A Deep Q-Learning Approach for GPU Task Scheduling

Ryan S. Luley, Qinru Qiu

Abstract: Efficient utilization of resources is critical to system performance and effectiveness for high performance computing systems. In a graphics processing unit (GPU) -based system, one method for enabling higher utilization is concurrent kernel execution - allowing multiple independent kernels to simultaneously execute on the GPU. However, resource contention due to the manner in which kernel tasks are scheduled may still lead to suboptimal task performance and utilization. In this work, we present a deep Q-learning approach to identify an ordering for a given set of tasks which achieves near-optimal average task performance and high resource utilization. Our solution outperforms other similar approaches and has additional benefit of being adaptable to dynamic task characteristics or GPU resource configurations.