Files
Abstract
Distributing data is one of the fundamental problems in implementing efficient distributed-memory parallel programs. The problem becomes more difficult in environments where the participating nodes (processors) are not dedicated to a parallel application. Such environments increase the difficulty of the data distribution problem, which is to determine an assignment of data elements to each node to minimize completion time. We are investigating this problem in the context of explicit message-passing programs.|We have designed and implemented an extension to the popular Message Passing Interface (MPI) that efficiently supports adaptive programs by providing the necessary infrastructure to redistribute data dynamically. (1) an efficient memory allocation mechanism, (2) techniques for accurately determining systems load and computation time, and (3) a heuristic for determining efficient data distributions, including the removal of nodes whose participation degrades the performance of an application. Performance results show that programs that use A-MPI can produce significant improvements over previous load-balancing systems.