“The Future of Supercomputing”
Cynthia A. Patterson, CSTB (Study Director)
Charles Brownstein, Director
Computing and Telecommunications Board
National Research Council
1
CSTB Members
David D. Clark
Eric Benhamou
Elaine Cohen
Thomas E. Darcie
Mark E. Dean,
Joseph Farrell,
Joan Feigenbaum
Hector Garcia-Molina
Randy Katz
Wendy A. Kellogg,
Sara Kiesler
Butler W. Lampson
David Liddle, U.S.
Teresa H. Meng
Tom M. Mitchell
Daniel Pike
Eric Schmidt
Fred B. Schneider
Burton Smith
William Stead
Andrew Viterbi
Jeannette M. Wing,
Massachusetts Institute of Technology, Chair
3Com Corporation
University of Utah
University of Victoria
IBM Systems Group
University of California, Berkeley
Yale University
Stanford University
University of California, Berkeley
IBM T.J. Watson Research Center
Carnegie Mellon University
Microsoft Corporation
Venture Partners
Stanford University
Carnegie Mellon University
GCI Cable and Entertainment
Google Inc.
Cornell University
Cray Inc.
Vanderbilt University
Viterbi Group, LLC
Carnegie Mellon University
2
CSTB Sponsors
Federal Government
Air Force Office of Scientific Research
Defense Advanced Research Projects Agency
Department of Commerce / National Institute of Standards and Technology
Department of Energy
National Aeronautics and Space Administration
National Institutes of Health/National Library of Medicine
National Science Foundation
Office of Naval Research
Foundations and Other Independent Sources
AT&T Foundation, Carnegie Corporation of New York, Rockefeller Foundation Sloan
Foundation, Vadasz Family Foundation, W.K. Kellogg Foundation
Industry (Unrestricted Funding)
Cisco Systems, Intel Corporation, Microsoft Research
3
Underway
•
•
Authentication Technologies and Their Privacy Implications
Digital Archiving and the National Archives and Records Administration
•
•
•
•
•
•
•
•
•
•
Frontiers at the Interface of Computing and Biology
Fundamentals of Computer Science: Challenges and Opportunities
The Future of Supercomputing
Improving Cybersecurity Research in the United States
Internet Navigation and the Domain Name System
Privacy in the Information Age
Radio Frequency Identification (RFID) Technologies: A Workshop
Review of the FBI Program for IT Modernization
Sufficient Evidence? Building Certifiably Dependable Systems
Telecommunications Research and Development
•
•
Whither Biometrics?
Wireless Technology Prospects and Policy Options
4
Past work on HPC
Evolving the High Performance Computing and Communications Initiative to Support
the Nation's Information Infrastructure (Jan 1993)
High Performance Computing and Communications: Status of a Major Initiative (July
1994).
Workshop on Advanced Computer Simulation and Visualization (November 1992)
Computing and Molecular Biology: Mapping and Interpreting Biological Information
(April 1991)
Supercomputers: Directions in Technology and Applications (Sept 1988)
White Paper: Survey on Supercomputing and Parallel Processing (May 1989)
5
“Future” Study Committee
SUSAN L. GRAHAM
University of California, Berkeley, Co-chair
MARC SNIR
University of Illinois at Urbana-Champaign, Co-chair
WILLIAM J. DALLY
Stanford University
JAMES DEMMEL
University of California, Berkeley
JACK J. DONGARRA
University of Tennessee, Knoxville
KENNETH S. FLAMM
University of Texas at Austin
MARY JANE IRWIN
Pennsylvania State University
CHARLES KOELBEL
Rice University
BUTLER W. LAMPSON
Microsoft Corporation
ROBERT LUCAS
University of Southern California, ISI
PAUL C. MESSINA
Argonne National Laboratory (part-time)
JEFFREY PERLOFF
Department of Agricultural and Resource Economics,
University of California, Berkeley
WILLIAM H. PRESS
Los Alamos National Laboratory
ALBERT J. SEMTNER
Oceanography Department, Naval Postgraduate School
SCOTT STERN
Kellogg School of Management, Northwestern University
SHANKAR SUBRAMANIAM Departments of Bioengineering, Chemistry and
Biochemistry, University of California, San Diego
LAWRENCE C. TARBELL, JR. Technology Futures Office, Eagle Alliance
STEVEN J. WALLACH
Chiaro Networks
6
Scope
• Sponsored by DOE Office of Science and DOE
Advanced Simulation and Computing
• Study Topics:
–
–
–
–
use of supercomputing for sci. and eng. applications
implications for systems design and for the market
needs and the role of the U.S. government
options for progress/recommendations (Final Report)
• Emphasis on “one-machine-room” systems
• Interim (July 2003) and Final (end 2004) Reports
• Input: 4 Meetings; Computational Science
Workshop; Town Hall (SC2003); Site Visits (NSA,
DOE, Japan)
7
Interim Report
http://www7.nationalacademies.org/cstb/
pub_supercomp_int.html
• Summary of earlier reports
• Reviews state of supercomputing
• Identifies issues
– Evolution
– Innovation
– The role of government
• Makes no specific findings or
recommendations
8
Supercomputing Today (Top500 view)
User Countries
Manufacturing Country
1%
1%
7%
13%
1%
1%
2%
3%
4%
50%
7%
USA
Japan
Europe
South-East Asia
8%
91%
11%
U.S. is in pretty good shape!
