Experimental Study of Network Contention Effects on All-to-All Operation

Interconnection networks of modern highperformance distributed computer systems are now deep hierarchical. In such systems, communication time between processors depends on their replacement in a computer system. In large-scale NUMA/SMP computer clusters network switches form the first level of this hierarchy and the second level is represented by a shared memory of computer nodes. In this paper we present a benchmark for estimating the message passing time when MPI-processes share the interconnection network. We studied the dependence of the execution time of the All-to-all collective operation on the message size and the number of processes sharing the communication channel. Authors developed a software for predicting the completion time of the All-to-all operation depending on the nodes allocation determined by the Resources and Jobs Management System. The software uses the results of an experimental estimate of the performance degradation for the MPI_Send/MPI_Recv operations during simultaneous (concurrent) usage of the communication channel by a set of processes. In future, we will develop structurally oriented algorithms for the determined nodes allocation.