in the graphics performance analysis, as it compares the different X11perf result files and represents the result in a tabular form. As a result, the X11perfcomp is used to compare the performance of different platforms.

X-bench

X-bench is a graphical benchmark which is used to analyze the performance of the X server. To analyze the performance it executes 40 tests which include drawing of lines, circles, windows etc on the X server. The X-benchmark measures the performance by calculating the time it takes to complete each and every individual test on the X server. Each test sends the instructions with awaiting known time interval normally ten seconds. Xbench remain silent till the server executed the given instruction with the help of XSync, then the server interacts with a graphic-controller using another instruction via fifo/pipe/buffer. In the end, Xbench examines back part of the screen. Examining back part of the screen assures that each pixel is

19

displayed on the X server and does not reside in the instruction queue of server. The Xbench execute tests in three levels which can be set via command line interface, however, if no level is defined the default level is used. Each test executes three times and the best evaluation is taken. As a result, it removes or reduces the effects of daemons or other background processes [22].

SPECviewperf

The SPECviewperf is a benchmark developed by the Standard Performance Evaluation Corporation (SPEC) that evaluates an OpenGL graphic performance. It is a software program written in C language which is used to evaluate system performance in higher-quality graphics modes and also measure the scalability of graphics subsystems while running multithreaded graphics content. SPECviewperf performs various tests and measures the performance in frames per second. It renders the data set in the predefined frame numbers, with animation between the frames displayed on the output screen. The benchmark use viewset to characterize the graphics representation of an application. Viewset comprises of tests, data sets and weights which are grouped with independent software vendors (ISV) that provide weight-age for each test through which performance is reported in the output. Viewset include cataia, ensight, maya, pro, maya, solid works, ugx-nx, 3ds-max. For Detailed viewset description see [23]; however we have only discussed Ensight, Maya and Proe viewset. Ensight is a viewset which is used to evaluate display list and immediate mode paths with the help of OpenGL API. Maya viewset uses a blend of the immediate mode and the OpenGL to transfer the data through an OpenGL API. Proe viewset uses two models and three rendering modes for evaluating the graphics. It also uses gradient background for the efficient modeling of workload [23].

SPECviewperf measures the performance for the following entities:

 3D primitives, including points, lines, line_strip, line_loop, triangles, triangle_strip, triangle_fan, quads and polygons

 Attributes per vertex, per primitive and per frame

 Lighting  Texture mapping  Alpha blending  Fogging  Anti-aliasing  Depth buffering

Alpha blending is used for combining a transparent foreground color with a background color, as a result creates a new blended color. Fogging is a process which is used to improve the acuity of distance. Anti-aliasing refers to a process that reduces the appearance of aliased or jagged edges, produced due to fixed resolution of the computer display.

21

4

Typical-load

performance

analysis

4.1 Introduction

Performance analysis of a load is described in chapter 3; the main ideas can be summarized as follows: Load measurement refers to the practice of assessing a system’s behavior under the load. The Load measurement may be performed during design, development, integration, and quality assurance phase. Performance analysis of a load evaluates the functional and performance problems of a system under the load. Performance problems include those scenarios in which the system experience low throughput or long response time. Functional problems include bugs in the system and deadlocks situations. Load is applied on the system with the help of the programs or the applications which utilizes system resources. The load of the system is measured by means of logs and reports generated by the monitoring application, which is already installed in the system. The monitoring application uses operating system kernel tables and counters to calculate the system resource usage information which includes CPU utilization, memory consumption, Disk I/O requests, paging activity etc.

This chapter describes the old and new computer system configurations at MR unit and performance analysis of typical load in these two configurations.

22

Old computer environment at the MR-unit

The computer environment at the MR-unit consisted of servers, workstations and thin clients, see tables 3 and 4 and figure 10. Nestor and milou were computational servers, anka was a file server, abdallah and tchang were workstations. Three thin clients were connected to nestor via a private network; they were not assigned any DNS names.

