Platform Performance Optimization for AI - A Resource Management Perspective

Explore a comprehensive conference talk that delves into platform performance optimization for AI workloads from a resource management perspective. Learn how node resource management impacts AI workload performance through a systematic approach that covers goal setting, data collection, analysis, and visualization techniques. Discover practical insights for LLM inference optimization, including methods to instrument PyTorch without source code modifications, techniques for flexible measurement adjustments without costly benchmark reruns, and advanced visualization approaches that provide deeper insights than traditional numeric metrics. Follow along as speakers demonstrate a real-world case study where resource management strategies led to a 3.5x improvement in token throughput per worker node compared to baseline performance. Master the trade-offs between total throughput and latency while gaining actionable techniques for optimizing AI platform performance.