Hardware accelerators (such as the Cell Broadband Engine) have recentlyreceived a significant amount of attention from the computational sciencecommunity because they can provide significant gains in the overall performanceof many numerical simulations at a low cost. However, such accelerators usuallyemploy a rather unfamiliar and specialized programming model that oftenrequires advanced knowledge of their hardware design. In this article, wedemonstrate an alternate and simpler approach towards managing the maincomplexities 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 theperformance results. We obtain gains in the overall performance of genericsimulations that are close to the theoretical maximum that can be obtainedthrough our parallelization approach.
展开▼
