Improving the Performance of an AstroPhi Code for Massively Parallel Supercomputers Using Roofline Analysis

Abstract

Astrophysics is the branch of astronomy that employs the principles of physics and chemistry “to ascertain the nature of the heavenly bodies, rather than their positions or motions in space”. Numerical modeling plays a key role in modern astrophysics. It is the main tool for the research of nonlinear processes and provides communication between the theory and observational data. New massive parallel supercomputers provide an opportunity to simulate these kinds of problems in high details. Our astrophysics code AstroPhi was written for new massive parallel supercomputers based Intel Xeon Phi architecture. The original numerical method based on the combination of the Godunov method, operator splitting approach and piecewise-parabolic method on local stencil was used for numerical solution of the hyperbolic equations. The piecewise-parabolic method on local stencil provides the high-precision order. After the transition of AstroPhi to KNL architecture, we obtained abnormally low performance of solver on KNL cores. In this paper, we will show the roofline analysis using Intel Advisor application and the results of the AstroPhi optimizations.