Scholarlink

Optimizing supercomputer software requires identifying parameter configurations that maximize performance. However, the wide range of parameter values and their varying impact across systems make traditional identification methods insufficient, highlighting the need for new approaches to performance prediction and parameter tuning. We propose Surrogate-Based Modeling (SBM) as an efficient method for characterizing performance across the parameter landscape. Using data from the RAJA Performance Suite’s computational kernels (RAJAPerf), we show that SBM outperforms the standard k-Nearest Neighbors (kNN) model, achieving predictions up to 54% more accurate while requiring 33% less data. Thus, SBM emerges as a powerful tool for enhancing performance predictions across diverse parameter combinations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Surrogate Models for Analyzing Performance Behavior of HPC Applications Using the RAJA Performance Suite

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Befikir Bogale,
  • Ian Lumsden,
  • Dalal Sukkari,
  • Dewi Yokelson,
  • Stephanie Brink,
  • Olga Pearce,
  • Michela Taufer

摘要

Optimizing supercomputer software requires identifying parameter configurations that maximize performance. However, the wide range of parameter values and their varying impact across systems make traditional identification methods insufficient, highlighting the need for new approaches to performance prediction and parameter tuning. We propose Surrogate-Based Modeling (SBM) as an efficient method for characterizing performance across the parameter landscape. Using data from the RAJA Performance Suite’s computational kernels (RAJAPerf), we show that SBM outperforms the standard k-Nearest Neighbors (kNN) model, achieving predictions up to 54% more accurate while requiring 33% less data. Thus, SBM emerges as a powerful tool for enhancing performance predictions across diverse parameter combinations.