理化学研究所 計算科学研究センター

メニュー
メニュー
Events/Documents イベント・広報

R-CCS Cafe

R-CCS Cafe は、異分野融合のための足掛かりとして、計算科学研究センター(R-CCS)に集う研究者が井戸端会議的にざっくばらんに議論する場として、毎月2回程度予定しております。興味をお持ちの方は原則どなたでも参加可能です。

  • 目 的: 異分野間の壁を超えた研究協力を促進し、新しい学問分野の開拓を目指すため、 研究者間の情報交換・相互理解の場を提供し、研究協力のきっかけを作る。
  • 会 場:R-CCS 6階講堂(予定)
  • 言 語:講演は日本語/英語、スライドは英語
  • その他:講演者は他分野の方にも理解できる発表を心掛け、参加者は積極的に質問しましょう。

「AICS Cafe」は2018年度より「R-CCS Cafe」に名称が変わりました。


第145回
日時: 2018年8月29日(水)、15:00 – 16:20
場所: R-CCS 6階講堂

・講演題目:Fracture Modeling in Particulate Rafts: Surface Difference Tension and Packing Fraction Variability
・講演者:Christian Peco (Assistant Professor at Penn State University (Department: Engineering Science and Mechanics))
※発表・スライド共に英語

講演要旨:

Hydrophobic particles at a liquid-air interface tend to aggregate and form a monolayer, known as a particulate raft. These particles, otherwise non-cohesive, deform the liquid surface and interact through capillary bridges, enabling the monolayer to withstand tensile and shear deformations. The study of particulate monolayers is of great interest to scientists and engineers due to their potential for stabilizing liquid drops and emulsions via jamming. Experiments show that when a certain particle density is achieved, the raft can be characterized as a two-dimensional elastic solid, the properties of which depend on the liquid layer’s surface tension. Consequently, a practical way to further study particle rafts is through the introduction of a controlled quantity of surfactant into the system. Injected at a particular point, the surfactant decreases the surface tension and generates a front, fracturing the monolayer. The goal of this work is to determine the critical physics underlying the surfactant-driven fracture of particulate rafts. We propose a continuum approach based on a phase-field model to describe the damaged zones in the system. In the model, the fracture evolves as a result of the interplay between the pressure exerted by the surfactant and the elastic response of the monolayer. We model the monolayer behavior, accounting for the fracture toughness, the solid rigid nature of the particles, and their initial distribution. The pressure is proportional to the surface tension difference between the surfactant and the liquid layer. A comparison between the experimental observations and the numerical results indicates a qualitative match in both the fracture patterns and temporal scaling of the process. We explore the influence of particle distribution on secondary features (e.g., crack bending). Importantly, we find a dimensionless parameter that characterizes the number of cracks in the final configuration, separating different fracture regimes. We support our findings with new experimental results that confirm the trends inferred from the simulations.

第144回 第1部
日時: 2018年8月24日(金)、14:00 – 15:00
場所: R-CCS 6階講堂

・講演題目:TBA
・講演者:Daniel Patric Howard(大規模並列数値計算技術研究チーム)
※発表・スライド共に英語

講演要旨:

TBA

第144回 第2部
日時: 2018年8月24日(金)、15:15 – 16:00
場所: R-CCS 6階講堂

・講演題目:Performance Evaluation of Application Kernel using ARM Scalable Vector Extension
・講演者:小田嶋 哲哉(アーキテクチャ開発チーム)
※発表・スライド共に英語

講演要旨:

Recent CPU architectures support wide SIMD instructions with extended vector lengths. As many operations can be executed in one instruction, high performance processing is expected. ARM has announced the Scalable Vector Extension (SVE), which is an extended SIMD instruction set. The most significant feature of SVE is uniform support for vector lengths from 128 bits to 2048 bits. It realizes Vector Length Agnostic programming which does not depend on vector length.
The Arm processor designed for the post-K supercomputer will be equipped with this SVE features for SIMD acceleration.
In this talk, I will introduce an overview of ARM SVE. Then, I examine the effect of vector length and hardware resources in the performance and the runtime power consumption of benchmarks with multiple vector lengths.

第144回 第3部
日時: 2018年8月24日(金)、16:00 – 17:00
場所: R-CCS 6階講堂

・講演題目:The Evaluation of the SA-AMG method by applying Hybrid Parallelization
・講演者:野村 直也(利用高度化研究チーム)
※発表・スライド共に英語

講演要旨:

A smoothed aggregation algebraic multigrid (SA-AMG) method is among the fastest and scalable solver for large-scale linear equation Ax = b. In this talk, we demonstrate the convergence and parallel performance of our SA-AMG library on the Oakforest-PACS and K computer. SA-AMG was proposed by Petr Vanek et al. in 1992. SA-AMG achieves good convergence and scalability by damping various wavelength components efficiently. To damp these components, SA-AMG creates multi-level matrices with arbitrary policies. The created matrices are hierarchically smaller than the dimension of the original coefficient matrix. By using these multi-level matrices, error components are damped efficiently. To parallelize the SA-AMG method, we applied an OpenMP/MPI hybrid parallelization with a domain decomposition technique. Target applications of our study are large-scale and have geometrical information. On these applications, the hybrid parallelization and the domain decomposition technique are suitable. In this talk, we demonstrate numerical evaluations using Oakforest-PACS (JCAHPC) and K (RIKEN R-CCS) supercomputer system. Each supercomputer system uses cluster architecture by Intel(R) Xeon Phi(TM) and SPARC64 VIIIfx, respectively. The numerical evaluations showed good convergence and parallel performance of our SA-AMG library.

