Featherlight Speculative Task Parallelism

Abstract

Speculative task parallelism is a widely used technique for solving search based irregular computations such as graph algorithms. Here, tasks are created speculatively to traverse different search spaces in parallel. Only a few of these tasks succeed in finding the solution, after which the remaining tasks are canceled. For ensuring timely cancellation of tasks, existing frameworks either require programmer introduced cancellation checks inside every method in the call chain, thereby hurting the productivity, or provide limited parallel performance. In this paper we propose Featherlight, a new programming model for speculative task parallelism that satisfies the serial elision property and doesn’t require any task cancellation checks. We show that Featherlight improves productivity through a classroom-based study. Further, to support Featherlight, we present the design and implementation of a task cancellation technique that exploits runtime mechanisms already available within managed runtimes and achieves a geometric mean speedup of 1.6 over the popular Java ForkJoin framework on a 20 core machine.