Local and Global Data Distribution in the Filaments Package

It is generally agreed that using threads and shared memory provides a desirable parallel programming model. However, to achieve scalability it is often necessary to execute these programs on a distributed-memory multicomputer. The Filaments package provides an interface of threads and shared memory with an implementation on a distributed memory machine through a software distributed shared memory (DSM). This focus of this paper is on the problem of finding effective data (thread) distributions both within and between phases, taking data redistribution into account, in scientific applications composed of multiple phases. Our approach, which is implemented within the Adapt run-time data distribution system, takes measurements on the first iteration of the outermost loop in the application, and uses them to find the global distribution (over a reasonable set of distributions) that leads to the best completion time. Intial results are encouraging; for a flame code simulation where behavior is largely dependent on input data, Adapt finds an effective global data distribution with reasonable overhead.