Working Set Model for Multithreaded Programs
summary
A statistical pipeline predicts multi-threaded working set size (WSS) using 80 measured points from 20 SPEC and PARSEC programs on SPARC T4/T44 running Solaris, selects RSS and TLB misses per instruction via AIC, and trains models (linear, regression tree, k-nearest neighbor). kNN gives ~96% cross-validated accuracy, 93% on a memcached workload and 88% on SPECjbb, with measurement overhead ~24 microseconds versus Solaris’ clockfs costing up to 2.5 hours to compute WSS.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:53Problem Statement
- 10:44Model Selection
- 11:48Model Evaluation
- 12:46Oral Approach Summary
- 13:44Work in Progress
- 15:10Grouping Threads
- 16:36Conclusion
- 17:10Questions
related talks
38mTask Scheduling for Multicore Embedded Devices
55mInteractive Model-based Resource Analysis for Systems Engineers
44mRestartable Sequences: Scheduler-Aware Scaling of Memory Use on Many-Core Systems
01h 18m“Work Stealing"
39mUnlocking Data-Intensive Applications with Virtual Threads
51mTutorial: Performance Tools
48m“Is Parallel Programming still hard? PART 2 of 2”
15mBWoS: Formally Verified Block-based Work Stealing for Parallel Processing