A scalable locality-aware adaptive work-stealing scheduler for multi-core task parallelism
Year: 2010
Type: dissertation
Abstract: Recent trend has made it clear that the processor makers are committed to the multicore\nchip designs. The number of cores per chip is increasing, while there is little or\nno increase in the clock speed. This parallelism trend poses a significant and urgent\nchallenge on computer software because programs have to be written or transformed\ninto a multi-threaded form to take full advantage of future hardware advances.\nTask parallelism has been identified as one of the prerequisites for software productivity.\nIn task parallelism, programmers focus on decomposing the problem into subcomputations\nthat can run in parallel and leave the compiler and runtime to handle\nthe scheduling details. This separation of concerns between task decomposition and\nscheduling provides productivity to the programmer but poses challenges to the\nruntime scheduler.\nOur thesis is that work-stealing schedulers with adaptive scheduling policies and\nlocality-awareness can provide a scalable and robust runtime foundation for multicore\ntask parallelism. We evaluate our thesis using the new Scalable Locality-aware\nAdaptive Work-stealing (SLAW) runtime scheduler developed for the Habanero-Java\nprogramming language, a task-parallel variant of Java.\nSLAW's adaptive task scheduling is motivated by the study of two common\nscheduling policies in a work-stealing scheduler, specifically, the work-first and the\nhelp-first policy. Both policies exhibit limitations in performance and resource usage\nin different situations. The variances make it hard to determine the best policy a\npriori. SLAW addresses these limitations by supporting both policies simultaneously\nand selecting policies adaptively on a per-task basis at runtime. Our results show\nthat SLAW achieves O.98x to 9.2x speedup over the help-first scheduler and O.97x\nto 4.5x speedup over the work-first scheduler. Further, for large irregular parallel\ncomputations, SLAW supports data sizes and achieves performance that cannot be\ndelivered by the use of any single fixed policy.\nSLAW's locality-aware scheduling framework aims to overcome the cache unfriendliness\nof work-stealing due to randomized stealing. The SLAW scheduler is designed\nfor programming models where locality hints are provided to the runtime by the\nprogrammer or compiler. Our results show that locality-aware scheduling can improve\nperformance by increasing temporal data reuse for iterative data-parallel applications.
Authors: Vivek Sarkar, Yi Guo
Institution: Rice University
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2127811544
