Home // ENERGY 2018, The Eighth International Conference on Smart Grids, Green Communications and IT Energy-aware Technologies // View article

Performance Prediction of Geophysics Numerical Kernels on Accelerator Architectures

Authors:
Vı́ctor Martı́nez
Matheus da Silva Serpa
Philippe Olivier Alexandre Navaux
Edson Luiz Padoin
Jairo Panetta

Keywords: Machine Learning; Geophysics Applications; Many- core Systems; Performance Model

Abstract:
In order to develop geophysics tools for exploration of energetic resources, numerical models are proposed to understand complex geological structures. They are solved from the discretization of Partial Differential Equations by the Finite Differences Method. This method creates a pattern that solves each point in a 3D domain, and it replicates the same calculation to compute all the data domain. Because of the quantity of calculations, solving the numerical kernels requires High Performance Computing. However, the complexity of current architectures may reduce the efficiency. In some cases, applications tuning have been used to improve the performance. In this context, predicting the performance from input parameters is a critical problem. This is particularly true regarding the high number of parameters to be tuned both at the hardware and the software levels (architectural features, compiler flags, memory policies, multithreading strategies). This work focuses on the use of Machine Learning to predict the performance of geophysics numerical kernels on manycore architectures. Measures of hardware counters on a limited number of executions are used to build our predictive model. We have considered three different kernels (7-point Jacobi, seismic and acoustic wave propagation) to demonstrate the effectiveness of our approach. Results show that the performance can be predicted with high accuracy.

Pages: 76 to 81

Copyright: Copyright (c) IARIA, 2018

Publication date: May 20, 2018

Published in: conference

ISSN: 2308-412X

ISBN: 978-1-61208-635-4

Location: Nice, France

Dates: from May 20, 2018 to May 24, 2018