Table 3: Servers and workstations at the MR-unit (old computer environment)

Host name Architecture CPU RAM Operating

System

milou 64 bit 2 x Opteron 1.8

GHz

4 GiB openSUSE 10.3

abdallah 64 bit Pentium D 2.8

GHz

1 GiB openSUSE 10.3

tchang

32 bit Intel Pentium 4 2 GHz

1 GiB openSUSE 10.3

nestor 64 bit Quad core Xeon

5150 2.66 GHz

32 GiB openSUSE 10.3

Table 4: Old thin clients at the MR-unit

Machine CPU RAM CPU cache Network

IGEL-2110 LX Smart

Via Eden 400 MHz

236 MiB 128 KiB Ethernet,

23

Figure 10: Schematic view of the network configuration at the MR-unit

Nestor was the main computational server at the MR-unit. It was used to run applications like jMRUI, Spm8, Idl, Analyze, Lcmodel, Sage, and Matlab. Thin client users used the XDMCP protocol to connect to the nestor. Workstation users used SSH to run commands remotely on nestor.

Milou had been the main computational server at the MR-unit before 2007. In the beginning of 2007, it was replaced in this role by the nestor. Since then it was used as a secondary computational server.

Thin clients were connected to the Gigabit Ethernet switch, see figure 10. A remote session via the XDMCP protocol was automatically opened on the nestor when the thin client was turned on.

Anka provided file services at CMIV and the MR-unit; nestor, milou, tchang and abdallah mounted user home directories via NFS version 3. It also provided the database of unix users via the NIS service. Hostnames were resolved via DNS; the master and slave servers at the MR-unit were nestor and milou, respectively.

24

New computer environment at MR unit

In the new computer environment the computational server nestor was replaced with tintin, see table 5. The new server was selected as the fastest server-class machine that fitted in to the limited budget see table 19 in Appendix C for considered alternatives.

Table 5: New server configuration at the MR-unit

Host name Architecture CPU RAM Operating

System

tintin 64 bit 2 x Intel Xeon Quad core E5560 2.8 GHz

48 GiB openSUSE 11.2

The new thin clients were selected according to a recommendation by Peter Ästrand from Cendio AB, for having good price-performance ratio and software support from the manufacturer; see table 20 in Appendix C for available models and table 6 for the parameters of the chosen system.

Table 6: New thin client at the MR-unit

Machine CPU RAM CPU cache Network

Fujitsu S550 AMD Sempron 200U, 1 GHz

1 GiB 256 KiB Ethernet,

1 Gbps

4.2 Methods

To evaluate the performance of a typical load, SAR was used. The ISAG utility was used to display the data graphically. To create a typical load on both the servers, a MATLAB script was used. The script comprised of numerous calculations based on Magnetic Resonance Imaging (MRI) of brain and different parts of the human body. The load tests were executed on the nestor and the tintin during a weekend to eliminate the users’ influence on the system load.

4.3 Results

25

CPU Utilization

Figure 11a shows the CPU utilization during the load test. The process was started at 16.30 and took 2.12 hours to execute. The utilization of 35% indicates that the one CPU core was fully loaded while another one was loaded up to 50% only. We theorize this was caused by the inner working of how Matlab executed the script. The script put a low demand on network and disk resources. Therefore the task was mainly CPU intensive.

Similar results were obtained for tintin, see figure 11b. The test started at 18:30 and took 1.32 hours to execute. The process fully utilized one core and 18% of another core. The speed up in total run time was 2.12 1.32 1.6.

(a)

(b)

26

PAGING STATISTICS

Figure 12a shows paging and fault statistics obtain during the load test on nestor. The amount of paging and major faults was negligible. The rate of minor faults reflected the activity of a data processing job.

For tintin, the rate of minor faults was substantly lower compared to the nestor, see figure 12b. The physical memory size was sufficient to meet the dynamic memory requirements of the load test.

(a)

(b)

27

I/O Transfer Rate

Figure 13a shows the Input Output transfer rate during the load test on nestor. Substantly low disk IO requests were reflected. On the average, the network IO requests transmitted and received per second during the load test were 208.03 and 150 Kbps.

For tintin, the similar results were obtained, see figure 13b. The disk IO requests were slightly reduced compared to the nestor. The average network IO request transmitted and received per seconds during the load test were 313.15 and 271.21 Kbps.

