Home > PC >
Originally appeared in
Clusterworx 3.2 Whitepaper
Learn how the comprehensive features of Clusterworx help
organizations lower the total cost of ownership, increase
cluster performance, and provide better return on investment.
- Linux Networx -
[ ] M481601 ) HPC User Forum Explores Leadership Computing...............17.0K
At Oak Ridge National Laboratory this week, 131 HPC User Forum participants
from the U.S. and Europe discussed current examples of leadership
computing and challenges in moving toward petascale computing by the
end of the decade.
Vendor updates were given by Cray Inc., Hewlett Packard Co., Intel Corp.,
Level 5 Networks Inc., Liquid Computing Corp., Panasas Inc., PathScale
Inc., Silicon Graphics Inc. and Voltaire Inc.
According to IDC vice president Earl Joseph, who serves as executive director
of the HPC User Forum, the buying power of users at the meeting exceeded $1
billion. In his update on the technical market, he noted that revenue grew 49
percent during the past two years, reaching $7.25 billion in 2004. Clusters have
redefined pricing for technical servers. The new IDC Balanced Rating tool (<http://www.idc.com/hpc>)
allows users to custom-sort and rank the performance of 2,500 installed
HPC systems on a substantial list of standard benchmarks, including the
HPC Challenge tests.
Paul Muzio, steering committee chairman and vice president of government
programs for Network Computing Services, Inc. and Support Infrastructure
Director of the Army High Performance Computing Research Center, said
the HPC User Forum's overall goal is to promote the use of HPC in industry,
government and academia. This includes addressing important
issues for users.
Jim Roberto, ORNL Deputy for Science and Technology, welcomed participants
to the lab and gave an overview. ORNL is DOE's largest multipurpose
science laboratory, with a $1.05 billion annual budget, 3,900 employees
and 3,000 research guests annually. A $300 million modernization is
in progress. ORNL's new $65 million nanocenter begins operating in
October and complements the lab's neutron scattering capabilities.
Thomas Zacharia, ORNL's associate director for Computing and Computational Sciences,
said computational science will have a profound impact in driving science
forward. ORNL, selected to be the DOE's main facility for Leadership Computing,
plans to grow its machines to 100 teraflops, then to a petaflop by the
close of the decade. Researchers have made fundamental new discoveries
with the help of the Cray X1 and X1E systems. The lab expects to put
its Cray XT3 into production in the October-November timeframe. Based on
estimates from vendors, Zacharia expects a petascale system to have
about 25,000 processors, 200 cabinets and power requirements of 20-40
According to Jack Dongarra, University of Tennessee, the HPC Challenge
benchmark suite stresses not only the processors, but the memory system
and interconnect. The suite describes architectures with a wider range
of metrics that look at spatial and temporal locality within applications.
The goal is for the suite to take no more than twice as long as Linpack
to run. At SC2005, HPCC Awards sponsored by HPCwire and DARPA will
be given in two classes: performance only and productivity (elegant
implementation). Future goals are to reduce execution time, expand
the set to include additional things such as sparse matrix operations,
and develop machine signatures.
Muzio chaired a session on government leadership and partnerships, asking
each speaker to comment on organizational mission, funding and outreach.
Rupak Biswas, from NASA Ames Research Center, reviewed NASA's four mission
directorates and said his organization, which hosts the Columbia system,
has special expertise in shared memory systems.
Cray Henry said the DoD High Performance Computing Modernization Program
(HPCMP) focuses on science and technology for testing and evaluation.
HPCMP wants machines in production within three months of buying
them and uses funds for specific projects, software portfolios (applications
development), partner universities, and the annual technology insertion
process, which expends $40 million to $80 million per year to acquire
large HPC systems for the HPCMP centers. The program works with other
agencies on benchmarking, partners with industry and other defense
agencies on applications development, and maintains academic partnerships.
Steve Meacham said NSF wants input from the HPC community on how best
to develop a census of science drivers for HPC at NSF, and on how
the science community would like to measure performance. NSF's goal
is to create a world-class HPC environment for science. HPC-related
investments are made primarily in science-driven HPC systems, systems software,
and applications for science and engineering research. In 2007, NSF
will launch an effort to develop at least one petascale system by
2010 and invites proposals from any organization with the ability
to deploy systems on this scale.
