Enabling OS Research by Inferring Interactions in the Black-Box GPU Stack

Explore a groundbreaking methodology for uncovering interactions within black-box GPU stacks in this 18-minute conference talk from USENIX ATC '13. Delve into the research conducted by Konstantinos Menychtas, Kai Shen, and Michael L. Scott from the University of Rochester, which aims to enable OS research by inferring interactions in the opaque GPU environment. Learn how their approach produces a state machine capturing transitions among semantically meaningful states, providing valuable insights for understanding and optimizing application performance. Discover how this methodology allows OS kernels to intercept and manage GPU requests, potentially revolutionizing whole-system resource management. While not yet production-ready, understand how these tools open new avenues for researchers to explore GPU resource management outside of vendor laboratories.