Hybrid parallelization of discrete element software for heterogeneous resources

Abstract

This paper presents the hybrid parallelization of DEM software developed by using OpenCL for shared-memory architectures and MPI for distributed-memory heterogeneous resources. Resource-aware partitioning based on the weighted RCB method adapts computational workload to heterogeneous Docker containers of the OpenStack cloud. The execution time of benchmark on 7 heterogeneous containers, including the container equipped by GPU, is reduced up to 32.6% of the execution time obtained by using unweighted repartitioning. The speedup of parallel computations up to 6.0 is measured on 8 heterogeneous containers. The replacement of 3 faster containers by 3 slower ones slightly decreases the speedup up to 7.4% of the speedup measured on 5 homogeneous containers.