Gary Wohl explained NOAA is a purely operational shop that does numerical weather
prediction and short-term numerical climate prediction. The primary HPC goal
is reliability for on-time NOAA products. NCEP and IBM share responsibility
for 99 percent on-time product generation. Changes in the HPC landscape include
greater stress on reliability, a dearth of facility choices, and burgeoning
In the ensuing panel discussion, participants stressed that the federal government
needs to recognize HPC as a national asset and a strategic priority. Non-U.S.
panelists echoed the message.
Suzy Tichenor, vice president of the Council on Competitiveness, showed
a video produced in collaboration with DreamWorks Animation to explain
and excite non-technical people about HPC. Meeting attendees applauded
the video, which can be ordered at www.compete.org <http://www.compete.org/>.
Tichenor reviewed the Council's HPC Project and its surveys that
found, among other things, that HPC is essential to urvival for U.S.
businesses that exploit it.
DARPA's Robert Graybill updated attendees on the HPCS program, noting
Japan plans to develop a petascale computer by 2010-2011 that will
have a heterogeneous architecture (vector/scalar/MD).
In related presentations, Michael Resch of HLRS, Michael Heib from T-Systems
and Joerg Stadler of NEC described their successful partnership in
Germany, which includes a joint venture company to buy and sell CPU
time and the innovative Teraflop Workbench Project, whose goal is
to sustain teraflop performance on 15 selected applications.
Sharan Kalwani from General Motors reviewed the auto maker's business transformation,
noting that GM is involved with one of every six cars in the world.
Today, GM can predict how much compute time and money it will need to develop
a new car. Senior management is convinced about the value of HPC,
David Torgersen's role is to bring shared IT infrastructures to Pfizer. Challenges
include vendors selling directly to business units for point solutions
that don't reflect the company's needs; differing business needs at various
points in the drug development process; and the fact that grid technology
is mature in some respects, not in others.
Jack Wells of ORNL, Thomas Hauser of Utah State, Jim Taft of NASA and
Dean Hutchings of Linux Networx explored possibilities for partnering to
boost the performance of the Overflow code on clusters. They explained
why none of their organizations would do this on its own, then reviewed
the challenges and potential next steps.
Jill Feblowitz of IDC's Energy Insights group said the financial health
of the utility industry has been slowly improving since Enron. In contrast,
the oil and gas industries have had a run-up in profits, although
these profits have not yet translated into an increased appetite for
technology and investments. The Energy Policy Act of 2005 specifically
includes HPC provisions for DOE. She described the concepts of "the
digital oilfield" and the Intelligent Grid.
Marie-Christine Sawley, director of the Swiss National Supercomputer Center (CSCS),
described her organization and its successful, pioneering use of the HPC Challenge
benchmarks in the recent procurement of a large-scale (5.7 teraflops) HPC
system in conjunction with Switzerland's Paul Scherrer Institute.
Thomas Schulthess reviewed ORNL's material science work on superconductivity, which
has revolutionary implications for electricity generation and transmission.
Two decades after the discovery of higher-temperature superconductors,
they still remain poorly understood. Using quantum Monte Carlo techniques,
the team of ORNL users explicitly showed for the first time that superconductivity
is accurately described in the 2D Hubbard model.
Bill Kramer said NERSC focuses on capability computing, with 70 percent
of its time going to jobs of 512 processors or larger. NERSC has won numerous
awards for its achievements in the DOE's INCITE and predecessor "Big
Splash" programs. In the related panel discussion, participants from
industry, government and academia stressed the need for better algorithms
Frank Williams from ARSC is chair of the Coalition for the Advancement
of Scientific Computation, whose members represent 42 centers in
28 states. CASC disseminates information about HPC and communications and
works to affect the national investment in computational science and
engineering on behalf of all academic centers. Williams invited HPC
User Forum participants to attend a CASC meeting and to contact him
IDC's Addison Snell moderated a panel discussion on leadership computing
in academia. HPC leaders from the University of Cincinnati, Manchester
University (UK), ICM/Warsaw University and Virginia Tech discussed
their organizations, leadership computing achievements and the challenges
of moving toward petascale computing. In another panel discussion
moderated by Snell, HPC vendors debated the issues with cluster data management
and what needs to be done to improve the handling of data in large HPC
Phil Kuekes of HP gave a talk on molecular electronics and nanotechnology
("nanoelectronics"), summarizing HP's progress toward developing a
nanoscale switch with the potential to overcome the limitations of existing
IC circuit technology.
