Performance analysis of message-passing libraries on high-speed clusters

摘要

This paper presents a performance analysis of message-passing overhead on high-speed clusters. Communication performance is critical for the overall high-speed cluster performance. In order to analyze the communication overhead, a new linear model proposed in this work is used for its characterization. Performance models have been derived using our own micro-benchmark suite on MPI C and emerging Java message-passing libraries. These models predict communication overhead quite accurately. Representative performance metrics have also been obtained in order to evaluate message-passing performance and establish comparisons among different message-passing libraries and clusters. Besides the evaluation process, communication models are useful to optimize parallel applications. Several model-based performance optimizations have been reported. Thus, inefficient primitives have been replaced by more efficient equivalent combinations of primitives. Through a kernel benchmarking it has been analyzed the influence of multiple processor nodes and the message-passing overhead on the overall application performance. From this analysis, it has been concluded that current message-passing implementations do not fully benefit from multiple processor nodes.