A Study of Evidence-Based Performance Tuning

2017 Fiscal Year Final Research Report

• Project/Area Number: 26540031
• Research Category: Grant-in-Aid for Challenging Exploratory Research
• Allocation Type: Multi-year Fund
• Research Field: Software
• Research Institution: Chiba Institute of Technology (2017); Institute of Physical and Chemical Research (2014-2016)
• Principal Investigator: HASHIMOTO Masatomo 千葉工業大学, 人工知能・ソフトウェア技術研究センター, 上席研究員 (60357770)
• Project Period (FY): 2014-04-01 – 2018-03-31
• Keywords: 高性能計算 / 性能チューニング / 最適化パターン推定 / 計算カーネル同定 / 目標実行効率推定 / プログラム構造理解支援 / プログラム解析 / 機械学習

Outline of Final Research Achievements

Performance tuning is still a demanding manual task. To improve an application's efficiency, we have to identify its computational kernels, each of which is typically composed of one or more loops. Then various empirical attempts such as loop transformations are made. Thus, it is crucial to learn from the experience of performance tuning experts. As a proof-of-concept, we extracted various facts from performance tuning histories of a few real-world scientific applications, and then constructed a database of that facts. Based on the database, we constructed a few experimental predictive models for promising loop transformation patterns. We also explored a thousand computation-intensive applications to reveal the distribution of kernel classes, each of which is related to expected efficiency and specific tuning patterns. In addition, we constructed a binary classifier for identifying loop kernels and a multi-class classifier for predicting kernel classes.

Free Research Field

情報学