AN EXPLORATION OF OPENCL ON MULTIPLE HARDWARE PLATFORMS FOR A NUMERICAL RELATIVITY APPLICATION

摘要

Currently there is considerable interest in making use ofrnmany-core processor architectures, such as Nvidia andrnAMD graphics processing units (GPUs) for scientiu0002c computing.rnIn this work we explore the use of the OpenrnComputing Language (OpenCL) for a typical NumericalrnRelativity application: a time-domain Teukolsky equationrnsolver (a linear, hyperbolic, partial differential equationrnsolver using u0002nite-differencing). OpenCL is thernonly vendor-agnostic and multi-platform parallel computingrnframework that has been adopted by all major processorrnvendors. Therefore, it allows us to write portablernsource-code and run it on a wide variety of compute hardwarernand perform meaningful comparisons. The outcomernof our experimentation suggests that it is relatively straightforwardrnto obtain order-of-magnitude gains in overall applicationrnperformance by making use of many-core GPUsrnover multi-core CPUs and this fact is largely independentrnof the speciu0002c hardware architecture and vendor. We alsornobserve that a single high-end GPU can match the performancernof a small-sized, message-passing based CPU cluster.