Date and Place

日時,会場

12th(Mon.), March, 2018, 10:00-17:00
CyberMediaCommons, Cybermedia Center 1F, Suita Campus, Osaka University, JAPAN. access

2018年 3月12日(月) 10:00--17:00
大阪大学サイバーメディアセンター 1F, サイバーメディアコモンズ
アクセス



Reception

懇親会

12th(Mon.), March, 2018, 18:00-
Salvatore Cuomo & Bar SenriChuo, SenriChuo, Osaka,
4000 JPY/person.
* Please mail to Daisuke Furihata (furihata@cmc.osaka-u.ac.jp) by 4th March, to participate this reception. All of speakers are already included in the participant list.

2018年 3月12日(月) 18:00-
Salvatore Cuomo & Bar 千里中央, 千里中央, 大阪
4,000円/人.
* 懇親会参加を希望される方は、降旗 大介 (furihata@cmc.osaka-u.ac.jp) 宛に 3/4 までにメールをいただければ幸甚です． なお、講演者は既に参加者としてリストアップしてあります．

Program (AM)

プログラム(午前)
10:00- (opening)

10:05-10:50
Speaker:
Pierre Jolivet
(ENSEEIHT)
Title:
Finite element analyses with domain specific languages and high-performance computing
Abstract:
Mixing domain specific languages and high-performance computing enables researchers to perform complex simulations. In this talk, I will present FreeFem++, a finite element language, and how it may be used to discretize various complex applications, such as radiative transfer problems, or computational fluid dynamics. On top of this kernel, users may access various linear algebra backends to solve the discretized PDEs using disitributed-memory parallelism on large-scale clusters.


10:55-11:40
Speaker:
Kensuke Aihara
(Faculty of Knowledge Engineering, Tokyo City University)
Title:
Alternative implementation of the residual smoothing technique for improving the convergence of CGS-type methods
Abstract:
We treat the CGS-type iterative methods for solving large sparse linear systems. The methods often have large oscillations in the residual norms, which affect the maximum attainable accuracy of the approximate solutions. The residual smoothing technique is useful to obtain a smooth convergence behavior, but it does not help to improve the attainable accuracy. In this talk, we propose an alternative implementation of combining the CGS-type methods and the smoothing technique to avoid severe propagation of rounding errors. Numerical experiments demonstrate that our proposed methods attain more accurate approximate solutions. This is joint work with Ryosuke Komeyama and Emiko Ishiwata.


11:45-12:30
Speaker:
François-Xavier Roux
(ONERA/LJLL, Sorbonne Université)
Title:
Schur complement and Dirichlet to Neumann operator for domain decomposition solvers in structural mechanics and fluid-structure coupling
Abstract:
For domain decomposition methods or for solving strongly coupled problems in mechanics, generalized Robin boundary conditions allow imposing in one domain a relation between the trace of the solution field and its flux which takes into account the interaction with the neighboring domains. Optimal conditions are based on the Dirichlet to Neumann mapping of the external domain. The discrete form of the Dirichlet to Neumann matrix is the Schur complement on the interface. Both analytical and linear algebra approaches can be used to find good approximations, leading to fast convergence of the coupling iterations.

Program (PM)

プログラム(午後)
13:40-14:25
Speaker:
Toshiyuki Imamura
(RIKEN Advanced Institute for Computational Science)
Title:
Communication-Avoiding approach for acceleration of a dense-eigenvalue solver
Abstract:
Emerging post-Moore’s era, we might meet strangeness in a deeper memory-network hierarchy, which becomes one of the walls of the future parallel programming. We recognized that the typical approaches to overcome such strangeness in numerical linear algebra are known as so-called block algorithm and communication avoiding technique. So far, we applied both techniques in dense eigenvalue solver, the EigenExa library, and confirmed 15 to 20 % performance improvement on the K computer system.


14:25-15:10
Speaker:
Seiji Fujino
(Emeritus Professor, Kyushu University)
Title:
On parallel Cache-Cache (Hide and Seek) computing and its performance estimation
Abstract:
We consider to solve efficiently a linear system of equations by preconditioned Krylov subspace methods on parallel computers with distributed memory. We want to improve parallel performance of Eisenstat type of SSOR preconditioning owing to adoption of strategy of Cache-Cache (= hide and seek in English) computing.
Through many numerical experiments, we will make clear that parallel performance of E-SSOR preconditioning outperforms with other preconditionings in view of convergence rate and robustness.


15:10-15:25 (break)

15:25-16:10
Speaker:
Ning Zheng, Ken Hayami†, and Nobutaka Ono‡
(† National Institute of Informatics, speaker,
‡ Tokyo Metropolitan University)
Title:
Fast solution of nonnegative matrix factorization via a matrix-based active set method
Abstract:
For the solution of nonnegative matrix factorization, which is a low rank matrix approximation problem with nonnegative constraints, we propose an alternating nonnegative least squares method by utilizing modulus-type inner outer iteration to solve the nonnegative constrained least squares problem with multiple right hand sides at each iteration. We also propose a fast matrix-based active set method to improve the convergence. Numerical experiments show that the proposed methods converges faster than the state-of-the-art methods.


16:10-16:55
Speaker:
Atsushi Suzuki
(Cybermedia Center, Osaka University)
Title:
A sparse direct solver : Dissection, in quadruple precision, with kernel detection, and on parallel architecture
Abstract:
A software, Dissection to perform LDU factorization of large sparse matrix on parallel multi-core system is developed. The given matrix is decomposed into a union of sparse and dense sub-matrices by nested-dissection algorithm, and each sub-matrix is factorized with symmetric pivoting and postponing computation to recover accuracy under multi-frontal factorization. Quadruple precision arithmetic is feasible thanks to implementation with C++ template. Dissection has kernel detection capability of the singular matrix and these two features are helpful to understand ill-conditioned linear system.


16:55- (closing)