Research On Communication Issue On Public High-Performance GPU Cloud Computing Platform

With the rapid development of processor technology, GPU is more and more applied in scientific computing, and gradually becomes one of the mainstreams of high performance computing. CUD A from NVIDIA provides simple and easy used tools for GPU development. Many high-performance computing applications on the GPU platform appear. GPU-based cloud computing platform is not a new concept, and several foreign web service providers have provided GPU hosting and GPU cloud computing services. In China, GPU-based cloud computing platform has been recognized and supported by government, industry and academia. The current existed GPU-based cloud computing services have limits in user configuration, payment, network latency, and thus affect the usage by users and developers. This paper designs and realizes the communication when building high-performance public GPU cloud computing platform, and tests the availability of communication scheme on MeeGo platform.Firstly, this thesis realizes GPU parallel program development and operation for the users without GPU hardware by modifying the original CUDA runtime library routines. Secondly, we use Socket to realize the communication among GPU agents, users and GPU sub-cloud nodes on the public GPU cloud computing platform. This platform allows multiple user access, job submission, GPU execution. Finally, we test the public platform using a terminal device in MeeGo system.This public GPU cloud computing platform can improve the GPU computing performance, save the user investment, and allow users accessing transparently and efficiently. We will continue improving this platform and let GPU resources more conveniently, efficiently and transparently being used by users (especially ones without GPU hardware) in the future.