SC17 Denver, CO

P75: Model-Agnostic Influence Analysis for Performance Data


Authors: Rahul Sridhar (University of California, Irvine; Lawrence Livermore National Laboratory), Rushil Anirudh (Lawrence Livermore National Laboratory), Jayaraman J. Thiagarajan (Lawrence Livermore National Laboratory), Nikhil Jain (Lawrence Livermore National Laboratory), Todd Gamblin (Lawrence Livermore National Laboratory)

Abstract: Execution time of an application is affected by several performance parameters, e.g. number of threads, decomposition, etc. Hence, an important problem in high performance computing is to study the influence of these parameters on the performance of an application. Additionally, quantifying the influence of individual parameter configurations (data samples) on performance also aids in identifying sub-domains of interest in high-dimensional parameter spaces. Conventionally, such analysis is performed using a surrogate model, which introduces its own bias that is often non-trivial to undo, leading to inaccurate results. In this work, we propose an entirely data-driven, model-agnostic influence analysis approach based on recent advances in analyzing functions on graphs. We show that the problem of identifying influential parameters (features) and configurations (samples) can be effectively addressed within this framework.
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index