Skip to Main Content
In recent years, Linux-based clusters have become more prevalent as a basis for high performance computing (HPC) systems. Network performance analysis is crucial to the management and administration of such clusters. To assist in this process, we developed Hpcbench to measure UDP, TCP and MPl communications over high performance networks. Hpcbench records and tracks experiment results and system statistics, facilitating detailed analyses of network behaviour. In this paper, we introduce the design and prototype implementation of Hpcbench, and demonstrate Hpcbench in evaluating the network performance of three high performance interconnects in HPC clusters: Gigabit Ethernet, Myrinet, and Quadrics' QsNet.