Latency reduction from runtime-interference to the parallel Quantum Chemistry program GREMLIN in heterogeneous and homogeneous environments

摘要

The runtime behaviour of the PVM version of the Quantum Chemistry program GREMLIN is monitored and analyzed with respect to the latency of the different participating hosts. A priori parallel performance is far from optimal, because of the multifold variabilities influencing the heterogeneous cluster environment. The article discusses the improvement gained when periodically re-partitioning the load based on the current actual latencies of the parallel nodes. It is also shown that even on homogeneous parallel architectures this dynamic version of Speed Weighted Load Balancing (SWLB) may aid to optimize parallel performance considerably.