Contention of Communications in Switched Networks with Applications to Parallel Sorting

Abstract

Contention of communications across a switched network that connects multiple compute nodes in a distributed-memory cluster may seriously degrade performance of parallel code. The InfiniBand network is the most popular interconnect for compute clusters. While one may correctly assume that increased resource contention leads to decreased application performance, alternate methods such as virtual channels and adaptive routing have obscured the point at which inter-job interference becomes a major issue. This contention is maximized when communicating large blocks of data among all parallel processes simultaneously. This communication pattern arises in many important algorithms such as parallel sorting. We use the cluster tara in the UMBC High Performance Computing Facility (HPCF) with a quad-data rate Infini-Band network which provides an opportunity to test the case if the capacity of a switched network is a limiting factor in algorithmic performance.