第11回ASE研究会のお知らせ

日時：2012年3月16日（金）13時00分～17時25分（懇親会：18時00分～）
場所：東京大学　情報基盤センター（本郷キャンパス弥生地区）４階遠隔講義室（地図）
主催：東京大学情報基盤センター　スーパーコンピューティング部門

懇親会参加の方は、末尾の参加票をお送りください。

11th Advanced Supercomputing Environment (ASE) Seminar

16th March 2012 (Friday)
Information Technology Center, The University of Tokyo (Hongo Campus)
4th Floor, Telecommunication Lecture Room

13:00 - 14:00

Invited Speaker

　Professor Serge G. Petiton (University of Lille, CNRS / LIL and INRIA)

Title

　Toward future smart auto-tuned Parallel Krylov methods and programming for Exascale Computing

Abstract

　Auto-tuning Krylov subspaces sizes at runtime for the GMRES(m) methods may be efficient to minimize the global computing time toward convergence. We introduce a general algorithm to auto-tuned the Krylov subspace sizes at run time, and we also introduce the first algorithm to auto-tune at run-time the number of vectors targeted by an incomplete orthogonalizations of Krylov basis associated with the GMRES(m) method, minimizing in particular the number of dot-products for a fixed subspace size. We present and analyse in this talk these new algorithms and we propose some experimental results. As a conclusion, we introduce some future projected researches which would lead to smart auto-tuned Hybrid Krylov methods for exascale computing, including language issues.

14:05 - 14:25

Speaker

　Junichiro Shiomi (Department of Mechanical Engineering, The University of Tokyo)

Title

　Calculations of phonon transport in semiconductors based on first principles.

Abstract

　Encouraged by the recent progress in nanostructuring techniques to enhance the conversion efficiency of thermoelectric materials by reduction of lattice thermal conductivity, we aim to develop a numerical tool that evaluates microscopic heat conduction characteristics, namely phonon transport properties, based on first-principles. The capabilities and limits of the new framework for designing thermoelectrics will be discussed.

14:30 - 14:50

Speaker

　Kimio Kuramitsu (Yokohama National University)

Title

　A scripting language design and implementation for HPC computing

Abstract

　This talk presents our attempt to extending Konoha scripting language for HPC computing. Konoha is a newly designed scripting language that has static typing feature. As a static typing helps produce well-optimized execution code, Konoha is expected to run scripts much faster than existing dynamic scripting languages, such as Python and Ruby. The speaker will report our experimental results with several efforts, including GPGPU integration, LLVM-base JIT compilation, and MPI-based extension. The myth of HPC scripting will be revealed.

14:55 - 15:15

講演者

　宮下秀樹（山梨大学大学院　医学工学総合教育部）

タイトル

　クラスタシステムにおける高速な特異値分解のための前処理、後処理の実装

概要

　特異値分解はさまざまな分野で利用される重要な数値計算アルゴリズムであり、コンピュータの性能向上に伴ないますます大規模化の要求が高くなっている。最新の特異値分解アルゴリズムは主として二重対角行列向けなので、一般行列の特異値分解には前処理として二重対角化が必要となる。この前処理には QR 分解が使用される。
　本研究の目的は T2K オープンスパコン上で、OpenMP と MPI による Hybrid プログラミングにより QR 分解を実装することである。

15:20 - 15:40

Speaker

　Hidetaka Muguruma and Yuichi Tsujita (Department of Electronic Engineering and Computer Science, Faculty of Engineering, Kinki University)

Title

　Study of A Pipelined Processing for High Throughput Collective MPI-IO

Abstract

　One of the MPI-IO implementations named ROMIO provides high performance collective MPI-IO with the help of its Two-Phase I/O protocol. Two-Phase I/O consists of repetitive operations including file accesses and data communications. We propose a pipelined processing implementation by overlapping file accesses with communications for further throughput improvements. We have evaluated its performance on the T2K Open Supercomputer with a Lustre file system. We observed effectiveness of our implementation. Furthermore, we found effective memory utilization in some access patterns compared with the original Two-Phase I/O.

15:45 - 16:30

Speaker

　Yoshikazu Kamoshida (Information Technology Center, The University of Tokyo)

Title

　Jitter Quantification on a Supercomputer in Operation

Abstract

　Analyzing and preventing system noises on the parallel computing environment are one of very important problems to run the parallel applications with better scalability. We introduce an analysis method of system noises which estimates the processes that are causing the delays of the parallel application and the degree of them. We also discuss about our light-weight monitoring technique which enables our method to work efficiently on the running supercomputer.

16:35 - 17:20

Speaker

　Takahiro Katagiri (Information Technology Center, The University of Tokyo)

Title

　New Auto-tuning Functions and Its Effect of Xabclib and OpenATLib Ver.1.0

Abstract

　We have developed a sparse iterative solver with an auto-tuning (AT) facility, named Xabclib. In this presentation, we present new functions of the AT ver. 1.0. The main functions in Xabclib and OpenATLib Ver.1.0 are summarized as follows:
　First, we have developed new AT for selection of preconditioners and solver algorithms. It is known that selection of preconditioners is crucial function for iterative solvers; however, automatic selection was very difficult in general. We propose an algorithm for detecting stagnation of series of residual errors. By utilizing the algorithm of stagnation detecting, the solver establishes high convergence ratio to that of conventional solvers.
　Second, but this is not new, we have developed an automatic selection of implementations for sparse matrix vector multiplications (SpMV). Especially, we provide a new implementation of segmented scan (SS) method for scalar processors. This enables us a dynamic parameter change for the new SS on SpMV.
　We show several results of performance evaluation with the ver.1.0 with one node (16 threads) of the T2K Open supercomputer (U.Tokyo).

17: 25　Closing

　Takahiro Katagiri (The University of Tokyo)

18:00　A Banquet near Nedu station

• センターユーザに限定せず、研究会は一般公開とします。
• 参加費は無料で、基本的に事前登録は不要です。

また、今後の開催予定を確実に知りたい方は、メーリングリストへの登録をお願いします。登録依頼については、下記問い合わせ先までお願いします。

〒113-8658　東京都文京区弥生2－11－16
東京大学　情報基盤センター

ASE研究会幹事　准教授　片桐孝洋
E-mail：katagiri＠cc.u-tokyo.ac.jp
（”＠”を半角にしてからお送りください。）