Expenses $50 Taxi Nov 3 $15 November 10 $21 Meal on Hotel room $2 tip Boucher -- Dataflow for QR improves LAPACK or Linpack (by hand dependence analysis to get better scheduling etc.) Interval Analysis: Group in Denmark has used for chaos (ODE's) Highly tuned to this problem to get reasonable answers US3 (UltraSparc III) 100% binary compatible US2 Plus Interval arithmetic Media processing (SIMD on 8/16 bit) Fast 64 byte fetching -- don't destroy cache can get 4 instructions per cycle (6 pipelines in processor) 14 stage pipeline 750 Mhz .15 micron 243 mm**2 die 23 million transistors (12 of these RAM) 64 kbytes 4 way data cache 32 kbytes 4 way instruction TLB on chip Off chip L2 Cache 1 to 8 megabytes (tags on chip) Nice Picture of chip Software and Hardware prefetch 170 ns memory latency on 750 ns CPU -------------------------------------------------------------- Ultragrid Myrinet Ultra 80 LSF Java management 72 GB disk 4 GB menmory 16 CPU's 155,000 Euro LIST 14 Gigaflops Aiming at minimum price per gigaflop Technical computing farm Gridware is free for this system --------------------------------------------------------------- Sun Grid: Focus on ASP model with GRD follow-on managing pool of resources -- new way of thinking for campus computing Technical Compute Farm Jini access to storage Free Open source Standard API People who are desperate will pay for early access to new developments Visual Instruction Set -- specialialized Matrix multiplication on UltraSparc Compute Farms Internally 4000 CPU's >600 systems; >97% efficiency (Used in Chip Designing - testing) Can plug in other Sun Products Has Sun Grid Engine Limitation is 100 megabit Next Year T3 Disks Myrianet Ultra3 sun.com/suntcf ken.tallman@sun.com Salvini NAG Nice foils MPI and OpenMP OpenMP gives parallelism easily but performance is hard Lapack -- need to do more than parallelism of BLAS Sectors divided into 4 groups Excellence Centers Porting Applications Portals Resource Management Data-mining Java for HPC Large Data Sets -- Computational Finance, Weather, Bioinformatics This is a marketing collaboration Bizarre choice of disciplines Cryptography is academic link to eCommerce Enabling Technologies Portal Interfaces Tools Security Resource Management Visualization etc. Data Management Big emphasis on using idle compute cycles Bruce Jastremski is in charge of weather, climate, geoscience, finance joerg.schwarz@eng.sun.com Hammond has HPC Environments ----------------------------- Salvini CandC paper Gentszsch Portal White paper --------------------------------- Gannon ITR Radio Astronomy Gryphon "Virtual Data" ----------------------------- PSE DISCOM2 Milepost Zosel PSE Hale Discom 2 day April/May 2001 Meeting at Livermore First for these areas 15% ASCI budget for these 2 efforts Distibuted Multi vendor All mileposts are externally reviewed ASCI White 512 by 16 processors (Night Hawk NH-2) IBM working on 40 Severity 2 PRM's Current white switch bad -- fix end of january snow unclassified 128 CPU's NH 1 white classified 136 nodes NH-2 frost unclassifoed 376 nodes NH-2 PSE Zosel Note PSE is not John Rice PSE Whole infrastructure -- storage This is part of "white" some IBM some UNIX some Pathforwardsome ASCI small runs 3 codes (1 per lab) large runs on 1500 CPU's I/O Tools TotalView -- very popular Vampir MPI etc. Use Screendumps to document Some interactive demos Varied role for PSE in each tool DISCOM-2 Extends PSE capability to non LLNL users OC-3 ESNET + encrypters All ON classified Will be 4 times OC12 (no OC48 encrypters till june 01) Web Documentation Data movement -- using parallel OC12 Job submission Coordinated Support 2.4 gigabit linkhope to 100 megabytes per sec for viz Gigabit Ethernet locally ATM is WAN Uses Globus XML scripts for job submission wizard -------------------------------------------------------------- Grid Panel Foster: Globus is for ever Snir: e-business is