Flight Recorder 應對 NCCL Watchdog 超時

🔑 增強重點摘要

• •Flight Recorder operates as a circular buffer within the NCCL communication backend, capturing a rolling window of recent collective operations to minimize memory overhead during normal execution.
• •The tool specifically addresses the 'silent failure' problem in distributed training where NCCL timeouts often mask underlying network congestion, hardware faults, or deadlocks caused by mismatched collective calls across ranks.
• •Integration with PyTorch's distributed package allows for automated dumping of the trace buffer upon detection of a timeout, enabling post-mortem analysis without requiring manual reproduction of non-deterministic hangs.

🛠️ 技術深入

• •Implements a ring buffer mechanism to store metadata for the last N collective operations, where N is configurable to balance memory footprint and diagnostic depth.
• •Captures per-rank state including sequence numbers (SeqNum), operation types (AllReduce, AllGather, ReduceScatter, Broadcast), input/output tensor shapes (numel), and timestamps for latency tracking.
• •Utilizes the NCCL 'Watchdog' thread to trigger the dump, ensuring that the trace is captured at the exact moment the timeout threshold is exceeded.
• •Designed to be low-overhead, utilizing atomic operations for buffer updates to ensure minimal impact on the critical path of high-performance collective communications.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線