Evaluating QLogic's dispersive routing on high performance clusters

As high performance clusters (HPC) expand in terms of nodes count and processing cores, internet work congestion and bottlenecks at the host and network levels become one of the main concerns in clustered computing. Considerably, the Infiniband Architecture (IBA) Specification provides six different routing algorithms to better optimize the HPC internet work traffic. In this paper, we present these algorithms and then evaluate QLogic's dispersive routing using a large-scale Infiniband cluster, equipped with Intel's latest West mere processor. The paper presents the cluster configuration and evaluates its performance using High Performance LINPACK (HPL) and Intel MPI (IMB) benchmarks. Our results show that whilst the default Min Hop algorithm suits most of the serial and point-to-point benchmarks, the dispersive routing algorithm exhibits improved performance when running specific computational and parallel transfer routines.