•
•
Most of the systems are U.S. products
Most of the installations are in the U.S.
But,
•
The Japanese Earth Simulator is on top
9
USA
Germany
Japan
UK
France
Italy
Korea
Canada
China
Other
Modern Supercomputer Structures
Clusters of shared-memory
multiprocessors with an interconnect
(switch)
Varying combinations of off-the-shelf
and custom components
Each combination has its uses
10
Hardware ecology: 3 “Species”
•
•
•
All custom (9%)
– Within TOP500, synonymous with
vector
• NEC Earth Simulator (#1), Cray
X1
Commodity microprocessor, custom
interface & switch (50%)
• IBM ASCI White (#4), SGI
Origin, Sun Fire, Cray T3E
Commodity microprocessor, standard
interface (40%)
– Most use custom (non LAN) switch
– Half use 32-bit processors today
• HP ASCI Q (#2), LLNL Linux
Network (#3), VaTech G5 cluster
•
•
•
•
Best processor performance even
for codes that are not “cache
friendly”
Good communication performance
Simplest programming model
Most expensive
•
•
Good communication performance
Good scalability
•
Best price/performance (for codes
that work well with caches and are
latency tolerant)
More complex programming model
•
Each species has own niche!
11
Supercomputer software
is in bad shape
• Supercomputers are too hard to program
• Software developers have inadequate environments and tools
• Legacy software is difficult to adapt and evolve to newer
platforms and more ambitious problems
Contributing factors
•
•
•
•
Inadequate investment
Few 3rd party supercomputing software vendors
Lack of standards (IO, tools)
Lack of perseverance: e.g., High Performance Fortran
12
The Case for Evolution
• Current platforms do important work
– Need both capability and capacity
• No near-term silver bullet in the offing
– Technology pipeline is fairly empty
• It’s never one size fits all
– Relative ranking of architectures is problem/time/cost
dependent; expect three main species to be around for a
while
• Technology evolves over decades
– X1 inherits from 25 years of Cray and T3D/E
– Clusters inherit from 20+ years of
MPP/NOW/COW/Beowulf research
13
The Case for Evolution (2)
• Commodity based clusters will continue to play an
important role
– Cost/performance advantage (for suitable applications)
– Scalable technology – compatible with technology used in
academic research groups and departments
• Software and application work done on today’s
machines will be adapted to tomorrow’s machines
– Need massive parallelism to scale up on any architecture
• Legacy codes have continuing utility
– Often need to run on old-style architecture
• Uncertainty and inconsistent policies are expensive
– Companies disinvest, R&D teams disappear, researchers
move to greener pastures
14
The Case for Sustained Research
Investment
• Field is not mature: base technology continues fast
evolution
– Non-uniform scaling causes major dislocations (e.g.,
processor vs. memory speeds)
– Supercomputers are early victims of non-uniform scaling
– Solutions require both hardware and software innovation
• New applications challenges abound
– Scaling and coupling
– Massive amounts of data
15
Need Variety of Approaches
• Continuous, steady investments at all stages of
technology pipeline
• Continuous, steady investment in all major
communities
• Mix of small science (individual projects) and
large science (collaborative teams)
• Avoid linear view (successive elimination) and
maximize flow of ideas and people across projects
and concepts
16
Examples of Research Directions
• Architecture
– Better memory architecture (higher bandwidth,
latency hiding)
• Software
– Programming environments and tools
• Applications and algorithms
– Interdisciplinary challenges (e.g., coupling)
17
The Role of Government
Government is main user of highest end machines
Government must ensure that corresponding technology evolves at a rate and
in a direction that serves government missions
Supercomputers are essential to national security
Government must ensure strong supercomputing technology base in the U.S.
Market-based incentives are insufficient
Government needs to support development of supercomputing technology
and supercomputer use in support of science
18
Interim Report Summary
• Supercomputing remains important
• Need balance between custom and
commodity approaches
• Need balance between evolution and
innovation
• Need continuity and sustained investment
• Government role is essential
19
Grand Challenge Questions
Investment
Management
Priorities
Intellectual resources
Validity and quality
Equity
20