Workshops Archive

An Omnistat-Guided Workflow for Energy-Efficiency Analysis


Workshop: 8th Workshop on Programming and Performance Visualization Tools (ProTools)

Authors: Antigoni Georgiadou (Oak Ridge National Laboratory); Nicholas Frontiere (Argonne National Laboratory); Bronson Messer (Oak Ridge National Laboratory); J.D. Emberson (Argonne National Laboratory); James White III (Oak Ridge National Laboratory); Jordà Polo, Bruno Alvarez, and Karl Schulz (Advanced Micro Devices (AMD)); Woong Shin (Oak Ridge National Laboratory); and Arthur Lorenzon (Federal University of Rio Grande do Sul)

Abstract: Large-scale accelerator applications can exhibit variation in device activity and completion times as workloads evolve. Allocation-wide averages may obscure whether reduced activity reflects improved energy efficiency or devices waiting for slower ranks. We present an Omnistat-guided workflow that aligns timestamped node- and accelerator-level telemetry with application timers and algorithmic steps. Omnistat operates in user mode and requires no modification or recompilation. We demonstrate the workflow using a 64-node CRK-HACC cosmological hydrodynamics simulation spanning 512 GCDs and 625 steps. Its evolving particle distribution illustrates how computational heterogeneity manifests in application timings and system telemetry. Dynamic load balancing is disabled to provide a controlled configuration. The analysis shows how cross-rank timing variability propagates to accelerator utilization, frequency, power, and energy. Relating step energy to cross-device power dispersion identifies intervals for load-distribution analysis and phases that may benefit from power management, making Omnistat a practical first-pass diagnostic layer for subsequent profiling and optimization.


Back to 8th Workshop on Programming and Performance Visualization Tools (ProTools) Archive Listing

Back to Full Workshop Archive Listing