2023 Fiscal Year Final Research Report
Development of a dynamic load balancing method based on prediction by cooperative use of simulation and machine learning
| Project/Area Number |
20K21787
|
|---|
| Research Category |
Grant-in-Aid for Challenging Research (Exploratory)
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Medium-sized Section 60:Information science, computer engineering, and related fields
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2020-07-30 – 2024-03-31
|
| Keywords | ステンシル計算 / 高性能計算 / 機械学習 / 適合細分化格子法 / 高生産フレームワーク / 動的負荷分散 |
|---|
| Outline of Final Research Achievements |
Recently, adaptive mesh refinement (AMR), which is well suited for GPU computation, has been attracting attention because it can locally refine regions where high accuracy is required. This research aims to develop a method to predict "future" results of simulations by machine learning and to dynamically balance the load of the simulations based on these predictions. We have realized the prediction of fluid simulations by deep learning, the construction of optimal domain decomposition methods based on the amount of computation and communication, the improvement of the AMR method framework, and applied it to the lattice Boltzmann method. We have found machine learning to be helpful in predicting simulations.
|
| Free Research Field |
格子法に基づいた大規模物理計算
|
| Academic Significance and Societal Importance of the Research Achievements |
格子計算はスパコンを利用する代表的なアプリケーションで、局所的に高精細な大規模計算を実現させる意義は大きい。 米国エネルギー省は、AMR法は所謂「エクサスケール」でのマルチスケール問題解決の鍵となる技術と位置付けている。本研究では、機械学習という全く異なるアプローチで数値計算結果を予測する。本研究の目標は予測に基づいた動的負荷分散の実現であるが、近似的ではあるが超高速な予測が可能である機械学習は計算科学分野の様々な要素技術で従来手法を凌駕する可能性を秘めており、本研究でその有用性を示した意義は大きい。
|
