Abstract: Alltoall is an important collective operation of MPI (message passing interface), which impacts the parallel efficiency of many parallel numerical computing applications. Since theoretic analysis and evaluation of Alltoall collective communications on the massive parallel computer is still insufficient, improper communication module design and poor communication usage are widespread in many applications. The test of basic MPI communication performance shows that the communication latency is instable due to network contention and the variety increases with the number of MPI processes. In order to reduce the gap between the predicted run-time and the measured run-time, this paper proposes a new performance model to evaluate Alltoall operation. The model considers not only the standard parameters such as bandwidth and latency, but also takes into account network communication contention. Results on the BXJ supercomputer show that the performance prediction model accurately captures the Alltoall communication behavior even for the operations on a large number of processors and manifests network competition cost on the Alltoall communication.

Key words:  collective communication, communication performance, Alltoall

摘要： Alltoall是一种重要的MPI（message passing interface）集合通信类别，是影响许多并行程序并行效率的重要因素。但对于大规模并行计算机上Alltoall集合通信的评测和理论分析仍较为缺乏，导致许多应用程序的通信模块设计和使用不合理。首先，开展了MPI基本通信性能的测试和分析，发现随着MPI进程数的增加，其性能波动也增加，而这种波动源自网络竞争。为此，在传统的Alltoall性能评估模型中引入了网络竞争因素，新模型不仅考虑传统的通信带宽和通信延迟参数，还考虑了通信竞争因素。某国产并行机平台上的测试结果显示：引入网络竞争模型的新Alltoall性能评估模型可以较为准确地预估Alltoall性能，体现出网络竞争开销对Alltoall性能的影响。

关键词: 集合通信, 通信性能, Alltoall