Multithreading: Fundamental Limits, Potential Gains, and Alternatives

Abstract

Multithreading as a means of tolerating latency, enabling powerful parallel languages, and exposing parallelism is critically examined in order to identify its fundamental limits and potential gains. A simple analytical model shows how the performance gain due to multithreading is related to switch cost, remote reference frequency, and outstanding message capacity. Examination of current networks shows that they support only limited multithreading, due to overhead, channel, and volumetric constraints. Compiler-controlled multithreading is proposed as an alternative to hardware multithreading to make effective use of the processor with a limited number of communication threads. The approach is illustrated by a simple parallel language, Split-C, with split-phase remote references and a novel compilation methodology, TAM, for powerful parallel languages which require dynamic scheduling of a large number of threads.