Extending eBPF to GPU Device and Driver Contexts

on YouTube ↗34m

" summary

Prototypes offload eBPF into GPU device contexts by compiling eBPF to PTX/SPIR-V and adding a verifier, helpers and map support built on bpftime, with driver-side verified attach points for CUDA device function entry/exit, thread begin/end, memory ops and scheduling hooks. Implementation includes gpu_ext and gpu_ext-kernel-modules, roughly 100 lines of instrumentation inside the NVIDIA open driver, and shows runtime wins (about 3–10× faster than NVBit, 4× speedup in some tests and 1.5× on VRM). Current limits: injection uses LD_PRELOAD or library injection before kernel launch and cannot attach into an already-running GPU kernel.

AI-written summary · about

" chapters (8)

1. 0:00Introduction to eBPF on GPU
2. 1:06GPU stack background and challenges
3. 7:11Need for extensible GPU interfaces
4. 8:16Extending eBPF to GPU driver
5. 14:28Scheduling control with eBPF
6. 18:04Device-side eBPF for observability
7. 26:47Implementation and limitations
8. 31:50Q&A on GPU profiling