Abstract
Large-scale DL on HPC systems like Frontier and Summit uses distributed node-local caching to address scalability and performance challenges. However, as these systems grow more complex, the risk of node failures increases, and current caching approaches lack fault tolerance, jeopardizing large-scale training jobs. We analyzed six months of SLURM job logs from Frontier and found that over 30% of jobs failed after an average of 75 minutes. To address this, we propose fault-tolerance strategies that recache data lost from failed nodes using a hash ring technique for balanced data recaching in the distributed node-local caching, reducing reliance on the PFS. Our extensive evaluations on Frontier showed that the hash ring-based recaching approach reduced training time by approximately 25% compared to the approach that redirects I/O to the PFS after node failures and demonstrated effective load balancing of training data across nodes.
