A Case for Remote GPUs over 10GbE Network for VR Applications

VR technology that enables users to experience environments made by computer similar to real environments has become popular. In VR technology, computation cost of graphic processing is high and thus it requires a high-end GPU because high-quality pictures that look like real environments are processed while reflecting sensor information. Therefore, users have to prepare a computer with a high-end GPU, which requires a high cost. In this paper, we propose to connect GPU cluster and user-side computers via 10GbE (10Gbit Ethernet) network so that graphic processing for HMDs is done by the GPU cluster via a network. In this way, users can use HMD with inexpensive computers because they do not have to prepare computers with high-end GPUs. In this paper, we propose an index to evaluate performance of VR processing tasks in remote GPU environment and evaluate the performance in the remote GPU environment based on this index. We also propose a condition that does not degrade user experience in the remote GPU environment and allocation methods of multiple tasks to GPUs under this condition. The evaluation results of the proposed allocation methods show that if there is a high-load task, the method that preferentially allocates tasks to GPUs with low bandwidth achieves high performance; otherwise the method that preferentially allocates tasks to GPUs with high utilization achieves a high performance.