ACWS: Adaptive Cache-state Aware Warp Scheduling Based on Cache Feature Analysis

Abstract

General-purpose computing on graphics processing units (GPGPU) is widely used in high-performance computing, AI, cloud computing and other fields depending on its powerful parallel computing speed and performance. The on-chip cache in the GPGPU has a significant impact on the overall performance, while the warp scheduling directly affects the utilization of the GPGPU cache, thereby affecting the performance of the GPGPU. We found that for kernels with different workloads, a static single warp scheduling strategy cannot well adapt to multiple types of workloads and provide good performance. Therefore, we propose an adaptive cache-state aware warp scheduler (ACWS) based on cache feature analysis, which can dynamically and adaptively select the appropriate warp scheduling strategy for the current load among different warp schedulers according to the state feedback of L1D. The experiment shows that ACWS can take advantage of other warp schedulers and provide appropriate warp scheduling for cache. It can adapt to various GPGPU workloads with a small amount of overhead. It has been improved to varying degrees on different test benchmarks.