parca
parca copied to clipboard
NVIDIA GPU profiling support
Feature Request
Support for GPU profiling similar to what's currently being offered for CPU.
Alternative solutions
- Sentry's (OSS) OpenTelemetry Collector collects GPU metrics
- Framework-specific like PyTorch Profiler or TensorFlow Profiler that could be incorporated into this product to provide it as a service out of the box.
Is there a use case or business reason for this request?
The CPU market is growing at a compound annual growth rate (CAGR) of 4.36% while the GPU market grows at a CAGR of 33.4%. Also NVIDIA has the biggest market share at 80%.
I love this idea! No clue if we can use eBPF for this, but pprof is a generic format, so as long as we can get it into that format we can make it work!