Stability Engineering for Large-Scale Distributed Training: Spike Theory in Autonomous Driving
The article analyzes why performance degrades when scaling single‑machine training to thousands of GPUs, attributing it to the straggler effect, exponential spike probability, and system reliability limits, and presents a three‑layer theoretical framework and concrete engineering practices—including HyperAcc, GPU tracing, NUMA binding, and async DataLoader redesign—to keep per‑node spike rates below 0.5 % and achieve stable, 50 % higher throughput in autonomous‑driving model training.
