Workshops Archive

Fantastic Hardware Counters and How to Use Them for Noise-Resilient Performance Modeling


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

Authors: Alexander Geiß, Ahmad Tarraf, and Felix Wolf (Technical University of Darmstadt)

Abstract: Performance models are a key instrument in HPC, exposing scalability bugs early and predicting scaling behavior. Most empirical models, however, rely on time measurements, which can be distorted by noise. Hardware counters offer a robust alternative, but not every counter is noise-resilient and reflects scaling behavior. We close the loop between noise-resilience analysis and performance modeling. Building on NORC, we introduce a selection stage combining a knee-based cutoff on noise-resilient counter rankings with a modeling-quality assessment that compares counter models from a benchmark suite against ground truth via lead-exponent deviation, then clusters redundant counters. The selected counters feed two strategies added to Extra-P that derive a model's term structure from the counters and fit only the coefficients from runtime. Evaluated using PolyBench, Kripke, and RELeARN, our approach reduces model error by up to 75% while keeping measurement effort at a standard Extra-P level.


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

Back to Full Workshop Archive Listing