Impact of OpenMP Optimizations for the MGCG Method

This paper shows several optimization techniques in OpenMP and investigates their impact using the MGCG method. MGCG is important for not only an e cient solver but also benchmarking since it includes several essential operations for high-performance computing. We evaluate several optimizing techniques on an SGI Origin 2000 using the SGI MIPSpro compiler and the RWCP Omni OpenMP compiler. In the case of the RWCP Omni OpenMP compiler, the optimization greatly improves performance, whereas for the SGI MIPSpro compiler, it does not a ect very much though the optimized version scales well up to 16 processors with a larger problem. This impact is examined by a light-weight pro ling tool bundled with the Omni compiler. We propose several new directives for further performance and portability of OpenMP.