HSA System Virtualization and GPU Scheduling

• Main Authors: Che-Wei Chang, 張哲瑋
• Other Authors: Shih-Hao Hung
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/83033771144861607197

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 102 === Heterogeneous computing has been proposed to incorporate specialized processing capabilities (e.g. GPU, DSP and FPGA) in order to handle particular tasks. However, there are some drawbacks of current heterogeneous computing. In order to improve the performance, HSA foundation proposed the Heterogeneous System Architecture (HSA). In this thesis work, we investigate on the techniques for virtualizing an HSA platform to support various types of usages with virtual machines. We first delivered a system that emulates system virtualization of HSA. Second, we proposed some GPU scheduling policies of HSA to utilize the GPU. Third, we implement the GPU context switch and apply to scheduling policies. Finally, We designed a simple timing model to evaluate GPU scheduling policies. Experimental results show that GPU scheduling polices with context switch are important, and better policies increase the performance of GPU for HSA virtualizaion.