Analyses on Performance of Gromacs in Hybrid MPI+OpenMP+CUDA Cluster

The Tri-Level parallel programming pattern of MPI+OpenMP+CUDA, which enables better speedup for applications on popular multi-core architecture cluster, is increasingly admired by research institutions and companies. The interaction of particles of molecular dynamics simulation needs extensive calculation, which will also increases with the extension of system. Therefore higher performance for the computing capability and storage ability of current high performance computer is required. As one of main softwares of molecular dynamics simulation, GROMACS can be used for the simulation of hundreds of or even millions of atoms. Based on this kind of hybrid parallel programming pattern, the latest version of GROMACS has higher operation efficiency and shorter simulation time. In this paper, we take advantage of two different sizes of protein simulation system as the data of GROMACS for the molecular simulation of different parallel granularity on mixed cluster which is based on Intel Xeon5650 and NVIDIA C2050. By the result of experiment, we have obtained the best mechanism of hybird CPU-GPU cluster and analyzed the advantage of MPI+OpenMP+CUDA hybrid parallel programming pattern. In the end, we compared GROMACS4.6 with GROMACS4.5. The results of the test also provide a reference for scientists who work on building large-scale molecular dynamics simulation platform and observe the molecular dynamics simulation.