This course aims to enable OS research by uncovering interactions within black-box GPU stacks. The learning outcomes include understanding the methodology to reveal these interactions and creating a state machine to capture them. The course teaches skills such as analyzing GPU interactions, interpreting application performance, and enabling OS kernels to control GPU resource management. The teaching method involves systematizing a methodology and presenting a case study. The intended audience for this course includes researchers interested in GPU resource management and OS kernel control.
Overview
Syllabus
Intro
The GPU software/hardware stack Application
Motivation
Quick how-to
From traces to state machine
State machine reduction
Selecting
Case study: Nvidia NVS295, k = 35
The GPU driver state machine distilled
Conclusions, future work
Taught by
USENIX