Cases1 case

1 to 1 of 1

1. IL-0249 Large training jobs on a cluster of over 10,000 GPUs slowed for tens of seconds to nearly ten hours without crashing

   Alibaba Group2024

   Symptoms

   On a shared production cluster of over 10,000 GPUs, large hybrid-parallel training jobs kept running but slowed down, sometimes for hours, without crashing. Among 27 jobs of 512 to 1024 GPUs each submitted in July 2024, 16 slowed this way for 72 minutes on average, and their job completion time grew by 34.59% on average.

   Throughput degradation

   Root cause

   Capacity exhaustion

   Component

   InfiniBand fabricGPU hardware and memory

   Contributing factor

   Overlapping failuresMonitoring blind spot

   Detection

   The authors found fail-slows with small probe training jobs that tracked their own performance at runtime, and by inspecting by hand a one-month trace of large training jobs.

   Remediation

   • Found fail-slows by hand and handled them as fail-stops with checkpoint-and-restart

   Action items

   • Detect fail-slows at runtime from iteration times and locate the slow GPUs and links with Falcon-Detect
   • Mitigate fail-slows with Falcon-Mitigate, escalating from cheap strategies to checkpoint-and-restart only if the slowdown persists

   Lessons

   Fail-slows are usually transient and come from computation, through slow GPUs or CPU contention, or from communication, mostly network congestion. Computation fail-slows are rarer and shorter and cost less, while cross-node communication fail-slows are more common, last longer and slow training more. As training grows, several performance problems are more likely to occur at once, and together they can slow training by more than 90%.

   View case

By root cause1