@article{Soga_Yamaguchi_Mathur_Watanabe_Musa_Egawa_Kobayashi_2020, title={Effects of Using a Memory Stalled Core for Handling MPI Communication Overlapping in the SOR Solver on SX-ACE and SX-Aurora TSUBASA}, volume={7}, url={https://superfri.org/index.php/superfri/article/view/318}, DOI={10.14529/jsfi200401}, abstractNote={<p>Modern high-performance computing (HPC) systems consist of a large number of nodes featuring <span style="font-size: 10px;">multi-core processors. Many computational fluid dynamics (CFD) codes utilize a Message </span><span style="font-size: 10px;">Passing Interface (MPI) to exploit the potential of such systems. In general, the MPI communication </span><span style="font-size: 10px;">costs increase as the number of MPI processes increases. In this paper, we discuss performance </span><span style="font-size: 10px;">of the code in which a core is used as a dedicated communication core when the core </span><span style="font-size: 10px;">cannot contribute to the performance improvement due to memory-bandwidth limitations. By </span><span style="font-size: 10px;">using the dedicated communication core, the communication operations are overlapped with computation </span><span style="font-size: 10px;">operations, thus enabling highly efficient computation by exploiting the limited memory </span><span style="font-size: 10px;">bandwidth and idle cores. The performance evaluation shows that this code can hide the MPI </span><span style="font-size: 10px;">communication times of 90% on the supercomputer SX-ACE system and 80% on the supercomputer </span><span style="font-size: 10px;">SX-Aurora TSUBASA system, and the performance of the successive over-relaxation (SOR) </span><span style="font-size: 10px;">method is improved by 32% on SX-ACE and 20% on SX-Aurora TSUBASA.</span></p>}, number={4}, journal={Supercomputing Frontiers and Innovations}, author={Soga, Takashi and Yamaguchi, Kenta and Mathur, Raghunandan and Watanabe, Osamu and Musa, Akihiro and Egawa, Ryusuke and Kobayashi, Hiroaki}, year={2020}, month={Oct.}, pages={4–15} }