An Efficient Task Partitioning and Scheduling Method for Symmetric Multiple GPU Architecture

This paper presents an efficient task partitioning and scheduling method for symmetric multiple GPU architecture which enables parallel execution of data sending from host to device, kernel execution and data receiving from device to host. Our method divides an individual GPU application into subtasks and overlaps the kernel execution with the bi-directional data transfer from different subtasks on symmetric multiple GPU architecture. To make full use of the bandwidth between the host and the device, we provide a time optimal data transfer algorithm for the data transfer configuration. We divide GPU applications into six basic types. Basing on the classification, we provide six scheduling and modeling sub-methods. We have applied our work to a set of benchmarks and GPUs to measure the performance. The results show that our method successfully hides the communication latency between the host and the device to achieve high performance.