Function Shipping in a Scalable Parallel Programming Model

[摘要] Increasingly, a large number of scientific and technical applications exhibit dynamically generated parallelism or irregular data access patterns. These applications pose significant challenges to achieving scalable performance on large scale parallel systems. This thesis explores the advantages of using function shipping as a language level primitive to help simplify writing scalable irregular and dynamic parallel applications. Function shipping provides a mechanism to avoid exposing latency, by enabling users ship data and computation together to a remote worker for execution. In the context of the Coarray Fortran 2.0 Partitioned Global Address Space language, we implement function shipping and the finish synchronization construct, which ensures global completion of a set of shipped function instances. We demonstrate the usability and performance benefits of using function shipping with several benchmarks. Experiments on emerging supercomputers show that function shipping is useful and effective in achieving scalable performance with dynamic and irregular algorithms.

[发布机构] Rice University

[效力级别] sciences