Profiling the communication workload of an iPSC/860

摘要

Many tools are available for profiling the communication traffic in individual applications on distributed memory message passing machines, but little is known about the communication traffic of the entire workload of a given machine. This work describes the results of using a modified kernel for the Intel iPSC/860 hypercube to collect message passing statistics for each application, run on that machine. The modified kernel does this transparently so that the entire workload can be profiled without requiring special efforts by each user. This kernel was used to profile the communication workload of the 128-node Intel iPSC/860 at the NASA Ames Research Center. Results from this study are presented.