第143回
日時: 2018年8月7日(火)、15:30 – 16:30
場所: R-CCS 6階講堂

・講演題目:Software for Linear Algebra Targeting Exascale (SLATE) on Oakforest-PACS
・講演者:Yaohung Mike Tsai (University of Tennessee, Knoxville)
※発表・スライド共に英語

講演要旨: 詳細を見る

The objective of the Software for Linear Algebra Targeting Exascale (SLATE) project is to provide fundamental dense linear algebra capabilities to the US Department of Energy and to the high-performance computing (HPC) community at large. To this end, SLATE will provide basic dense matrix operations (e.g., matrix multiplication, rank-k update, triangular solve), linear systems solvers, least square solvers, singular value and eigenvalue solvers.
In this talk, I will first introduce the programming model of SLATE.
Including software dependency requirements, programming interface, current development status, and a matrix multiplication code snippet as example.
The next section would be the detail of Oakforest-PACS, an Intel Knights Landing (KNL) based system and its performance running SLATE BLAS level 3 routines.
The last part is going to be the process of implementing Cannon's Algorithm in SLATE, to showcase the advantage of the programming model, and the tracing tool for fine tuning.

第142回 第一部
日時: 2018年8月3日(金)、13:00 – 14:00
場所: R-CCS 6階講堂

・講演題目:Current Progress on Biomolecular Dynamics Simulations in Cellular Environments
・講演者:杉田 有治(粒子系生物物理研究チーム チームリーダー)
※発表・スライド共に英語

講演要旨: 詳細を見る

To understand structure-dynamics-function relationship of various biomolecules, such as proteins, nucleic acids, membranes, oligosaccharides, and many other metabolites, is essential in molecular biology as well as drug discovery. For this purpose, classical molecular dynamics (MD) simulation was first applied to a small protein, BPTI, in gas by M. Karplus et al in 1977. Since then, the simulation method together with molecular models have been continuously improved and updated by many computational biophysicists in the world.

The computational biophysics research team has developed high-performance MD software, GENESIS (GENeralized Ensemble SImulation System), for performing cutting-edge MD simulations of biological systems on supercomputer K or other computational platforms. One of the key features in GENESIS is the excellent weak scaling on massively parallel supercomputers. This allows us to perform all-atom MD simulations of the cytoplasm in small bacteria, Mycoplasma Genitarium.
This study showed the importance of non-specific protein-protein and protein-metabolite interactions on structure, dynamics, and functions of proteins in the cellular environment.

We are also developing various advanced simulation methods in GENESIS, which are the other key feature in the program. This is important to investigate slow conformational dynamics of biomolecules, such as protein folding, large-scale domain motion of membrane proteins, protein-ligand binding, and so on. In GENESIS, various enhanced conformational sampling methods, including gREST, RSE-MTD, and String method were implemented.

The accuracy of classical MD simulations highly relies on the quality of molecular force field, which consists of a number of parameters that describe bonded and nonbonded interactions. In some cases, the artifact based on the low-quality of molecular force field cannot be neglected. We are now developing a novel machine learning approach linking MD simulations with single-molecule experimental data for overcoming the problems due to the force-field biases. The machine learning approach is applicable to different experimental data for providing reliable information on three-dimensional conformational dynamics.

第142回 第二部
日時: 2018年8月3日(金)、14:00 – 15:00
場所: R-CCS 6階講堂

・講演題目:Progress report/Introduction of Large-scale Parallel Numerical Computing Technology Research Team
・講演者:今村 俊幸(大規模並列数値計算技術研究チーム チームリーダー)
※発表・スライド共に英語

講演要旨: 詳細を見る

The Large-scale Parallel Numerical Computing Technology Research Team conducts R&D of large-scale, highly reliable, and high-performance parallel mathematical software for the K computer (at the moment post-K is also another target system). Our activities are direct and indirect supports to simulation users because simulation programs require various numerical techniques to solve systems of linear equations, to solve eigenvalue problems, to compute and solve non-linear equations, and to do fast Fourier transforms. In addition, we have started to investigate data analysis tools, for example, high dimensional-sparse FFT, tensor decomposition, and acceleration techniques on some emerging devices such as an FPGA.

