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.<
>