A Method for Estimating Task Granularity for Automating GPU Cycle Sharing

In this paper, we propose a method for estimating task granularity to reduce guest's effort in graphics processing unit (GPU) cycle sharing systems. A cycle sharing system should maximize both the frame rate for resource donators (i.e., hosts) and the acceleration effect for scientific job submitters (i.e., guests). This can be realized by selecting the appropriate granularity for guest tasks. However, a previous cooperative multitasking method requires manual interactions to find the appropriate task granularity. To avoid such interactions, the proposed method automatically estimates the appropriate granularity by measuring the length of idle periods on the host. We show how this measurement can be realized on donated resources, where the host program cannot be instrumented due to the lack of code availability. In experiments, the proposed method automatically estimated the appropriate task granularity, which not only maintained the frame rate for an image filter (i.e., host task) but also maximized the performance of matrix multiplication (i.e., guest task).