【摘要】 针对分布式多节点多GPU的系统环境,实现一种基于CUDA框架的多GPU通用计算虚拟化平台。应用程序可以如同使用本地GPU一样方便地使用多个远程GPU,原来的CUDA应用程序可以不经过修改或者只进行少量的修改就可以运行在该虚拟化GPU平台上,从而实现单机多GPU和多机多GPU在编程模式上的统一,并通过一个基于高斯混合模型的数据聚类程序来进行实验验证。实验结果表明,在不影响程序正确性的前提下,相对于原来使用CPU的程序,使用两个远程GPU可以获得十倍左右的加速比。更多还原

【Abstract】 Aiming at the distributed multi-node and multi-GPUs system environment, a general computing virtualization platform of multi-GPUs based on CUDA framework is implemented. The application program can use the remote GPUs in the same way as the local GPUs. The original CUDA application program can be run on the virtual platform of GPUs without modification or with only a few changes,in order to achieve unity in the programming model of single multi-GPUs and multi-machine multi-GPUs. In the end,we verify the correctness of the experiment through a Gaussian mixture model for data classification by CUDA. The experiment shows that the result of a program using two remote GPUs can get about ten times faster than using the original CPU without affecting the correctness.更多还原