Research on Mathematical Methods and Development of Libraries for Combined and Hierarchical Autotuning

2017 Fiscal Year Final Research Report

• Project/Area Number: 15H02708
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Research Field: High performance computing
• Research Institution: The University of Tokyo
• Principal Investigator: SUDA Reiji 東京大学, 大学院情報理工学系研究科, 教授 (40251392)
• Co-Investigator(Kenkyū-buntansha): 藤井 昭宏 工学院大学, 情報学部(情報工学部), 准教授 (10383986) ; 美添 一樹 国立研究開発法人理化学研究所, 革新知能統合研究センター, ユニットリーダー (80449115)
• Co-Investigator(Renkei-kenkyūsha): TANAKA Teruo 工学院大学, 情報工学部, 教授 (90622837) ; YAMAMOTO Yusaku 電気通信大学, 情報理工学研究科, 教授 (20362288) ; KATAGIRI Takahiro 名古屋大学, 情報基盤センター, 教授 (40345434)
• Project Period (FY): 2015-04-01 – 2018-03-31
• Keywords: 自動チューニング / 複合的自動チューニング / ベイズ統計 / 実験計画 / 線形モデル / 相関モデル / モデルフィッティング / コード生成

Outline of Final Research Achievements

Autotuning is technology that aims to attain good performance under various conditions, by letting software controls its own parameter. In case multiple parameter exist, most of previous research chose either exhaustive search or heuristic pruning. In this research, we aim mathematically founded method using Bayesian statistics, which gives practically good and asymptotically optimal solutions.
In survey of previous works, we found that linear models and correlation models have such properties, and they can be combined. From description of such models, we create a software that generates a code that constructs performance model from a priori information and observations. Also we apply autotuning mathematical libraries to various computations, and confirms their effectiveness.

Free Research Field

高性能並列数値計算