Waste-Efficient Work Stealing
Name
3774934.3786452.pdf
Size
806.34 KB
Format
Adobe PDF
Checksum (MD5)
56f1bb14d1ec46991a242f91e62e45cb
Author(s)
Singer, Kyle
Agrawal, Kunal
Schardl, Tao B.
Date Issued
January 28, 2026
Publisher
ACM|Proceedings of the 31st ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming
Citation
Kyle Singer, Kunal Agrawal, and Tao B. Schardl. 2026. Waste-Efficient Work Stealing. In Proceedings of the 31st ACM SIGPLAN Annual Symposium on Principles and Practice of Parallel Programming (PPoPP '26). Association for Computing Machinery, New York, NY, USA, 68–80.
Version
Final published version
Abstract
Although randomized work stealing is effective at automatically load-balancing task-parallel programs, it can waste computational resources when scheduling programs that lack sufficient parallelism to use all available threads. For such programs, threads will waste cycles attempting to steal parallel tasks when none are available. This waste can reduce the machine’s efficiency by wasting computational resources and energy and needlessly burdening the operating system.
This paper introduces WEWS, a simple, practical, and provably efficient extension to randomized work stealing that mitigates waste. WEWS dynamically adjusts the number of active threads to reduce the waste of randomized work stealing. WEWS executes a parallel computation with the same asymptotic running time as traditional randomized work stealing while bounding the waste to O(min{PT∞, T1 + P2}) instructions. WEWS also follows the work-first principle to perform well in practice.
WEWS requires no special support from the operating system or hardware, which simplifies its implementation. We implemented WEWS within the OpenCilk runtime system and compared it to other common waste-mitigation strategies. Across 10 parallel benchmarks, we find that WEWS has minimal impact on parallel running times while, on programs with limited parallelism, substantially reducing waste.
Description
PPoPP ’26, Sydney, NSW, Australia
MIT Department
Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory
Terms of Use
Creative Commons Attribution
Persistent DSpace Link
DOI of Published Version
https://doi.org/10.1145/3774934.3786452