(a)

(b)

28

4.4 Discussion

The overall throughput of the new system was higher compared to the old system for the following reasons: (i) individual cores of the new system were faster than cores of the old system. (ii) The new system had 8 cores while the old one had only four cores.

The load test showed that tintin was 1.6 times faster than nestor. This result did not match with the standard test result of SPEC-CPU 2006 because the SPECcint and SPECfp benchmarks evaluates the integer and floating-point performance, however, the Matlab load test was the combination of integer and floating-point. Some benchmark results were closer to the result which includes hmmer, xalancbmk, gcc in SPECcint and tonto benchmark in SPECfp, see table 7 and 8 for SPECcint and SPECfb benchmarks result respectively.

Table 7: Ratio of SPECcint benchmark results on nestor and tintin

SPECcint benchmarks Performance

nestor in t P P_int Perlbench 1.26 bzip2 1.25 Gcc 1.54 Mcf 2.40 Gobmk 1.24 Hmmer 1.57 Sjeng 1.52 Libquantum 12.11 h264ref 1.25 Omnetpp 1.77 Astar 1.35 Xalancbmk 1.55

Table 8: Ratio of SPECfp benchmark results on nestor and tintin

SPECfp benchmarks Performance

nestor in t P P_int Bwaves 4.37 Games 1.29 Milc 3.44 Zeusmp 2.12 Gromacs 1.40 Cactusadm 9.26 leslie3d 2.45 Namd 1.34 Dealii 1.38 Soplex 2.34

29 Povray 1.54 Calculix 1.94 Gemsfdtf 3.90 Tonto 1.73 Ibm 3.68 Wrf 2.42 sphinx3 2.23

4.5 CONCLUSION

The results showed that the load test was mainly CPU intensive. The range of load at the MR unit is very wide. Some applications are purely CPU intensive and some of them only perform mathematical operations on large number of files. The results generated from SAR and visualized by ISAG indicated that the new computer system of the MR unit is 1.6 times faster than the old one.

31

5

Graphic

performance

analysis

5.1 Introduction

Graphics performance may affect the overall performance of applications. This aspect is especially important in the server based computing, because the speed of the server, thin clients, network has to be properly balanced. One way of measuring the overall graphics performance is the usage of benchmark programs, which evaluate the performance of both the hardware and the software implementations of visualization routines.

It is well known that desktop environment may play a significant role in the graphical performance; the more visually rich the environment is the slower the response of applications may be. On Linux, typically used desktop environments are KDE and Gnome. The X window system, however, does not require the usage of a full desktop environment; a full featured window manager like IceWM may be used instead. To keep the discussion in this chapter simple, we use the term “desktop environment” for IceWM also. List of all tested desktop environments is in table 11.

32

5.2 Methods

Graphics performance was evaluated using graphics benchmarks X11perf, Xbench and SPECviewperf, refer to chapter 3 for more information about the benchmarks.

Xbench and X11perf simulations were executed from KDE, Gnome and IceWM on the old thin client with the original software (OTC-O), new thin client with the original software (NTC-O), new thin client with the Thinlinc (NTC-T), the computational servers were nestor and tintin. In the case of Xbench, the AWK programming language was used to process the output files. For X11perf, the X11perfcomp was used to merge the output data in a tabular form. Additionally, an R script (see appendix A and B) was written to visualize the data.

Table 9: Experimental design for benchmark tests with Xbench and X11perf. Values in the

table give the number of repetition for each test.

Desktop Environment

Xbench X11perf

OTC-O NTC-O NTC-T OTC-O NTC-O NTC-T

KDE 3 3 3 2 2 2

Gnome 3 3 3 2 2 2

IceWM 3 3 3 2 2 2

To estimate the effect of remote display i.e how the network transfer slows down the graphics device performance, the Xbench benchmark was also used in the configuration described in table 10.

Table 10: Experimental design for the estimation of remote display. Values in the table give

the number of repetition for the Xbench test.

X server/ Display device Server

Tintin Nestor

33

Table 11: Desktop environments at MR unit.

Server KDE version Gnome version IceWM version

Tintin 4.3 2.28 1.12

Nestor 4.0 2.20 1.00

SPECviewperf benchmarks were executed on computational servers tintin and nestor. The benchmark failed to execute on NTC-O, OTC-O because some OpenGL extensions were missing. However it successfully ran on NTC-T and it was also executed on tintin with KDE, Gnome and IceWM to compare the three desktop environments.

