Kairos for AI: Analysis, Projections, Pathfinding
Antonio Valles, Rebecca David · 2025
Kairos advances the state-of-the-art for analysis of AI/HPC applications and performance projections aimed at future platforms for single and multi-node. Kairos constructs an application model from unique repeating behaviors by applying a novel combination of phase detection and runtime behavior pattern identification. The Figure-of-Merit region is then expressed by these repeating portions, enabling targeted analysis and filtering out what is not important (e.g., startup/warmup/epilogue). Kairos creates multi-level dependency graphs per region across components executing in sequential and parallel execution layers: CPU, GPU, memory transfers, and communication. These graphs, in combination with the application model, are used to project performance and enable detailed insights: critical-path analysis, bottleneck analysis, and concurrency behavior. These insights equip the user to understand the application behavior on various platforms, compare different runs, narrow down performance bottlenecks, and target software/hardware optimizations. This paper will describe the Kairos methodology for analysis and what-if studies.