Assessing MPI Performance on QsNet~(II)

摘要

To evaluate the communication capabilities of clusters, we must take into account not only the interconnection network but also the system software. In this paper, we evaluate the communication capabilities of a cluster based on dual-Opteron SMP nodes interconnected with QsNet~(II). In particular, we study the raw network performance, the ability of MPI to overlap computation and communication, and the appropriateness of the local operating systems to support parallel processing. Experimental results show a stable system with a really efficient communication subsystem which is able to deliver 875 MB/s unidirectional bandwidth, 1.6 μsec unidirectional latency, and up to 99.5% CPU availability while communication is in progress.