for ever Pancake:Remember the users and the business model Fox: Grids are great but computing is not so important Basic architecture Beginning -- Middle -- End Services and ASP model IU Chemical engineering: Coupling Portals IBM Power4 is a miracle with initially 32 way SMP parallelism We get 16 such multi-module systems MPI with colony switch Use as a black box ---------------------------------------------------------- Originally MSRC --> CTA Now CTA ---> MSRC MSU Single point of contact MSI: JSU, CAU, Morgan State, Central State, Tribal Colleges AF DC to Dayton Army DC to ARL Navy DC to NAVO Other DC to ERDC UCF CEN Should be very strong EOT NPACI/Alliance 4 White Papers -- each 10 pages SDSC Document Sharing OSU: Portal front ends Data management ---------------------------------------------------------------- Reynders Conversation Java Grande Succeeded: Generics Agreed Overloading will happen Immutable classes will happen floating point ? Portals To Grid Forum Java MPI Alive -- direction unclear RMI Performance/Jini etc. Commercial Benchmarks Alive? Grande Programming Environment People ask me what to use to program MPP I reply F77 plus MPI Java no performance C++ future unclear Hybrid Java-C++ (F90?) Parallel Object Oriented -- F90 is object based Generic Security Reliability Component Compiler Better MPI NUMA Multi threaded OpenMP Involve funding agencies e.g Frederica -------------------------------------------------------------------- Java Grande ------------- 1. Java Grande/ISCOPE Meeting How many if any tutorials? Affects financials Location of Conference ? ISCOPE Sensitive to this as traditionallly moved internationally Student Funding IBM Sun DoE NSF Keynotes Guy Steele Start process to agree Reach out to component community Rooms for breakouts and exhibits Java University for tutorials -- Eric Sharakan -- overlaps meeting Target NSF (Koelbel) and Doe (Mary Ann Scott) Have best paper and best student paper Target applications Parabon interested in submitting paper(s) Possible Themes/panels High performance Java Environments Hybrid Java/C++ Environments -- Strengths and Weaknesses Applications Bridge science to e-commerce 2. Status 3. Future Activities Meet at Java Grande/ISCOPE Meeting Maybe an earlier february/march meeting Numerics Area Generics Agreed outside forum Overloading ? Joe Darcy Immutable/value classes ? Fast FP IBM (Snir/Moreira) Arrays IBM F90 Syntax NIST/IBM Complex Visual Numerics/Michael Phillipsen Library Framework NEW effort Concurrency etc. OpenMP Deliver "obvious" standard (Bull) MPI Declare finished (Getov) RMI Cannot change specification but can develop fast implementations on specific platforms Need to test (Gannon/Phillipsen) Benchmarking EPCC working away and making progress with comparative (C) implemedntations Portals Keep in Grid Forum go for JSR in MPI and later openMP JSR needs a champion and a reference implementation Jose E. Moreira IBM jmoreira@us.ibm.com John F. Brophy Visual Numerics brophy@vni.com J. Mark Bull EPCC markb@epcc.ed.ac.uk Jim Gannon Parabon Computation jgannon@parabon.com Thomas Ballos Parabon Computation tballos@parabon.com Dustin Lucien Parabon Computation dlucien@parabon.com Matthew J. Hunt Parabon Computation Denis Caromel University NICE - Inria caromel@unice.fr John Reynders Sun Labs john.reynders@sun.com Vladimir Getov University Westminister V.S.Getov@westminister.ac.uk Eric Sharakan Sun Microsystems eric.sharakan@east.sun.com Dennis Gannon Indiana University gannon@cs.indiana.edu Geoffrey Fox FSU fox@csit.fsu.edu ------------------------------------------------------------------------ Gregor Likes betterportalML Dennis a little worried Tom Haupt did not come to SC00 -- NOT brought up by Thompson at all ------------------------------------------------------------------------- Concurrency and Computation:Practice and Experience 2 meetings: Kennedy Baker Walker Johnsson Issues: Name Concurrency is not a "hot" name Puts off application people Rather jaDed even for Computer Science Replace by Communication or put computation first Could lose subscriptions Online Administration Agree to Wiley online Software Profile Submission Process Delay is at most 4 months refereeing Then 4 to 8 months publication if necessary delay special issues So total <=8 months unsolicited, <=12 months special issues Editorial Board Tapos Ask them. Lieberherr not so enthusiastic Zicari more energetic Remember they suggested Harold Ossher from IBM for board They should be asked for around 3 more with some geographic distribution Suggested Names for boatd Xiaoming Li Fran Berman Ian Foster Larry Synder Gregor von Lazweski HAS BEEN ASKED BY BAKER Manish Parashar Omer Rana Tom Sterling (or his JPL colleague Daniel Katz) Vladimir Getov David Callahan Duties Suggest succesful special issue Monitor workshops for special issues Cite CandC:PandE papers solicit 1 paper take advertising material to meetings Have a telecon with board to nominate upto 5 names classified by field and geography (US Europe, Asia) and experience (50-50 wise owls and up and coming) List will be Anonymous Need balance Other topic editors Need Japan Editorial on new name End of November Suggested Meetings Portal working group HPF Users group SC00 Best Papers Special Issues Grid Peer to Peer Computing Clusters Baker ASCI Messina OAk Ridge -- Climate Walker Java Performance -- Cherri Pancake Produce CandC:PandE list of citations in bibtex make possible to automatically download or to copy and paste needs to add link or text to Wiley official page Increase Impact Factor Ensure Search Engines find CandC:PandE papers Need right type of keywords -- extract from abstract e.g. "multi-level blocking" not "compiler optimization" add in review process Analyse Web Hits and use to persuade universities to subscribe e.g.number of people accessing abstracts from university Make it easy to bug universities to subscribe Messina becomes board member; not associate editor Have class of Topic Editors Object Systems -- TAPOS editors Performance -- Mark baker Add byline (strap line) explaining title High Performance Systems (7 or 8 words) ------------------------------------------------------------------- Contact Mary Thomas Terry Diesz and Access Grid Need Network and Placement details / contact people How to be used Support There is a portable AG but no personal one Can one drive multiple displays with same infrastructure ---------------------------------------------------------------------- Joe Thompson Use FCU and Georgia Tech not SAIC for FMS ----------------------------------------------------------------------- Linda Callahan Working with WebWisdom.com on NASA pork barrel ----------------------------------------------------------------------- IBM Briefing ****************************************************** AIX-L is Linux flavor of AIX ACTC division does training UNM has Linux Cluster plus SP Power3 Chip 23 Million Transistors Power 4 Chip 2 CPU per chip L2 on chip is 1,5 MB L3 on chip is 32 MB 4 chips on a mdule 4 modules gives a 32 way SMP Something is 170 million transistors Speed 1100 Mhz 2001 1800 Mhz 2003 32 way (one per "cabinet) is defintely a full SMP Eventually full system is a NUMA SMP LPAR Logical Partitioning down to level of CPU SP 2001 Colony Switch 15 microseconds latency -- Can have multiple NUMA 2003 Federation Switch 4 microseconds ia64 only has one CPU per chip Sledgehammer is AMD 64 bit which is sama architecture as iA32 Competition Sledgehammer AMD iA64 Intel PowerPC4 IBM Alpha Compaq PSSP SP Systems Management is better than current Linux Linux Systems Use 1 Ghz iA32 with 2 CPU's per node Turbo or Redhat Linux Using only 2 CPU per node is a Linux restriction Myrianet and SP switches are both flat omega networks about 9 microsecond latency 2 gigabits per second bandwidth