Table 12: Experimental design for SPECviewperf benchmark. Values in the table give the

number of repetition for each test. NTC-T did not run IceWM because ThinLinc didn’t support it. Desktop Environment SPECviewperf Nestor Tintin NTC-T KDE 2 2 2 Gnome 2 2 2 IceWM 2 2 -

As each benchmark was executed several times see table 9, 10 and12. The averageRavg



  n i i avg R n R 1 1 , (1)

of resulting values was taken to evaluate the performance of these benchmarks. Where Ri is

the time taken by each test of the benchmark to execute on the X server. The performance, P, was defined as the natural logarithm of the mean number of graphical objects plotted during one second interval. It was evaluated as

avg R

Pln , (2)

where Ravg is the average number of graphical objects plotted per unit time and s stands for

one second (the unit of time). The performance was evaluated in terms of natural logarithm to scale down the variability of results of individual test. Uncertanity of the average value was calculated with the standard deviation as



    n i avg i R R R n n 1 ) ( ) 1 ( 1  , (3)

34

Since each of the benchmarks executed tests multiple times the resulting uncertainties of averages were low (0.01%). They are not reported here.

5.3 Results

The results of the graphic benchmarks Xbench, X11perf and SPECviewperf are discussed in this section.

5.3.1 Xbench

The Xbench simulation executed from KDE environment revealed that the best performance was achieved with NTC-T. NTC-O had slightly higher performance compared to OTC-O, see Figure 14 (a). The average performance of NTC-O relative to the average performance of OTC-O was 12 times higher. The average performance was taken as the average of individual test performances. For NTC-T this ratio was 25. NTC-T was the fastest, followed by NTC-O and OTC-O in this order, see Figure 15(a).

The NTC-T had highest performance, when Xbench simulations were executed from Gnome. NTC-O had higher performance than OTC-O except for five tests 29, 30, 31, 35 and 38 out of forty-one tests executed in the simulation, see Figure 14 (b). The ratio of the average performances of NTC-T to OTC-O was 22. The same quantity for NTC-O and OTC-O was 10, see Figure 15 (b).

The OTC-O had slightly improved performance for IceWM as compared to KDE and Gnome. NTC-O had better performance except for ten tests 8, 9, 20, 22, 25, 29, 30, 31, 34 and 38 out of forty-one tests executed in the simulation, see Figure 14 (c). The ratio of average performance of NTC-O to OTC-O was 8, see Figure 15 (c).

35

(a)

(b)

(c)

Figure 14: Xbench performance on the old thin client (blue circles) and on the new thin client

with (green circles) and without (red circles) the ThinLinc software installed. The 40 Xbench tests are ordered so that the performance of the old thin client as a function of test number does not decrease in KDE (a), Gnome (b), and IceWM (c).

0 10 20 30 40 8 10 12 14 16 18 test number p e rf o rm a n ce NTC-O NTC-T OTC-O 0 10 20 30 40 8 10 12 14 16 18 test number p e rf o rm a n ce NTC-O NTC-T OTC-O 0 10 20 30 40 8 10 12 14 16 test number p e rf o rm a n ce NTC-O OTC-O

36 (a)

(b)

(c)

Figure 15: Xbench performance relative to that of OTC-O for the old thin client (blue circles)

and for new thin client with (green circles) and without (red circles) the ThinLinc software installed. The 40 Xbench tests are ordered so that the performance ratio of the old thin client as a function of test number does not decrease in KDE (a), Gnome (b), and IceWM (c).

0 10 20 30 40 1. 0 1. 2 1. 4 1. 6 1. 8 test number pe rfo rm an ce re la tive to O TC -O NTC-ONTC-T OTC-O 0 10 20 30 40 1. 0 1. 2 1. 4 1. 6 1. 8 2. 0 test number pe rfo rm an ce re la tive to O T C -O NTC-ONTC-T OTC-O 0 10 20 30 40 1 .0 1 .2 1 .4 1 .6 1 .8 2 .0 test number p e rf o rm a n ce r e la tive to O T C -O NTC-O_OTC-O