Hybrid MPI/OpenMP optimization in Linpack benchmark on multi-core platforms

With the increasing of the number of CPU cores, thousands of cores are used in the current supercomputers. The MPIJOpenMP hybrid programming model is popular in multicore systems. Some serial codes in the pure MPI programs turn to bottleneck and are easy to be neglected when these codes are ported to the MPIJOpenMP hybrid model. In the Linpack benchmark, we focus on the local swap algorithm and present an OpenMP optimization method to speedup the performance using multi-thread. On a cluster system with 36 multi-core CPUs, experiment results show that this method can decrease the time of lacal swap from 271.11 seconds to 202.53 seconds with the problem size 240, 000 and improve 5.23% of the whole Linpack program from 5.01 TeraFLOPS to 5.28 TeraFLOPS.