Simulation of collective operations hardware support for «Angara» interconnect

Abstract

JSC NICEVT develops the Angara high-speed interconnect with multi-dimensional torus topology. To evaluate the performance of the interconnect on a large number of nodes a cycleaccurate simulator is used. Angara interconnect supports two types of collective operations: broadcast and reduce. The paper describes the implementation of collective operations in the simulator and presents an early performance evaluation. Performance benchmarks include some basic broadcast and all-reduce tests, as well as several well-known computational applications, specificaly, sparse matrix-vector multiplication and numerical solution of the nonlinear heat conduction equation.