high performance computing on graphics processing units: hgpu.org

In this article, we consider hybrid architectures that consist of standard CPU cores associated with accelerators (such as GPUs). These architectures are increasingly employed in large computing centers. We develop a strategy designed to deal with hybrid computing architectures from the computing performance and programmability points of view. We focus on hybrid computing clusters that consist of a potentially high number of standard CPU cores combined with some accelerators. Although such hardware is increasingly being used, because of related programming difficulties, only a small number of large applications use all the computing resources of such hybrid systems (i.e. both CPU cores and accelerators). To allow large applications to use both CPU cores and accelerators, we introduce SGPU 2, a runtime system that offers a programming model designed to express the fact that a computation can be performed on both CPU cores and accelerators. We first show how SGPU 2 manages the different computing resources of a hybrid architecture and how it can be used with a classical existing seismic wave propagation code; we then use this simulation code to benchmark SGPU 2.