DL_POLY_3 Parallel I/O Alternatives at Large Processor Counts

Two methods to deal with the I/O bottleneck when performing classical Molecular Dynamics runs on large MPI task counts are presented. We discuss the advantages and drawbacks of both methods, and present performance data for a typical problem size on both HPCx and HECToR. It is found that one method is markedly superior in terms of time to solution, despite an apparently large communication overhead. However, the better performing method use significantly more memory, and the both the implications of this and possible solutions are discussed