In last year and this year, the technical issues on the post-K computer have become more explicit. We recognized that some of the optimization technologies studied on manycore architecture are also of significance in the team activities. For a short-term research milestone, we are going to focus on following three topics in FY2018-2021(the post-K computer is expected to be launched in regular operations);
1) `Communication avoiding' algorithm,
2) `Reproducible and maintainable numerical software framework,' and
3) `Precision-and-power aware computing.'

The talk at the CCS-Café also covers our team history, the background of numerical linear algebra and high-performance computing, and some innovative topics of numerical algorithms.

第142回 第三部
日時: 2018年8月3日(金)、15:15 – 16:15
場所: R-CCS 6階講堂

・講演題目:Development of a unified continuum mechanics simulation framework for industrial applications
・講演者:坪倉 誠(複雑現象統一的解法研究チーム チームリーダー)
※発表・スライド共に英語

講演要旨: 詳細を見る

High-Performance Computing(HPC) frameworks for the fluid-structure interaction problems with complicated geometry have been developed, considering its application to industrial problems. Two frameworks are considered here; one is based on the unstructured mesh system (FrontFlow/red-HPC), and the other is based on the hierarchically structured grid system (CUBE).

For the fully unstructured finite volume method developed here, arbitrary Lagrangean-Eulerian (ALE) method together with solving the governing equations on the non-inertial reference of frame was adopted to track the complicated motion of rigid body in fluids. The method was specially tuned on the K computer, and so far it can treat up to ten billion unstructured meshes for the coupled fluid-solid motion problems.
The advantage of this framework is its accurate prediction of aerodynamic forces acting on the complicated geometry by optimizing the unstructured meshes near the wall, especially in the case flow is in turbulence state. On the other hand its disadvantage is its workload required for the pre-processing including dirty CAD treatment and mesh generation. This problem in fact prevents its spread and utilization in industrial process.

To overcome this problem, alternative framework based on the hierarchically structured finite volume method was developed, in which both the fluid motion and structure deformation are solved in Eulerian manner. To achieve higher computational efficiency of parallelization and scaling on the massively parallel environment, Building Cube Method (BCM) proposed by Nakahashi was adopted. In the method, numerical domain is first decomposed into cubic sub-domains based on the Octree method.
Then the same number of numerical grids is allocated to each cubic subdomain. In the simulation framework, the solid surface with complicated geometry is represented by the immersed boundary method (IBM). In the fluid-structure interaction problems with structure surface in motion, accurate representation of the immersed body is indispensable. Thus Lagrangian description for tracking the moving solid body surface is adopted in the Eulerian framework of solving fluid and structure motions. The parallel scalability of the numerical method and the efficacy of the load balancer were evaluated through simulation with up to 32,768 cpu cores on the K-computer. So far, the framework can handle maximum of tens of billions of numerical meshes using hundreds of thousands of CPU cores on the K-computer.

第141回
日時: 2018年7月20日(金)、15:30 – 16:30
場所: R-CCS 6階講堂

・講演題目:Computing Matrix Functions on the K Computer
・講演者:William Dawson (量子系分子科学研究チーム)
※発表・スライド共に英語

講演要旨: 詳細を見る

In this talk, I will present a new library for computing matrix functions, and will do my best to present it in a way that should be of interest to researchers from R-CCS’s many disciplines. Matrix functions have a number of applications including materials science, the solution of partial differential equations, and the study of real world networks. In this talk, I will review the theory of matrix functions, and discuss several methods of computation. Then I will present the parallelization strategy of our library, including the use of communication avoiding algorithms and task based parallelization. I will discuss usability considerations, including how to build a library in Fortran which can be called from many different programming languages. I will finish by presenting a number of different applications of this library, including quantum chemistry, social network analysis, search engine optimization, and the calculation of the eigendecomposition of a matrix.

第140回 第一部
日時: 2018年7月6日(金)、13:00 – 14:00
場所: R-CCS 1階セミナー室

・講演題目:Innovating Computer Architectures to Survive the Forthcoming Post-Moore Era
・講演者:佐野 健太郎(プロセッサ研究チーム チームリーダー)
※発表・スライド共に英語

講演要旨: 詳細を見る

To Survive the forthcoming post-Moore era, we need to innovate computer architectures in order to overcome various device constrains for successive improvement of computational performance and capability. In the processor research team, we research and develop an FPGA (Field-Programmable Gate Array)-based system which is based on the data-flow computing model.
Data-flow computing has been re-spotlighted and considered as one of the promising approaches to further scale computational performance once the multi-core scaling becomes difficult in improving performance. This is because data-flow computing can provide availability of naturally-grained parallelism and constant complexity of communication and synchronization especially when data-flow is statically mapped onto hardware.
In this talk, in addition to introduction of our team, ultimate research goals, and needs/seeds for collaboration opportunities, I talk about the FPGA-based prototype system, the data-flow compiler, and the case study of of FPGA-based Tsunami Simulation, following recent trends in FPGA devices and researches for high-performance computing.