Abstract: The performance of collective communications impacts the efficiency of large scale parallel numerical computing application. Since the theoretic analysis and evaluation on every type of collective communications is still insufficient for the programmer, the communication modules of many applications are designed and used unreasonably. Using Intel IMB benchmark, this paper analyzes and evaluates the typical MPI (message passing interface) collective communication on BXJ supercomputer (a parallel machine platform), and gives the theoretic results based on the current model and algorithm. The results show that the performance of different MPI collective communications is different. There is an obvious gap between the actual value of some collective communication and theoretic value. The results reflect that there are still many researches, such as accurate theoretic performance model for some collective communication, optimization of collective communication, and optimization of application communication module, etc.

Key words: collective communication, communication performance, scalability

摘要： 集合通信性能是影响并行程序并行效率的重要因素之一，但对于大规模并行计算机上不同类别集合通信的评测和理论分析仍较为缺乏，许多应用程序的通信模块设计和使用不合理。基于某国产并行机平台，利用IMB测试程序，对各典型MPI（message passing interface）集合通信性能进行了分析，并基于现有通信模型和算法进行理论拟合。结果显示：不同类别的MPI集合通信操作的性能差异很大，并且许多集合通信的性能在超大规模下与理论差距很大。一方面反映出现有理论和模型的不足；另一方面也体现出，无论是集合通信的优化，还是基于集合通信的特征进行应用程序的通信模块设计，仍然大有可为。

关键词: 集合通信, 通信性能, 可扩展性