KubeCon + CloudNativeCon North America 2024

How much node resource management can affect AI workload performance? What options are there? What is the trade-off between total throughput and low latencies? In this talk we take a systematic approach to Platform Performance Optimization. We walk through the whole path from goal setting, gathering data, analysis, visualizations and conclusions. At each stop along the path we share our practical experiences in a case of LLM inference optimization. You will find many considerations, findings and practical tricks to take away. For instance, how to instrument PyTorch without touching the source or a container image, how to enable changing what we are measuring without new expensive benchmark reruns, and how much more we can learn from visualizations compared to numeric averages and percentiles. Finally we share real results from our case: how resource management increased total token throughput per worker node by more than 3.5x from the baseline.