Evaluating the Impact of GPU Partitioning on Kernel Execution Time

Abstract

GPU acceleration has become a critical strategy to satisfy the increasing requirements for computational capacity in real-time high-performance embedded systems. In this paper, we investigate the problem of providing accurate timing estimations for GPU-accelerated functions in real-time systems, in a context where a GPU device is concurrently being used by multiple applications. For this purpose, we experiment with the available possibilities to partition the resources on an Nvidia GPU to allow multiple concurrent uses, leveraging the libsmctrl library for hardware compute partitioning. We deployed a variety of compute kernels using different partition sizes, possibly along with other kernels that concurrently run in other partitions, and measured the resulting kernel execution times. We show an effective yet simple rule to predict the change in execution time of compute-bound kernels when using partitioning, and measure its accuracy against the obtained experimental data. Overall, the results presented provide new insight into the opportunities of using GPU partitioning for real-time systems.