Desktop Workload Characterization for CMP/SMT and Implications for Operating System Design Sven Bachthaler Fernando Belli Alexandra Fedorova Simon Fraser University Canada
Objectives Advanced scheduling algorithms for desktop systems? Data collection from live systems
Motivation First study for desktop systems (restricted to Windows XP) Should we address parallelism in periods of activity?
Approach Metric for parallelism Ready queue length Characterization of parallelism Zero parallelism(no threads waiting) Low parallelism (1-2 threads waiting) High parallelism(>2 threads waiting)
Outline Methodology and Data Collection Results Conclusions Future Work
Methodology Collect data from three groups 20 university lab computers 10 university staff computers 12 home computers
Methodology Local and remote data collection Remote data collection For university computers Less overhead No user interaction necessary Local data collection for home PCs
Tools Performance Monitor PsList PsInfo
Data Collection Collected every 15 seconds: Ready queue length Number of running processes Number of running threads Available main memory Percentage of time when CPU is busy
Results Presenting the results Each slide for specific hardware Several computers grouped according to hardware configuration
Results University lab computers…
Results Three groups of lab computers
Results Three lab computers
Results University staff computers…
Results Single staff computer
Results Six staff computers
Results Home computers…
Results Home computers without CMP/SMT
Results Three home computers with CMP/SMT
Results Special case…
Results Staff computer
Conclusion Low parallelism for a significant number of analyzed workloads Not too much benefit from performance-optimizing scheduling algorithms
Future Work Expand data collection to gain statistical significance Investigate better ways for local data collection
Acknowledgements We want to thank the department of Computing Science at SFU Special thanks to the volunteers for the data collection Thank you!