Memory Consumption of Highly Process-Parallel Jobs in Fugaku

Abstract

We examine the behavior of highly process-parallel jobs based on the MPI parallelization; the jobs were carried out in Fugaku. In particular, we focus our attention on the change of memory consumption of the jobs. When the degree of the process parallelization becomes large, significant memory consumption occurs and increases due to the execution of the MPI initialization. For even larger parallelization degrees, jobs eventually reach memory exhaustion. We clarify that the number of MPI processes that can be accomplished is about 3.5 million at most even when we carry out full-node execution in Fugaku. It is also clarified that users should be careful about a situation that utilizable memory as an explicit description in users' code is quite restricted, although the number of MPI parallelization should be input corresponding to the number of computational nodes used in a huge-scale job in Fugaku.