Description
Explore checkpoint and restore support for NVCF GPU functions as a way to reduce cold-start latency and improve recovery. The proof of concept should determine whether an initialized function can be checkpointed after expensive startup work and restored on a compatible GPU node to serve traffic faster than a normal cold start.
Definition of Done
- A representative GPU function is checkpointed after model or runtime initialization.
- A restored function instance serves at least one validated request successfully.
- Cold-start and restore latency are measured across multiple runs.
- Constraints involving CUDA, drivers, GPU type, container runtime, process model, storage, security, scheduling, and observability are documented.
- The proof of concept ends with a recommendation to proceed, pause, or reject a product implementation.
Description
Explore checkpoint and restore support for NVCF GPU functions as a way to reduce cold-start latency and improve recovery. The proof of concept should determine whether an initialized function can be checkpointed after expensive startup work and restored on a compatible GPU node to serve traffic faster than a normal cold start.
Definition of Done