Scheduling Parallel Futures By Work Stealing (2008)
BibTeX
@MISC{n.n.08schedulingparallel,
author = {n.n.},
title = {Scheduling Parallel Futures By Work Stealing},
year = {2008}
}
OpenURL
Abstract
The choice of scheduling policy can have dramatic effects on the performance of implicitly-threaded parallel programs. Work stealing has been shown to be an efficient method for scheduling fine-grained tasks in programs that use fork-join parallelism, but there has been little work studying this form of scheduling in languages that support parallel pipelining. Pipelining allows programmers to express a wider range of programs. We study futures as a means to implement pipelining. While there is an established implementation of work stealing for fork-join parallelism based on double-ended queues, there is no established implementation that supports parallel futures or other pipelining constructs. We identify a number of important design choices in any such implementation. To understand the effects of these choices, we have implemented several of these schedulers for a deterministic parallel language with futures. We evaluate these implementations by running a suite of benchmarks that use both a combination of fork-join parallelism and futures. We compare them by measuring not only steals and space use but also deviations, a generalization of steals.







