Workshops Archive

Characterizing Simulation-AI Coupled Workflows on Frontier: Benchmarks and Digital-Twin Replay


Workshop: AI on HPC: Performance Engineering, Challenges and Opportunities

Authors: Ketan Maheshwari (Oak Ridge National Laboratory (ORNL)), Riccardo Balin (Argonne National Laboratory), and Wesley Brewer (Oak Ridge National Laboratory)

Abstract: Simulation–AI workflows now consume a large share of HPC cycles, shifting usage away from traditional modeling and simulation and leaving users and facility policymakers without a clear picture of how these coupled workloads consume system resources. We present a framework that characterizes the HPC resource usage of distributed simulation–AI workflows across six machine learning architectures used for AI training and inference. The framework couples a simulation–AI benchmarking suite with a digital twin of the system, built by extending, customizing, and integrating existing components. We port the benchmarks to the OLCF Frontier supercomputer and run six 64-node experiments, one per architecture, capturing computation, I/O, and energy telemetry without instrumenting any component's source. A translator bridge makes the collected telemetry replayable by the digital twin. We further demonstrate what-if scenario evaluation—energy projections at larger scales and outcomes under alternative scheduling policies—illustrating the framework's broader utility.


Back to AI on HPC: Performance Engineering, Challenges and Opportunities Archive Listing

Back to Full Workshop Archive Listing