Docstoc

PowerPoint Presentation - Microsoft Research

Document Sample
PowerPoint Presentation - Microsoft Research Powered By Docstoc
					              The Future
                   of
         Scientific Computing

             Microsoft Faculty Summit, 2005

               Dr. Francine Berman
    Director, San Diego Supercomputer Center
Professor and HPC Endowed Chair, UC San Diego

 SAN DIEGO SUPERCOMPUTER CENTER

                                      UNIVERSITY OF CALIFORNIA, SAN DIEGO
                        Fran Berman
                Dan’s Questions
1. What are the key technical issues your centers
   currently face?

2. What research problems do you encounter in
   supporting scientists?

3. What will be the top challenges to
   Computational Sciences in 5-10 years?



    SAN DIEGO SUPERCOMPUTER CENTER

                                         UNIVERSITY OF CALIFORNIA, SAN DIEGO
                           Fran Berman
      Today’s Technologies:
      Enablingversion of the
the Branscomb Pyramid, circa 1993
       Branscomb Pyramid
                       100+TF Class
             TF TF
                  TF                                           Key function of the
           ClassClass
                 class
           Center
               Center                                                  NSF




                                        Scientific Computing
             Mid-large
           Super-
               Super-                                            Supercomputer
               scale
             computers
         computers
           clusters and                                              Centers:
         parallel machines                                      Provide facilities
           Mid-range
               Mid-range                                         over and above
         parallel processors
      parallel processors                                       what can be found
      networked workstations
  andand networked workstations
          Personal devices,                                       in the typical
   Home and desktop computers,
        High Performance
                                                                   campus/lab
           workstations
     High Performance
        High Performance                                          environment
           Workstations
       Workstations




  SAN DIEGO SUPERCOMPUTER CENTER

                                       UNIVERSITY OF CALIFORNIA, SAN DIEGO
                         Fran Berman
There’s More to the Story …
                                   Data-oriented
                                      Science
                                  and Engineering
                                    Environment
        Data (more BYTES)




                            Home, Lab,      Traditional
                             Campus,           HPC
                             Desktop       environment




                             Compute (more FLOPS)
 SAN DIEGO SUPERCOMPUTER CENTER

                                            UNIVERSITY OF CALIFORNIA, SAN DIEGO
                             Fran Berman
Today’s
scientific
applications                       Data Mgt. Envt.         Extreme I/O Environment
span the
spectrum of
                                                   Data-oriented Climate
                                                          SCEC
usage and                                     SCEC ScienceSimulation
                                            Visualization                    ENZO
requirements                       EOL          and Engineering              simulation
                                                 NVO            ENZO
                                                    Environment
               Data (more BYTES)
                                                              Visualization Turbulence
                                                   GridSAT                     field
                                   CiPres                       CFD

                                    Seti@Home                  MCell
                                    Home, Lab,                 Traditional
                                                                        Protein
                                                                       Folding/MD
                                     Campus,                      HPC
                                                                       CPMD                 Lends itself to Grid
                                     Desktop                  environment
                                                                            QCD              Could be targeted
                                              GAMESS                                         efficiently on Grid
                                                               Turbulence
                                                              Reattachment                   Difficult to target
                                   EverQuest                     length                      efficiently on Grid



                                      Compute (more FLOPS)
       SAN DIEGO SUPERCOMPUTER CENTER

                                                                 UNIVERSITY OF CALIFORNIA, SAN DIEGO
                                      Fran Berman
      SDSC: Using Data as a Driver
                               SDSC           in a nutshell
   DataStar                    •   National Cyberinfrastructure                          ENZO
   IBM Power 4                     Center                                                astrophysics
                               •   UCSD Organized Research Unit

 Data-      IntimiData         •   Staff includes 400 multi-
oriented    Blue Gene              disciplinary IT professionals,           TeraShake
              --Data               applied researchers,                     geosciences
HPC and                            technologists, and students
storage                                                                          Data-oriented
                               Projects include                              scientific applications
                                         NEES IT Center
HPWREN                                 (earthquake engineering)
              NVO                       Protein Data Bank
 Sensor                                     (life sciences)
  Data
                                              CAIDA
           Community                    (internet data analysis)
           Databases                            GEON
               and                            (geosciences)
              Data
                                               TeraGrid                            Data-oriented
           Collections                      (Grid Computing) ++                Software and Services
           SAN DIEGO SUPERCOMPUTER CENTER

                                                           UNIVERSITY OF CALIFORNIA, SAN DIEGO
                                   Fran Berman
Today’s Challenges for SDSC




 SAN DIEGO SUPERCOMPUTER CENTER

                                      UNIVERSITY OF CALIFORNIA, SAN DIEGO
                        Fran Berman
   Integration and Coordination
• Today’s “computer”
                                          Internet
  is an integrated and
  coordinated set of
  hardware, software,
  and services.




