Multi-GPU Parallel Computing and Task Scheduling under Virtualization

Abstract

General Purpose Graphics Units (GPGPUS) have seen a tremendous rise in scientific computing application. To fully utilize the powerful parallel computing ability of GPU， and combine the isolation characteristic of virtualization, a GPU virtualization method that supports dynamic scheduling and multi-user concurrency is proposed. For multi-task of GPU general computing programs in virtualization environment, the existing GPU scheduling algorithms have been improved for achieving a more fine-grained and more accurate load evaluation. For large-scale computing programs, we present a method for multi-GPU collaborative computing in virtualization environment, which can effectively deals with accelerating the large-scale program on multi-GPU within a single node. In the experiments, we make verifications by using the representative scientific computing examples, such as classical matrix calculation and discrete Fourier transformation. The experimental results prove that with the increasing of the calculation scale, the speedup can go up and finally close to the numbers of GPU.