Experiments about the Performance of GPU Oversubscription Strategy

GPU is gaining popularity in various applications due to its preeminent parallel computing power. However, GPU cloud is still far from mature, especially in resource supplying mode to guarantee possible highest yield. These deficiencies significantly impose restrictions on the potential markets of GPU. This paper introduces resource oversubscription and conducts experiments to evaluate the performance of oversubscribed GPU based on the cloud-style workloads with different features. The results of our experiments demonstrate that GPU oversubscription performs excellent improvement on economic benefits compared to original GPU supplying mode and a acceptable safety in QoS control. It is necessary and beneficial for datacenters to employ GPU oversubscription technology.