eGPU: Extending eBPF to GPUs for Low-Overhead Dynamic Observability
2025-04-10
With the surge in GPU-accelerated workloads, existing monitoring tools often suffer from high overhead or invasiveness. eGPU innovatively extends eBPF to GPU kernels via runtime PTX injection, enabling low-overhead dynamic observability. By compiling eBPF bytecode into PTX and injecting it into running GPU kernels, eGPU allows for dynamic addition, modification, and removal of instrumentation without interrupting execution. This not only improves the efficiency of GPU performance analysis but also opens up possibilities for programmable GPU computing, runtime optimization, and GPU security.