1. PERFORMANCE OF THIN-CLIENT MULTIMEDIA IN A LOW BUDGET ACADEMIC ENVIRONMENT Michael Gonzales Damian Clarke

2. Objective Tests the performance of web-based multimedia and video on thin clients and X-Windows sessions Focus was placed on multimedia platforms, HTML 5 and Flash 10 The hardware architecture included Sun Ray Thin Clients and RDP X- Windows session in a Linux environment connected to a Solaris server Assess the feasibility of both multimedia platforms as a potential education solution for content delivery 1

3. Motivation Evolution of Computing Environment (Back to the past?) Large and centralized >>> small and distributive >>>> back to small and centralized. Cost A distributive specialization approach could address issues of cost-effectiveness maintenance performance reliability security 2

4. Motivation Student Expectations convenient access to information; anytime, anywhere and on demand access Native vs. Web Application Zero footprint installation (Silverlight work well here too) Cross-platform compatible (Yes, you can use one of the cross-platform widget libraries too with native code, but Web is much easier.) Server-centralized control of logic, resources (data, CPU, et al), etc. Public facing application with highly sensitive data that needs to stay behind a firewall 3

5. Overview 4

6. Related Works J. Nieh, S. J. Yang, and N. Novik, “A Comparison of Thin-Client Computing Architectures” Tested 5 thin clients Multimedia-intensive workloads degrades significantly over WAN compared to LAN Higher-level encodings were shown to vary widely in graphic- intensive multimedia with rapidly changing images. 5

7. Related Works A. Lai and J. Nieh, “Limits of Wide-Area Thin-Client Computing” Introduced a non-invasive slow-motion benchmarking technique: packet monitoring and slow motion versions of application benchmarks. Non multimedia-intensive over WAN Thin client computing can deliver acceptable performance 6

8. Related Works Y. Chang, P. Tseng, K. Chen, and C. Lei, “Understanding The Performance of Thin-Client Gaming” Methodology for quantifying the performance of thin clients on gaming Display frame rates and frame distortion were critical factors in game performance. Differences in thin-client implementations may have different levels of robustness against network impairments 7

9. Related Works A. Y. Wong and M. Seltzer, “Operating System Support for Multi-User Remote, Graphical Interaction,” Effects of OS support for multi-user remote graphical interaction and the performance via user-perceived latency Resource scheduling for both the processor and memory in these systems is not well optimized for heavy concurrent interactive use, with latency and jitter well above human- perceptible levels. 8

10. Uniqueness of this Work Systematically focused on server instrumentation due to the proprietary and closed-source nature of the thin client platform used. None of the previous work analyzed the behavior of the server at runtime on the kernel level. Multimedia platforms: HTML5 and Flash 9

11. Background: HTML5 HTML5 is a synthesis of several new and existing technologies (primarily JavaScript, CSS3, Canvas, and SVG), Allows deployment when the Flash plugin is not available (Safari). Heavy graphics work: The tag allows bitmap graphics manipulation. Integration with other systems: WebSocket allows full-duplex, persistent connections. 10

12. Background: Flash Flash is a vector animation software, originally designed to create animations for display on web pages. Manipulates vector and raster graphics to provide animation of text, drawings, and still images Flash is everywhere. A gigantic community and secondary market.. 11

13. Background: GIUMark2 12 Vector Charting Test Simulating a streaming stock chart with different alpha fills Bitmap Gaming Test Simulates a tower defense game. Uses lots of bitmap resources and animations for sprites on screen at once. Text Column Test Tests rendering capability of various text organizations and character types.

14. Background: Instrumentation 13 UtilityDescription sar(1)General-purpose System Activity Reporter providing numerous system statistics vmstat(1M)Reports virtual memory statistics and aggregate system wide CPU utilization mpstat(1M)Per-CPU statistics netstat(1M)Network statistics kstat(1M)All available kernel statistics prstat(1M)Process/thread statistics

15. Experimental Setup 14 Sun Ray Server 5.2 running on Solaris 11, 2.8Ghz, 3GB RAM Apple Airport Express Router Sun Ray Thin Client 2 and a 2FS 1-2 X-Windows Remote clients for comparison purposes

16. PERFORMANCE ANALYSIS AND RESULTS 15

17. Performance Analysis and Results Evaluations based on GUIMark 2 frame rates for Adobe Flash 10.10 and HTML 5 using Firefox 7 Conducted five tests per scenario for each application on a given platform to normalize rates Assessed: Native Server Sun Ray Remote X Scalability case 16

18. The Native Case… Results based on our Solaris 11 Server Performed for a base case assessment for comparison for our Sun Rays and Remote session analysis 17

19. Native vs. Single Sun Ray 18

20. Remote X Session Performance Assessment based on a remote X session using Firefox Performed frame rate measurements of the remote session, as well as tests on the notebook natively 19

21. Native Laptop vs. Remote X 20

22. Remote X vs. Sun Ray 21

23. Multiple Sun Ray Sessions 22

24. D-Trace and Solaris System Tools Used to determine potential bottlenecks and drawbacks of each respective application Performed on server during Sun Ray Session Performed analysis on three main components: CPU Memory Network I/O 23

25. Flash Breakdown Flash utilized the most CPU time of the system Performed “yield” system call most frequently Lots of memory maps Most page faults on the system Xnewt, the Sun Ray X server, utilized most system calls Highest kernel function used were writes Utilized socket layer the most, higher than the scheduler 24

26. HTML 5 Breakdown Scheduler top process using CPU time, followed by Xnewt and Firefox Reads, writes, and mwaits were most common system calls Reads and writes generated evenly between Xnewt and Firefox Dtrace itself had most page faults, others unranked Similar to Flash, Xnewt largest process hitting network socket 25

27. Remote X Comparisons Remote X doesn’t use Xnewt Increases in performance where Xnewt hogged resources More CPU cycle utilization of sched Number of network reads/writes more even on X Xnewt shows as an additional process in HTML 5 Suffers in cases where HTML 5 shows CPU utilization between sched/Firefox as nearly equal 26

28. Bringing it All Together… Sun Rays perform relatively well in web-based multimedia Xnewt creates bottleneck Work best with HTML 5 in terms of scalability Remote X best on powerful client in HTML 5 Utilizes some of client resources due to X-server rendering Does not handle Flash well due to application dependencies Not targeted toward our intended environments Flash handles well on single Sun Rays, poor otherwise 27

29. In Conclusion… Assessed two different types of thin clients Analyzed various scenarios of Adobe Flash and HTML 5 rendering performance for different types of hardware organizations Assessed the application impact on the server under the hood Effectively determine what created bottlenecks Determine a best organization for such a framework 28

30. Further Areas of Exploration Better Remote X performance assessments How it handles multimedia rendering due to X-servers existing on the client Chromebook Assessments Browser focused OS Presents users only a login to their Google Account, and the web Costs of around $350+ How does this benefit if full netbooks cost less? Can have better performance, but pay more for hardware with less functionality at the moment 29

31. Questions ? 30