The CBE Hardware Accelerator for Numerical Relativity: A Simple Approach

TL;DR

This paper introduces a simple software caching approach to efficiently utilize hardware accelerators like the Cell processor for numerical relativity simulations, achieving near-maximum performance gains with less complex programming.

Contribution

The paper presents a novel, simplified method called software caching for programming hardware accelerators, demonstrated on a numerical relativity application to improve performance.

Findings

Performance gains close to theoretical maximum

Effective management of hardware complexity

Successful application to Kerr black hole simulations

Abstract

Hardware accelerators (such as the Cell Broadband Engine) have recently received a significant amount of attention from the computational science community because they can provide significant gains in the overall performance of many numerical simulations at a low cost. However, such accelerators usually employ a rather unfamiliar and specialized programming model that often requires advanced knowledge of their hardware design. In this article, we demonstrate an alternate and simpler approach towards managing the main complexities in the programming of the Cell processor, called software caching. We apply this technique to a numerical relativity application: a time-domain, finite-difference Kerr black hole perturbation evolver, and present the performance results. We obtain gains in the overall performance of generic simulations that are close to the theoretical maximum that can be obtained through our parallelization approach.