SeloGPU: A Selective Off-loading framework for High Performance GPGPU execution

Abstract

In general, GPU accelerated GPGPU application results in much higher performance than CPU application. However, to be accelerated by GPU, users should have GPGPU-enabled computation resources like recent GPU or CPU in their local machine. In this paper, we proposed selective GPGPU offloading framework named SeloGPU. SeloGPU not only supports remote offloading for GPGPU application but also supports target node selection among multiple GPGPU-enabled computation resources. We also proposed four optimization techniques to reduce additional overhead owing to remote execution. We implemented SeloGPU using OpenCL which is open standard heterogeneous language. The experimental result shows SeloGPU can choose best target node based on the history of execution information. The four optimization techniques reduce similar to 87% of network transmission overhead.