Task mapping on a dragonfly supercomputer

Abstract

The dragonfly network topology has recently gained traction in the design of high performance computing (HPC) systems and has been implemented in large-scale supercomputers. The impact of task mapping, i.e., placement of MPI ranks onto compute cores, on the communication performance of applications on dragonfly networks has not been comprehensively investigated on real large-scale systems. This paper demonstrates that task mapping affects the communication overhead significantly in dragonflies and the magnitude of this effect is sensitive to the application, job size, and the OpenMP settings. Among the three task mapping algorithms we study (in-order, random, and recursive coordinate bisection), selecting a suitable task mapper reduces application communication time by up to 47%.