A dynamic special-purpose scheduler for concurrent kernels on GPU

Abstract

GPUs are widely used as powerful accelerators for data-parallel applications such as financial and scientific applications in industrial and scientific areas. Effective scheduling of kernels can significantly enhance performance and utilization. In shared environments such as cloud, lots of kernels from users are being requested to be launched for execution. An effective kernel scheduling method can improve performance. In special environments such as space agency in which special tasks are processing separate fixed-size input data, special-purpose scheduling methods can be effective. In this paper, a dynamic special-purpose scheduler is proposed for scheduling specific tasks that are processing different fixed-size input data. Previous works mostly are static and can't schedule kernels that are launched in runtime. Experimental results show up to 25 percent improvement in execution time in the best case and 15 percent in average on NVIDIA GTX760.