GPUscout: Locating Data Movement-related Bottlenecks on GPUs

GPUs pose an attractive opportunity for delivering high-performance applications. However, GPU codes are often limited due to memory contention, resulting in overall performance degradation. Since GPU scheduling is transparent to the user, and GPU memory architectures are very complex compared to ones on CPUs, finding such bottlenecks is a very cumbersome process. In this paper, we present a novel method of systematically detecting the root cause of frequent memory performance bottlenecks on NVIDIA GPUs that we call GPUscout. It connects three approaches to analyzing performance - static CUDA SASS code analysis, sampling warp stalls, and kernel performance metrics. Connecting these approaches, GPUscout can identify the problem, locate the code segment where it originates, and assess its importance. This paper illustrates the capabilities and the design of our implementation of GPUscout. We show its applicability based on three commonly-used kernels, yielding promising results in terms of accuracy, efficiency, and usability.