DaReL: a portable data redistribution library for distributed-memory machines

Run-time data redistribution can be performed between algorithm phases when a different data decomposition is expected to deliver increased performance for a subsequent phase of computation. Redistribution, however, represents increased program overhead as algorithm computation is interrupted while data are exchanged among processor memories. A number of data parallel Fortran languages support run-time data redistribution primitives. We propose a portable data redistribution library, DaReL, specifically for High Performance Fortran (HPF). DaReL supports multi-dimensional data redistribution for HPF´s regular distribution patterns, BLOCK, CYCLIC, and *. Data exchange is performed with MPI primitives, facilitating DaReL´s portability among distributed memory platforms that utilize the emerging message passing standard. We present an overview of DaReL´s design and performance results on an IBM SP-1