Collective communication, which appears frequently in various parallel applications, is one of the most significant causes of scalability degradation in parallel computing, as the time required for it increases with the size of the computers, in some cases even more than linearly. Non-blocking collective communication is expected to be a means to overlap this collective communication with computation and hide the communication time. This ability was adopted in the MPI-3.0 standard approved in 2012 and has since been implemented in most MPI libraries. However, the use of non-blocking collective communication is currently limited to a small number of applications and is not fully exploited. One of the reasons why the use of non-blocking collective communication is not widespread is that users do not know how to apply it on their parallel computers, the effects of using it, or how well it is supported in the first place. In addition, since overlapping communication and computation requires algorithm modification, few programmers are willing to use it if they do not expect to obtain sufficient communication hiding effects. 

Therefore, in this study, we investigate the use of non-blocking collective communication in each resource provided by JHPCN and measure the performance using benchmark programs to clarify the characteristics of the communication hiding effect. In addition, we will create and/or collect several scientific and technical computing algorithms that use non-blocking collective communication and verify whether communication hiding improves their performance. Note that “non-blocking collective communication” is defined broadly in this study. In addition to the non-blocking collective functions, such as MPI_Iallreduce, defined in the traditional MPI standard we also study the impact of newly added features such as persistent collectives, which also enable communication hiding. Further, communication hiding techniques using one-sided communications for irregular group communication are also studied.