Strong and Weak Scalability Studies for the 2-D Poisson Equation on the Taki 2018 Cluster

Abstract

The new 2018 nodes in the cluster taki in the UMBC High Performance Computing Facility contain two 18-core Intel Skylake CPUs and 384 GB of memory per node, connected by an EDR (Enhanced Data Rate) InfiniBand interconnect. Parallel performance studies for the memory-bound test problem of the Poisson equation in two spatial dimensions yield several conclusions for the operation of the CPU cluster in taki. Strong scalability studies demonstrate excellent performance when using multiple nodes due to the low latency of the high-performance interconnect and good speedup when using all cores of the multi-core CPUs. Weak scalability studies confirm that best throughput is achieved by using all cores on a shared-memory node. Comparisons to results on the 2009 and 2013 nodes bring out that core-per-core performance of serial code improvements have stalled, but that node-pernode performance of parallel code continues to improve due to the larger number of cores available on a node. These observations compel the recommendations that serial code should use the 2009 and 2013 nodes of taki and parallel code is needed to take full advantage of the high-memory 2018 nodes. Comparisons between several compilers and several implementations of the MPI standard justify the choice of the Intel suite as the default on taki.