cuCatch: A Debugging Tool for Efficiently Catching Memory Safety Violations in CUDA Applications

Explore a groundbreaking debugging tool for CUDA applications in this 23-minute conference talk from PLDI 2023. Dive into cuCatch, a novel memory safety error detection tool specifically designed for the CUDA programming model. Learn how it combines optimized compiler instrumentation with driver support to implement an innovative algorithm for catching spatial and temporal memory safety errors with low performance overheads. Discover how cuCatch addresses the unique challenges of GPU software development, including multiple memory spaces and thousands of concurrently executing threads. Examine experimental results demonstrating cuCatch's superior error detection coverage and significantly lower runtime slowdown compared to existing debugging tools. Gain insights into how this tool can accelerate memory safety debugging for GPU applications, making it an essential resource for developers working with NVIDIA GPUs and CUDA programming.