Reproduce “CUTHERMO: Understanding GPU Memory Inefficiencies with Heat Map Profiling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yanbo Zhao, Jinku Cui, Zecheng Li, Shuyin Jiao, Xu Liu, Jiajia Li
CUTHERMO, developed at North Carolina State University, introduces a real-time profiling tool that visualizes fine-grained GPU memory access patterns using a "distinct warp count" metric and "word-sector" granularity. This enables the identification of critical inefficiencies like false sharing and strided access, leading to demonstrated performance improvements of up to 721.79% in various applications.
Key insights
- The "distinct warp count" metric, combined with "word-sector" granularity, effectively exposes micro-architectural memory inefficiencies like false sharing and strided access that are not readily apparent with traditional profiling methods.
- Five specific inefficient memory access patterns (hot spots, shared memory abuse, false sharing, misalignment, and strided access) were identified as portable and consistently observable across different NVIDIA GPU architectures.
- Visualizing these fine-grained memory access patterns through an interactive heat map offers an intuitive way for developers to pinpoint bottlenecks and derive concrete, actionable optimization strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.