A System-Wide Communication to Couple Multiple MPI Programs for Heterogeneous Computing.

This paper proposes a system-wide communication library to couple multiple MPI programs for heterogeneous coupling computing called h3-Open-SYS/WaitIO-Socket (WaitIO-Socket for short). WaitIO-Socket provides an inter-program communication environment among MPI programs and supports different MPI libraries with various interconnects and processor types. We have developed the WaitIO-Socket communication library and tested it on the Wisteria/BDEC-01 supercomputing system, including Odyssey (Fujitsu A64FX-aarch64/Fujitsu-MPI/Tofu) and Aquarius (Intel Xeon-x86_64+NVIDIA-A100/Intel MPI/InfiniBand). As a result of the evaluation, WaitIO-Socket can execute large-scale programs on the Wisteria system, our first target system. The Odyssey and Aquarius MPI programs are able to communicate using WaitIO-Socket and achieve 53.2 GB/s using multiple streams throughout the system. We also show that the application NICAM/ADA is able to run with the h3-Open-UTIL/MP coupler 35% faster on the combination of Odyssey with Arm CPU and Aquarius with NVIDIA GPU than Odyssey with Arm CPU.