学术文献库

Benchmarking is an important challenge in HPC, in particular, to be able to tune the basic blocks of the software environment used by applications. The communication library and distributed run-time environment are among the most critical ones. In particular, many of the routines provided by communication libraries can be adjusted using parameters such as buffer sizes and communication algorithm. As a consequence, being able to measure accurately the time taken by these routines is crucial in order to optimize them and achieve the best performance. For instance, the SKaMPI library was designed to measure the time taken by MPI routines, relying on MPI's two-sided communication model to measure one-sided and two-sided peer-to-peer communication and collective routines. In this paper, we discuss the benchmarking challenges specific to OpenSHMEM's communication model, mainly to avoid inter-call pipelining and overlapping when measuring the time taken by its routines. We extend SKaMPI for OpenSHMEM for this purpose and demonstrate measurement algorithms that address OpenSHMEM's communication model in practice. Scaling experiments are run on the Summit platform to compare different benchmarking approaches on the SKaMPI benchmark operations. These show the advantages of our techniques for more accurate performance characterization.