[san-tech][03197] Graph 500ベンチマークをソフトウェアで高速化, E.T. International, Inc. (ETI)

Date: Wed, 22 Jun 2011 08:26:47 +0900
------------------------------------------------
2011/06/23
[san-tech][03204] Graph 500 Complete Results - June 2011公開
[san-tech][03199] Re: Graph 500ベンチマークをソフトウェアで高速化, E.T. International, Inc. (ETI)
------------------------------------------------
何故か、Graph 500の 2011年 6月版 (最新版) が未だ公開されてないのですが、
今まで紹介してきた Cray XMTや Conveyはメモリアクセスをハードウェア的に
工夫して、Graph問題で強みを発揮してますが、ソフトウェアで高速化のプレス
リリースです：

"E.T. International’s SWARM Runtime System Achieves Significant
 Performance Improvements on Graph 500 for Intel Endeavor"
 June 14, 2011
  http://www.etinternational.com/index.php/news-and-events/press-releases/june-14-2011-e-t-international-s-swarm-runtime-system-achieves/


  "E.T. International, Inc. (ETI), a provider of system software
   solutions for hybrid and many-core hardware architectures, today
   announced significant speed improvements of the Graph 500 benchmarks
   on the Intel Endeavor supercomputer when using ETI's SWARM
   (SWift Adaptive Runtime Machine) system software. 

   ETI's SWARM runtime system improves the Graph 500 execution time by
   up to 11-fold over MPI.  When compared with the published 2010
   Graph 500 list, this result would put SWARM into the top two.
   Specifically, at 256 nodes (Scale 33), it delivers 6.9 giga-edges
   per second (GE/s), and at 128 nodes (Scale 32), 4.2 GE/s.

ちなみに、November 2010版の 1位 ～ 3位は
Complete Results November 2010
  http://www.graph500.org/Results.html
1: Intrepid (IBM BlueGene/P, 8192 nodes/32k cores), Argonne National Laboratory
   Scale 36 (Medium), 6.6 GE/s: Optimized
2: Franklin (Cray XT4, 500 of 9544 nodes), NERSC
   Scale 32 (Small), 5.22 GE/s: Optimized
3: cougarxmt (128 node Cray XMT), Pacific Northwest National Laboratory
   Scale 29 (Mini), 1.22 GE/s: Optimized

この時点では：
5: Endeavor (256 node, 512 core Westmere X5670 2.93, IB network),
   Intel Corporation, Scale 29 (Mini), 533 ME/s: Reference
9: Endeavor (128 node, 256 core Westmere X5670 2.93, IB network),
   Intel Corporation, Scale 26 (Toy), 615.8 ME/s, Reference

SWARMの開発は Intelと共同ですが、冒頭のプレスリリースで：
  "As a software partner in the Intel-led team for DARPA's Ubiquitous
   High Performance Computing (UHPC) program, ETI developed the SWARM
   runtime system technology for the UHPC initiative and improved the
   Graph 500 benchmark performance on the Endeavor supercomputer by
   using this system software instead of MPI.  Intel's UHPC team,
   jointly funded by Intel and DARPA, includes ETI Vice President of
   R&D Dr. Rishi Khan among its principal investigators."

※UHPCについては
[san-tech][02406] US DOE DARPA 新 HPCプロジェクト (OHPC) 公募
[san-tech][02527] Re: US DOE DARPA 新 HPCプロジェクト (OHPC) 公募
> プロトタイプ (第一フェーズだけの契約だと思います) 4チーム
>   Intel Corporation,
>   NVIDIA Corporation,
>   Massachusetts Institute of Technology Computer Science and Artificial Intelligence Laboratory,
>   Sandia National Laboratory,
※元メール時からリンクが削除されたりしています。追跡出来た分は Webで
更新してあります。

DARPAの予算ですが、SWARM Beta版は公開されています (要登録)
SWARM Beta Registration
  http://www.etinternational.com/index.php/register1/

White Paper
"ETI Many-Core Software Architecture for Large Scale Systems"
  http://www.etinternational.com/index.php/download_file/view/37/103/
"ETI Software System Technical Briefing"
  http://www.etinternational.com/index.php/download_file/view/17/103/

SWARM Documentation
"SWARM2c: Translating SWARM MLL To SWARM Low Level C API"
  http://www.etinternational.com/index.php/download_file/view/19/103/
※↑の 3編は全て PDFファイル

Publications
  http://www.etinternational.com/index.php/resource-center/publications/


実は、ET International, Inc.は興味深い会社です。設立は 2000年で、

Management Team
  http://www.etinternational.com/index.php/company/management-team/
Dr. Juan Del Cuvillo, Vice President of Engineering
  "...the technical lead for the $22M Cyclops-64 Software Development
   Project aimed to design and deliver the system software that powers
   the Petaflop 13,824-node Cyclops-64 system."

