FuPerMod

Computation models

Downloaded 42608 times

Authors:

Alexey Lastovetsky

Authors:

Robert Higgins

Authors:

Vladimir Rychkov

Authors:

David Clarke

Authors:

Zhong Ziming

Authors:

Jean-Noël Quintin

Authors:

Khalid Hasanov

Authors:

Tania Malik

Authors:

Ken O'Brien

Optimization 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. 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 optimize data-parallel applications for modern heterogeneous HPC platforms.