Network-aware data mapping on parallel molecular dynamics

Beginning with cluster or grid computing, recent parallel computing environments are involved with more heterogeneity or irregularity. The heterogeneity comes not only from processor performances but also from network topology and communication performances. It is essential to take into account the heterogeneity to derive the maximum benefit from such computing environments. In this research, we show the impact of considering such heterogeneity based on molecular dynamics (MD) with spatial decomposition method. We coordinate with complicated communication patterns caused from spatial decomposition of irregular data by using a newly developed graph mapping method. Previous researches on graph mapping have dealt with network heterogeneity by weighted edge-cut. However, the cost function ignores interference between communications and it is not proper estimation of communication time. Meanwhile, our method considers per-link bandwidths for correct estimation of communication time. In our experiments, the mappings derived by the proposed method reduced the communication completion times up to 20% and up to 60% with relayed communication.