RECURSIVE MULTI-FACTORING ALGORITHM FOR MPI_ALLREDUCE

摘要

This paper shows an effective all-reduction algorithm and its implementation on the Message Passing Interface. It performs comparatively stable in case not only composite numbers of processors but also prime numbers, since we introduce the process detachment strategy on each factorizing stage. On a preliminary test, we examine its efficiency, and we discuss and compare it with the existing algorithms by introducing a performance model of our algorithm.