Offloaded MPI persistent collectives using persistent generalized request interface

Abstract

This paper proposes a library with a persistent generalized request interface for the implementation of persistent communication operations. This interface allows developers to add persistent communication functions to the existing MPI library. We implemented a new generalized request interface which supports persistent communications because the generalized requests of the MPI standard lacks the features needed for persistent communications. We evaluate the expressiveness of the interface by developing five implementations of a persistent collective operation, namely, MPI_Neighbor_-alltoall_init: one utilizes the collective offload capability of Fujitsu FX100 Tofu2 interconnect and other four utilize the standard MPI functions and the Fujitsu-extended MPI functions. These implementations are evaluated on FX100 with micro-benchmark programs measuring latency. The results show that the offloaded version outperforms the existing implementations by more than a factor of two with data sizes up to 16 KiB, confirming that the proposed library interface facilitates the development of persistent collectives and the offloaded implementation exhibits the expected performance.