high performance computing on graphics processing units: hgpu.org

Due to their high computing performance, CPU-GPU heterogeneous computing platforms are widely used in mobile devices such as smart phones, tablet computers, and unmanned aerial vehicles. Because a mobile device is often powered by a battery, how to elegantly design a power-efficient real-time computing system becomes an important problem. In this paper, we propose a power-efficient scheduling approach to fully make use of the computation resource in a system with a CPU-GPU computing architecture to run its different functionalities using taskflow programming. The computation resource is modeled as heterogeneous computation units, and different functionalities are modeled as computing tasks whose runtime behavior with computation units can be profiled in advance. The scheduling problem is thus formulated as an integer nonlinear programming problem with the goal of minimizing power consumption. We derive a state-transition equation for such a problem and develop a dynamic programming approach to assign these computing tasks to different computation units and allocate the computation resource to them. Experimental results show that, compared to other approaches like particle swarm optimization, our proposed approach needs a shorter time to find a more power-efficient schedule. A case study is carried out on a physical platform, and the results confirm the effectiveness of our designed schedule.