Acceleration of a High Order Accurate Method for Compressible Flows on SDSM Based GPU Clusters

Abstract

The recent advent of multicore processors, and especially the introduction of many-core GPUs, opens new horizons to large-scale, high-resolution, simulations for a broad range of scientific fields. Among them, the scientific area of CFD appears to be one of the candidates that could significantly benefit from the utilization of many-core GPUs. In o rder to investigate such a potential, we evaluate the performance of a high-order accurate method for the simulation of compressible flows. Current implementation is taking place on a GPU cluster. Nevertheless, a novel approach is followed concerning the utilization of GPU clusters that does not involve explicit message passing. Instead, the presented implementation resides on Software Distributed Shared Memory (SDSM) to propagate changes across the simulation phases. The first results prove to be emboldening and lay grounds for further research along the use of shared memory abstraction in order to utilize future GPU clusters.