高性能计算集群运行时环境的配置优化

Runtime Environment Configuration and Optimization on High-performance Computing Clusters

本文关注如何在高性能计算集群上提供良好的运行时环境的问题,以使得并行应用程序获得更高的性能。指出了高性能计算集群运行过程中运行时环境配置优化需要考虑到的内容,包括跨节点资源的分配与选取、节点内进程及线程针对硬件资源的映射与绑定等两大类多方面因素,并分析了它们对并行应用程序性能带来的影响。通过在多个平台对基准程序和应用程序的实际测试来验证运行时环境对并行程序性能的影响,结果表明不同的运行时环境配置能够对应用程序造成约20%的性能差别。最后对运行时环境优化所需要进一步完成的各项具体工作进行了深入的讨论。

In this paper, it is discussed that how to provide good runtime environment for parallel applications so thus to obtain better performance on high-performance clusters. Some aspects within two categories are selected for analyzing their influence to parallel application performance, including inter-node resources allocation and selection and intra-node processes/threads mapping or binding with system resources. Several benchmarks and applications testing results on different platforms are given, which make it the evidence that runtime environment may cause -20% performance variance to parallel applications. Advanced discussion of the detail work need to be further performed is demonstrated at last.