Authors: Yuning Xia and John Mellor-Crummey (Rice University)
Abstract: More than half of the Top 500 supercomputers employ GPUs as accelerators. On GPU-accelerated platforms, developers face a key diagnostic gap: profilers show where stalls occur but provide limited dependency paths explaining why. Furthermore, the same kernel may have different stalls and underlying causes on different GPUs. This paper presents LEO, a root-cause analyzer for NVIDIA, AMD, and Intel GPUs that performs backward slicing from stalled instructions, considering dependencies arising from registers as well as vendor-specific synchronization mechanisms. LEO attributes GPU stalls to source instructions to explain performance inefficiencies. Across 21 kernel/region evaluation units on three GPU platforms, expert-designed, LEO-informed optimizations deliver per-platform geometric-mean speedups of 1.73×–1.82×; these are not end-to-end application speedups. Our case studies show that (1) the same kernel may require different optimizations for different GPU architectures, and (2) under our LLM protocol, prompts containing LEO's structured diagnostics yield better outcomes than code-only and raw-stall-count prompts.
Presentation: file