Board Of Directors
  http://www.etinternational.com/index.php/company/board-of-directors/
Dr. Guang Gao, Founder and Chairman, Distinguished professor at
the Department of Electrical and Computer Engineering at
the University of Delaware
  "... he is also the founder and leader of the Computer Architecture
   and Parallel Systems Lab (CAPSL)"

と、Cyclops-64の影響が強いようです。Cyclops-64については、

Cyclops64
  http://ja.wikipedia.org/wiki/Cyclops64
Cyclops-64, Ioannis E. Venetis, Post-Doctoral Associate, U of Delaware
  http://www.capsl.udel.edu/~venetis/Cyclops-64.html

Guang R. Gao
  http://www.capsl.udel.edu/~ggao/
Computer Architecture and Parellel Systems Laboratory (CAPSL)
  http://www.capsl.udel.edu/

さらに、Gao教授は、ISC'11で Gauss Awardを表彰されています：
"E.T. International Founder Wins 2011 Gauss Award at International
 Supercomputing Conference"
 June 10, 2011
  http://www.etinternational.com/index.php/news-and-events/press-releases/june-10-2011-e-t-international-founder-wins-2011-gauss-award-at/

  "... Dr. Guang R. Gao is the 2011 recipient of the annual Gauss Award,
   which recognizes the most outstanding paper in the field of scalable
   supercomputing at the International Supercomputing Conference (ISC)."
.....
  "... the Fresh Breeze memory model and system architecture as an
   approach to achieving significant improvements by supporting
   fine-grain management of memory and processing resources and
   utilizing a global shared name space for all processors and
   computation tasks.  Specifically, the Fresh Breeze memory model
   uses trees of fixed-size chunks of memory to represent all data
   objects."

Awards, ISC'11
  http://www.supercomp.de/isc11/Program/At-a-Glance/Awards
Gauss Award Winning Paper

"Experiments with the Fresh Breeze Tree-Based Memory Model"
 Jack Dennis, Massachusetts Institute of Technology;
 Guangrong Gao, University of Delaware;
 Xiaoxuan Meng, University of Delaware
  http://www.supercomp.de/isc11_ap/php/HTML/contribution_detail.php?con_id=1435&evid=189&request_day_id=13
  http://www.capsl.udel.edu/pub/doc/papers/ISC2011-Meng-FreshBreeze.pdf

Publications, CAPSL
  http://www.capsl.udel.edu/publications.shtml
※Cyclops-64関連も沢山公開されています。


最近の投稿から
[san-tech][03190] Slidecast: "Convey Hybrid Core Computing Rocks the Graph 500", 06.18.2011, insideHPC、(2011/06/20)

> この件に関して、
>
> "最近の話題 2011年6月18日", Ando's Processor Information Page
>   http://www.geocities.jp/andosprocinfo/wadai11/20110618.htm
>   ２．ConveyがGraph 500ベンチマークで773MTEPSを達成
>
> で詳しく解説しています。

> Graph 500一般に関しては
> [san-tech][02900] Cray XMT (Massively Multithreaded Platform) プレスリリース等
> > "まったく異なるスパコンの性能評価「Graph500」"
> >  2011/01/12
> >  SC10 - HPC Challenge、Green500とGraph500 (3)
> >   http://journal.mycom.co.jp/articles/2011/01/12/sc10_hpc_challenge/002.html
> ※これも安藤さんの解説記事です。
------------------------------------------------
[san-tech][03087] "Government Developing Data Mining Tools To Fight Terrorism", 2011/05/12
[san-tech][03111] Re:"Government Developing Data Mining Tools To Fight Terrorism", 2011/05/12
[san-tech][03129] Re:"Government Developing Data Mining Tools To Fight Terrorism", 2011/05/12