Super Computing: Creation of Large-Scale and High-Performance Technology for Processes of MCA by Utilizing Supercomputers

Abstract

In this study, we develop the parallelization and high-performance implementation of an application program for multidisciplinary computational anatomy (MCA) with a message passing interface (MPI). We also adapt the program to a supercomputer and evaluate its performance using advanced technologies for code optimization and parallel processing. To establish high-performance computations, we adapt loop collapse techniques for the target loops of a typical MCA program, called large deformation diffeomorphic metric mapping (LDDMM). The loop collapse enables us to obtain a high thread execution using OpenMP. In addition, we adapt an advanced parallel execution style, called hybrid MPI/OpenMP execution, to maximize the execution performance in a supercomputing environment. Hybrid MPI/OpenMP execution enables us to reduce the MPI communication time. This is crucial for massively parallel execution in a current supercomputer environment. The results of our implementation indicate that 18-day sequential processing can be reduced to a 43 h using eight nodes of a Fujitsu PRIMEHPC FX100, which is installed in the Information Technology Center, Nagoya University Japan.