Fupermod: A framework for optimal data partitioning for parallel scientific applications on dedicated heterogeneous HPC platforms

Abstract

Optimisation of data-parallel scientific applications for modern HPC platforms is challenging in terms of efficient use of heterogeneous hardware and software. It requires partitioning the computations in proportion to the speeds of computing devices. Implementation of data partitioning algorithms based on computation performance models is not trivial. It requires accurate and efficient benchmarking of devices, which may share the same resources but execute different codes, appropriate interpolation methods to predict performance, and mathematical methods to solve the data partitioning problem. In this paper, we present a software framework that addresses these issues and automates the main steps of data partitioning. We demonstrate how it can be used to optimise data-parallel applications for modern heterogeneous HPC platforms. © 2013 Springer-Verlag Berlin Heidelberg.