1. Operation System Support for Multi-User, Remote, Graphical Interaction Alexander Ya-li Wong, Margo Seltzer Harvard University, Division of Engineering and Applied Sciences Kim, Byeong Gil Software & System Laboratory @ kangwon Natl. Univ.

2. Content  Abstract  Introduction  Background of X Windows and TSE  Approach  Processor, Memory, Network User Behavior Compulsory Load, Dynamic Load Latency  Conclusion

3. Abstract  Processor and memory scheduling algorithms are not tuned for thin client service.  Under heavy CPU and memory load, user- perceived latencies is up to 100 times.  TSE’s network protocol outperforms X by up to six times.  Bitmap cache is essential for handling dynamic elements of modern user interfaces.  Use of bitmap cache can reduce network load by up to 2000%

4. Introduction  Modern computer system architecture Allow the processor, memory, disk, and display subsystems to be spatially extruded throughout network.  Thin Client consider cost and manageability interest in X Windows-like schemes introduction of thin client service into major commercial operating systems. accelerate as consumer products.

5. Background X WindowsTSE Library Xlib GUIWin32 GUI Class User-levelPass through the kernel Multi-user YES Protocol XRDP Compression & caching None Toolkit-specific, Usually none RLE Memory & Disk platforms Windows, Unix Macintosh Windows Unix (via third-party add-ons)

6. Background (con’t)  LBX (Low Bandwidth X) is a protocol extension to X is implemented as a proxy takes normal X traffic Applies various compression techniques

7. Approach  What is the maximum number of concurrent users and what impact on users yields this maximum value? Interactive  Latency, not throughput  Latency is a key performance criterion. ( “Using Latency to Evaluate Interactive System Performance” by Y., Wang, Z., Chen – 1996) Multi-User  Benchmarking on the multi-user system Graphical  need to consider with respect to latency Remote Access  The efficiency of the network protocol

8. The key Role of Latency  Lantency characteristic Latency tolerances for continuous operations are lower than for discrete operations. Humans are irritated by latencies 100ms or greater. ( “Providing A Low Latency User Experience In A High Latency Application” by Holden, L. – 1997)  Degree factor of latency continues to increase for any operation the number of operations that induce perceptible latency increases when perceptible latency continually changes

9. Effect factors of latency  Hardware resources relevant the processor, memory, disk, network resource scarcity - the speed of the memory hierarchy level  Operating system structure bad scheduling decisions inefficient context switches poor management of resource contention  User behavior Hardware resource limitations

10. Experimental Testbed  Composition Server - 333MHz Intel Celeron system - 96MB SDRAM - 4GB IDE hard disk - Bay Networks NetGear FA-310 Ethernet adapter Client - Intel Pentium II-400 - 128MB SDRAM - 11GB IDE hard disk - 3Com Ethernet adapter Network listening host - Intel Pentium-233 - 96MB EDO RAM - 2GB IDE hard disk - 3Com Ethernet adapter

11. Processor - Behavior  From Behavior to Load Multi-user support  incoming session connections  Additional per-user kernel state  Ownership information Remote-access support Interface operations  pass through the network subsystem  Compulsory Load are inherent in the operating system

12. Processor - Load  TSE observe greater overall idle-state CPU activity. Listen for and handle incoming client connections Session state management - NT Virtual Memory, Object, and process managers

13. Processor - Latency

14. Dynamic Latency  Methodology Sink C program - never voluntarily yields the processor. - should increase the scheduler queue length by one Testing program - TSE : Notepad - X Windows : vim Action - to engage character repeat on the client machine - the rate of which was set at 20Hz Measurement - using tcpdump

15. Dynamic Latency - Results  No load the server Sending a message to the client every 50ms  Load the sever

16. Memory  From Behavior to Load Compulsory Load  Dynamic memory usage of the kernel -The system is idle with no user sessions. -17MB for Linux and 19MB for TSE  Memory usage of each user session -To be a minimal login with no additional user activity

17. Memory – Compulsory Load (con’t)

18. Memory - Latency  From Load to Latency Opened a simple text editing application remotely TSE avg is about 40 times the threshold Linux avg is about 11 times the threshold

19. Network  How user behavior generates network load compare the ability of RDP, X, LBX Growing usage of animation  How network load translates to user-perceived latency Importance of network protocol efficiency  Terms “channel” – stream of network messages between the client and server “Input channel” – stream from the client to the server

20. Network - Behavior  From Behavior to Load depends on the design and implementation of the user interfaces increasing richness and sophistication of graphical interfaces is becoming increasingly network intensive

21. Network - Load  Compulsory Load session negotiation and initialization Any network traffic is exchanged after session setup Session setup costs -TSE : 45,328 bytes -Linux/X : 16,312 bytes  Dynamic Load RDP, X, LBX Testing Environment -Corel WordPerfect -Gimp -Netscape Navigator prototap – protocol tracing software based on the tcpdump

22. Network – Load (con’t)  RDP is the most efficient protocol less than 25% of LBX and less than 15% of X  Message of LBX is to be compressed.  Average message size is just 209 bytes.

23. Network – Load (con’t)  Virtual-IP (VIP) omit the IP header can reduce overhead  LBX has the smallest average msg size. still be less than half as efficient than RDP

24. Network – Animations: Bitmap Caching  RDP outperforms LBX and X.  X and LBX does not support bitmap caching.  TSE client reserves 1.5MB of memory for a bitmap cache using an LRU eviction policy

25. Network – Cache Effectiveness and CPU Load  is not only critical to reducing network load, but also processor load at the server.

26. Network – cache  Looping animations For values 25 ~ 65 - 0.01 Mbps For all values above 65 - 0.96Mbps  LRU is the wrong scheme for handling looping animations.

27. Network - Latency  Demonstration Two simple C programs - establish a TCP connection and send and receive random data Ran ping for 60 sec. Took the average and variance in RTT. Default ping size is 64 bytes ( keystroke size)

28. Network – Latency (con’t)

29. Conclusion  Latency is the paramount performance  Highlights important issues relevant to thin client performance  Resource scheduling is not well optimized  Resource saturation rise well above human- perceptible levels  performed a detailed comparison of the RDP, X and LBX protocols  RDP is more efficient in terms os network load ( animated UI elements)