Low contention load balancing on large-scale multiprocessors

Maurice Herlihy*, Beng Hong Lim, Nir Shavit

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Dynamic load balancing can have a dramatic effect on the performance of parallel programs. A good load balancing algorithm should allow high levels of concurrency without high levels of memory contention. In this paper, we investigate the performance of basic techniques for dynamic load balancing on large-scale multiprocessors. We consider concurrent data structures based on: (1) spin locks with exponential backoff, (2) 'queue' locks, (3) software combining trees, and (4) counting networks. We run a series of simple benchmarks on a simulated 64-processor Alewife machine, a distributed-memory multiprocessor currently under development at MIT. Although the two locking techniques are known to perform well on small-scale, bus-based multiprocessors, we found that they are severely affected by contention as concurrency increases. Both counting networks and combining trees substantially outperform the locking methods. Finally, although combining trees have the same scaling behavior, they are more susceptible to variations in the inter-arrival times of increment requests because they hold locks for long durations.

Original languageEnglish
Title of host publication4th Annual ACM Symposium on Parallel Algorithms and Architectures
PublisherAssociation for Computing Machinery (ACM)
Pages219-227
Number of pages9
ISBN (Print)089791483X, 9780897914833
DOIs
StatePublished - 1992
Externally publishedYes
Event4th Annual ACM Symposium on Parallel Algorithms and Architectures - SPAA '92 - San Diego, CA, USA
Duration: 29 Jun 19921 Jul 1992

Publication series

Name4th Annual ACM Symposium on Parallel Algorithms and Architectures

Conference

Conference4th Annual ACM Symposium on Parallel Algorithms and Architectures - SPAA '92
CitySan Diego, CA, USA
Period29/06/921/07/92

Fingerprint

Dive into the research topics of 'Low contention load balancing on large-scale multiprocessors'. Together they form a unique fingerprint.

Cite this