Detail

StarPU: A Unified Runtime System for Heterogeneous Multicore Architectures

Approaching the theoretical performance of heterogeneous multicore architectures, equipped with specialized accelerators, is a complex issue. While substantial efforts have been devoted to efficiently offload parts of the computations, designing a programming model that unifies all computing units (multicore CPUs, GPGPUs, Cell …) is the main challenge.
We present StarPU, a runtime system featuring a unified execution model and a library that automates data transfers within heterogeneous machines. We show that our approach makes possible to efficiently use all computing resources while hiding most of their complexity. Even though heterogeneity is usually considered as a major concern, we also demonstrate how the use of scheduling strategies makes possible to actually exploit the heterogeneous nature of the various processing units.