     SAN DIEGO SUPERCOMPUTER CENTER

                                               UNIVERSITY OF CALIFORNIA, SAN DIEGO
                            Fran Berman
Integration and Coordination Challenges:
       Sensors to Supercomputers

• Computational
  scientists and
  engineers
  need to
  integrate
  resources at
  all scales to
  support
  “end-to-end”
                                          SDSC’s Notebook Project
  applications                         Enables scientists to integrate data
                                        management, collaboration, and
                                      computation environments in a digital
                                             laboratory notebook
     SAN DIEGO SUPERCOMPUTER CENTER

                                              UNIVERSITY OF CALIFORNIA, SAN DIEGO
            Graph courtesyFran Berman
                          of Henri Casanova
                 Incorporating the “ilities”
 Scalability, Predictability




                                                                               Software engineering
                                                                               fundamental for
                                                                               modern scientific
                                                                               codes – managing
Predictable                                                                    distribution,
performance                                                                    accommodating
models key                                                                     multiple users and
to execution                                                                   web interfaces,
optimization                                                                   integrating with
for scientific                                                                 complex SW envt. Is
codes                                                                          key

                                                     Usability, Interoperability,
                                                             Flexibility
            SAN DIEGO SUPERCOMPUTER CENTER

                                                       UNIVERSITY OF CALIFORNIA, SAN DIEGO
                                 Fran Berman
                   Graphs courtesy of Jenny Schopf
    Incorporating the “ilities”: Reliability and
                  Sustainability
•   Computational scientists and
    engineers increasingly rely on           Entity at
                                                              What can go wrong                Frequency
                                               risk
    persistent and valuable
    community data collections                            Corrupted media, disk
                                            File                                          1 year
                                                          failure
•   Extreme data curation for
    100 years or more involves                            + Simultaneous failure of 2
                                            Tape                                          5 years
                                                          copies
    long-term planning and a
    strategic approach to support                         + Systemic errors in vendor
                                                          SW, or malicious user, or
•   Two approaches used by the              System                                        15 years
                                                          operator error that deletes
    preservation community:                               multiple copies

     1. Make lots of copies                               + Natural disaster,
                                            Archive                                       50 - 100 years
                                                          obsolescence of standards
     2. Make copies in
        heterogeneous SW
        environments
                                                   Data Reliability and Sustainability:
                                                          What can go wrong
           SAN DIEGO SUPERCOMPUTER CENTER

                                                         UNIVERSITY OF CALIFORNIA, SAN DIEGO
                                  Fran Berman
    Tomorrow’s Challenges
          for SDSC




SAN DIEGO SUPERCOMPUTER CENTER

                                     UNIVERSITY OF CALIFORNIA, SAN DIEGO
                       Fran Berman
Better Application Performance through Adaptivity

Everyware                                                                      Program Performance by
                                                                                                                              Legion
[Wolski et al., 1999]                                                          Infrastructure Type                            Condor
                                                                                                                              NT
                                                                                            5 Minute Averages                 Globus
-- a highly adaptive (non-                                          1.00E+10                                                  Unix
                                                                                                                              Java
embarrassingly parallel)                                                          NT           Unix         Legion            Netsolve
                                                                    1.00E+09
Grid application which
investigated solutions to the                                       1.00E+08




                                          Integer Ops. Per Second
Ramsey Number Problem:                                              1.00E+07

    What is the smallest                                            1.00E+06
    complete undirected two-
    colored (“red” and “green”)                                     1.00E+05
                                                                                         Condor
    graph R(m,n) such that                                                      Globus
                                                                    1.00E+04
    there a red clique of size m
                                                                                             Java                        Netsolve
    or green clique of size n?                                      1.00E+03




         SAN DIEGO SUPERCOMPUTER CENTER

                                                                                   UNIVERSITY OF CALIFORNIA, SAN DIEGO
                              Fran Berman
                 Graph courtesy of Rich Wolski
         The Dynamics of Sharing
• As scientific applications
  require more and more
  components, with more
  and more sophisticated
  interactions, models for
  effective group
  dynamics, policy,
  distribution of control,
  and other “social” issues
  will become critical for
  success.

       SAN DIEGO SUPERCOMPUTER CENTER

                                            UNIVERSITY OF CALIFORNIA, SAN DIEGO
                              Fran Berman
          Group Dynamics –
Innovation from the Commercial Sector




                                                            IM – Group
                                                          communication




   RPG – Group Entertainment                E-Bay – the group dynamics
                                                    of shopping
SAN DIEGO SUPERCOMPUTER CENTER

                                     UNIVERSITY OF CALIFORNIA, SAN DIEGO
                       Fran Berman
           The Next Generation of Scientists,
         Consumers, and Leaders is Tech-savvy
•   Assume
     •   That everything is available
         online and everything is free
         (the Web)
     •   That none of the resources
         have to be where you are
     •   That you can communicate
         with anyone anytime
         (email, IM)
     •   That you can adapt to things
         in real time
         (RPG)
     •   Some rudimentary level of
         competence with “business
         models” (Sim environments)
                                                Expect traditional ways of
•   That’s the baseline …                       doing things to change …
              SAN DIEGO SUPERCOMPUTER CENTER

                                                       UNIVERSITY OF CALIFORNIA, SAN DIEGO
                                      Fran Berman
                        Thank You




                                 www.sdsc.edu

SAN DIEGO SUPERCOMPUTER CENTER

                                      UNIVERSITY OF CALIFORNIA, SAN DIEGO
                       Fran Berman

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:4
posted:8/1/2011
language:English
pages:17