Design a high-performance checkpointing system for deep learning clusters with 10,000+ GPUs. The system must minimize training downtime (stall time), handle petabyte-scale data transfers to persistent storage, and ensure global consistency of model states across a massively parallel environment.