Optimization for Hybrid MPI-OpenMP Programs on a Cluster of SMP PCs

This paper applies a Hybrid MPI-OpenMP programming model with a thread-to-thread communication method on a cluster of Dual Intel Xeon Processor SMPs connected by a Gigabit Ethernet network. The experiments include the well-known HPL and CG benchmarks. We also describe optimization techniques to get a high cache hit ratio with the given architecture. As a result, the hybrid model shows performance prominence over the pure MPI model with about 27% for CG and 12% for HPL. Besides, with a relatively small programming effort, we have succeeded in reducing the cache miss ratio and thus significantly risen up performance for the CG benchmark as much as 4.5 times in some cases.