high performance computing on graphics processing units: hgpu.org

GPU (General-Purpose computation on Graphics Processing Units) offers an effective computing platform to accelerate a wide class of data-parallel computing. Multi-GPU’s appear as an attractive platform to speed up the computation of data-parallel GPU. This paper aims to explore the feasibility of relaxing the task-level restriction of single GPU use in multi-GPU real-time systems.We develop a multi-GPU runtime support system, called GPU-SPARC, where GPU applications can be automatically split and run concurrently over multi-GPU’s. We present the prototype of GPU-SPARC on OpenCL runtime that can provide the service to existing OpenCL applications without any modification to them unless global synchronization is employed. The multi-GPU parallel computing offers the potential for performance improvement but at the same time incurs additional resource consumption. Thereby, we analysis the benefit and cost of executing a GPU application on multiple GPU’s and propose a GPU execution mode assignment policy from the perspective of system-wide schedulability. Our experiment results show that GPU-SPARC is able to improve schedulability in real-time multi-GPU systems by relaxing the single-GPU-per-task restriction and choosing better GPU execution modes.