Speculative Evaluation for Parallel Graph Reduction

Speculative evaluation can improve the performance of parallel graph reduction systems through increased parallelism. Although speculation is costly, much of the burden can be absorbed by processors which would otherwise be idle. Despite the overhead required for speculative task management, our prototype implementation achieves 70% eeciency for speculative graph reduction, with little impact on mandatory tasks. Through speculative evaluation, some simple benchmarks exhibit nearly a factor of ve speedup over their conservative counterparts. 1 Motivation Graph reduction is a popular implementation technique for non-strict functional programming languages. Because graph reduction is free of side-eeects, it is well-suited to parallel processing. With conservative evaluation, an expression is not evaluated until its results are actually required. Speculative evaluation increases parallelism by assigning potentially useful computations to idle processors before their results are required. If a speculative computation later becomes necessary, the time required to complete the computation is reduced. On the other hand, if a speculative computation later becomes irrelevant, the time wasted on its evaluation is unimportant, because the only processors that engage in speculative computations are those that would otherwise have been idle. Unfortunately, speculative tasks are diicult to manage 2]. Speculative tasks compete with mandatory tasks for processors. A speculative task which becomes necessary should be upgraded to a mandatory task. A speculative task which becomes irrelevant should be discarded. Shared subgraphs may have to be repaired whenever a speculative task is terminated prematurely. Speculative evaluation may contribute additional hidden costs as well. For instance, even when processing elements are available, the memory and communications demands of speculative tasks may degrade the performance of mandatory tasks through increased contention.