Speaker
Description
With the increased adoption of AI workloads, efficient GPU checkpointing mechanisms are becoming crucial for inference, training, fine-tuning, and reinforcement learning workloads. One of the key challenges with GPU checkpointing today is the lack of memory-tracking support that enables incremental snapshots. When the GPU state is checkpointed into host memory, all pages appear modified, preventing CRIU from identifying which pages have changed since the previous checkpoint and resulting in full snapshot of the GPU memory for every iteration. In this talk, we will discuss extending the GPU plugins for CRIU with support for memory tracking that enables efficient incremental checkpointing. We will explore the benefits of this approach and the trade-offs between performance overhead and storage efficiency across different GPU workloads.