Predicting performance and energy efficiency for large-scale parallel applications on highly heterogeneous platforms

Predicting the performance of parallel programs for large-scale parallel platforms is challenging due to the disparity between development system and target platform. This is even worse now that energy efficiency is a universal concern and platforms move towards highly heterogeneous reconfigurable systems containing GPUs, FPGAs, and other unconventional processing elements. In this paper we present a simulative approach that predicts energy usage and performance of parallel software on large heterogeneous platforms. It simulates communication activity in detail while abstracting functional behaviour. This allows developers to quickly compare and optimise application designs, hardware configurations, and mapping alternatives even without a fully working target platform. Introduction