1. The Performance of Remote Display Mechanisms for Thin-Client Computing S.Jae Yang, Jason Nieh, Matt Selsky, and Nikhil Tiwari Department of Computer Science Columbia University Kim, Byeong Gil Software & System Laboratory @ kangwon Natl. Univ.

2. Introduction  Background shifted to the distributed model of desktop computing become unmanageble more centralized and easier-to-manage computing strategy  Purpose is to centralize computing resources. Maintain the same quality of service for the end user. Require less maintenance and less frequent upgrades. Can be shared server resources.

3. Introduction(con’t)  Improvement ~ Graphical computing environment  What do we analyze? to assess the general feasibility of the thin- client computing model to compare various thin-client platforms to determine the factors that govern their performance

4. Thin-client platforms

5. Measurement Methodology  Standard benchmarks Benchmark applications are executed on the server Benchmarks measure the server’s performance Benchmarks do not reflect the user’s experience at the client-side  slow-motion benchmarking Use network packet traces to monitor the latency and data Insert delays between the separate visual events

6. Slow-motion benchmark

7. Experimental Testbed  Composition Network emulator machine - ISDN(128Kbps), DSL(768Kbps), T1(1.5Mbps), 10BaseT(10Mbps), 100BaseT(100Mbps) Packet monitor machine - obtain the measurements for slow-motion benchmarking Thin client/server systems - used the same client/server hardware (except Sun Ray) - video resolution : 1024x768, 8-bit (Sun Ray : 24-bit) - compression and memory caching : ON - disk caching : OFF Web server

8. Web Benchmark  Modified i-Bench web benchmark introduce delays of several seconds displayed each page completely was no temporal overlap used the packet monitor  Environment Netscape Navigator 4.72 Browser’s memory cache and disk cache were enabled Netscape browser window was 1024x768

9. Video Benchmark  Playback rates 1 fps - establish the reference data size 24 fps - playback performance - video quality  Video quality(VQ)

10. Experimental Results  Default Configurations default settings demonstrate the performance of a traditional “fat” client system  Underlying baseline remote display encodings disabled configurable caching and compression mechanisms measure for experiments at 100Mbps  Caching and compression mechanisms

11. Default Configuration Web Performance

13. Default Configuration Video Performance

15. Baseline Display Encoding primitives Web Performance

16. Baseline Display Encoding primitives Video Performance

17. Caching and Compression  Environment All caching and compression options disabled All compression only options enabled All caching only options enabled All caching and compression options enabled

18. Caching and Compression Web Performance

20. Caching and Compression Video Performance

21. Memory versus DiskCaching  Memory caching provide much faster access times to smaller caches.  Disk caching provide larger amounts of local cache with relatively slower access times  Environments Platform – Citrix MetaFrame (ICA) Disk cache size – 39MB Minimum cacheable bitmap size – 8KB Memory cache size – 8MB

22. Memory versus DiskCaching(con’t)

24.  improves ICA performance at bandwidths below 768Kbps  is much faster to fetch data from the client disk cache than going across the network to the server

25. Conclusions  Higher-level graphics display primitives are not always more bandwidth efficient than lower-level- display encoding primitives.  The timing in sending display updates.  Display caching and compression are techniques which should be used with care as they can help or hurt thin-client performance.  Thin-client design and implementation choices across environments.

26. References  Primary The Performance of Remote Display Mechanisms for Thin-Client Computing - S. Jae Yang, Jason Nieh, Matt Selsky, and Nikhil Tiwari (June 2002)  Additional Measuring Thin-Client Performance Using Slow- Motion Benchmarking - S.J. Yang, J.Nieh, and N. Novik (June 2001)