Extending eBPF to GPU Device and Driver Contexts

Explore the groundbreaking extension of eBPF technology to GPU device and driver contexts in this 34-minute conference talk from the Linux Plumbers Conference. Learn how researchers are addressing the observability and customization challenges in GPU workloads by moving beyond traditional CPU-boundary tracing tools that treat GPUs as black boxes. Discover the innovative approach of offloading eBPF programs directly into GPU device contexts through GPU-side attach points for CUDA device functions, thread management, and memory operations, with compilation into device bytecode (PTX/SPIR-V) and full verifier, helper, and map support for on-device execution. Understand how this bpftime-based solution achieves 3-10× performance improvements over existing tools like NVBit while enabling fine-grained instruction-level profiling and programmable scheduling across streaming multiprocessors. Examine the gpu_ext framework that extends Linux GPU drivers with verified eBPF attach points for adaptive runtime customization of GPU scheduling and memory management policies. Gain insights into the technical challenges of implementing eBPF's programming model on SIMT accelerators with complex memory hierarchies and CPU-side orchestration, along with practical use cases for ML workload optimization and the lessons learned from bridging eBPF with GPU architectures.