Zoomer: Powering AI Performance at Meta’s Scale Through Intelligent Debugging and Optimization

Introduced Zoomer, an automated debugging and optimization platform for Meta's AI training and inference workloads. Zoomer provides deep performance insights to enable energy savings, workflow acceleration, and efficiency gains. Its architecture comprises three layers: an infrastructure layer for scalable trace data storage and processing (using Manifold), an analytics and insights engine with specialized analyzers (integrating Kineto, NVIDIA DCGM, StrobeLight, dyno telemetry) for GPU, CPU, communication, and memory analysis, and a visualization layer for interactive timelines, dashboards, and automated summaries. Zoomer automates profiling triggers, captures comprehensive data streams (GPU metrics, execution traces, host-level data, application annotations, inference-specific data, communication analysis), and performs distributed analysis including straggler detection, bottleneck analysis, critical path analysis, anti-pattern detection, parallelism analysis, memory analysis, and load imbalance analysis. It offers specialized features for LLM workloads (N-dimensional parallelism visualization, GPU efficiency heat maps) and inference optimization (single-click QPS optimization, request-level deep dives, real-time memory profiling).