Fingerprinting GPU Workloads with eBPF

Learn how to bridge the observability gap between host kernels and GPU workloads using eBPF technology in this 20-minute conference talk from the Linux Plumbers Conference. Discover a novel approach that leverages eBPF tracepoints attached to NVIDIA kernel drivers to non-intrusively monitor and profile GPU activity at the kernel level. Explore how this method captures rich metrics including frequency and timing of driver function calls to create unique "execution fingerprints" for different workloads. Examine practical demonstrations showing how eBPF-based profiling can reliably identify various real-world GPU tasks including machine learning training, inference, and cryptocurrency mining with high accuracy. Understand the limitations of current observability solutions that rely on application-level telemetry or proprietary vendor tools, and see how this kernel-level approach provides a more comprehensive view of GPU resource utilization for system administrators, security engineers, and resource schedulers.