Muzio and Joseph facilitated a session on architectural challenges in moving
toward petascale computing. According to Muzio, an application engineer's
ideal petascale system would have a single processor, uniform memory,
Fortran or C, Unix, a fast compiler, an exact debugger and the stability
to enable applications growth over time. By contrast, a computer scientist's
ideal petascale system would have tens of thousands of processors,
different kinds of processors, non-uniform memory, C++ or Java, innovative
architecture and radically new programming languages.
"Unfortunately, for many users, the computer scientist's system may
be built in the near future," he said. "The challenge is
to build this kind of system but make it look like the kind the applications
software engineer wants."
According to Robert Panoff of the Shodor Educational Foundation, math
and science is more about pattern recognition and characterization than
mere symbol manipulation. He said the lag time between discoveries
and their application.
"The people who will use petascale computers are now in high school
to grad school, while most of us are approaching retirement," he
added. "You don't need petascale computing for this teaching,
but this will help produce the people needed to do petascale computing."
David Probst of Concordia University argued scaling to petaflop capability
cannot be done without embracing heterogeneity. Global bandwidth is the
most critical and expensive system resource, so he said we need to
use it well throughout each and every computation. "Heterogeneity
is a precondition for this in the face of application diversity, including
diversity within a single application," Probst added. "Every
petascale application is a dynamic, loosely coupled mix of high thread-state,
temporally local, long-distance computing and low thread- state,
spatially local, short-distance computing."
Burton Smith, chief scientist at Cray, challenged the popular definitions
of "petascale," "scale" and "local." The
popular definition of scale "doesn't mean much, maybe that I
ran it on a few systems and it seemed to go fast," he said. "You probably
mean it message-passes with big messages that don't happen very often. Also,
people say 'parallel' when they mean 'local.'" He concluded that parallel computing
is just becoming important; we know how to build good petascale systems
if the money is there; and sloppy language interferes with our ability to
According to Michael Resch of HLRS, there needs to be a "move on
from MPI to a real programming language or model. I hear people complaining
about how hard it is to program systems with large numbers of processors.
What about buying systems with a smaller number of more-powerful processors?
Why not buy high-quality systems?"
Muzio introduced the companion panel discussion on "software issues
in moving toward petascale computing" by reviewing the HPC User
Forum's achievements in promoting better benchmarks and underscoring
the limited scalability and capabilities of ISV application software.
Suzy Tichenor reviewed the Council on Competitiveness' recent "Study
of ISVs Serving the HPC Market: The Need For Better Application Software." The
study found the business model for HPC-specific application software
has evaporated, leaving most applications unable to scale well. Market
forces alone will not address this problem and need to be supplemented
with external funding and expertise. Most ISVs are willing to partner
with other organizations to accelerate progress.
DARPA's Robert Graybill said the HPCS program is looking at how to measure
productivity, and that he believes new programming languages are needed. We
need time to experiment before deciding which are the right HPC language attributes.
The goal by 2008 is to put together an industry consortium to pursue
this. I/O is another major challenge.
BAE Systems' Steve Finn, chair of the HPCMP's User Advocacy Group, said continuous
improvements are still occurring to legacy codes and large investments
have been made in scalable codes. "We need to prioritize which codes
to rewrite first [for petascale systems]," he added. "UPC and
CAF won't be the final languages. It's good to try them out, but
if you rewrite them now, you may need to rewrite them again in a few
The next HPC User Forum meeting will take place April 10-12, 2006 in Richmond, Va.
The meetings are co-sponsored by HPCwire.
[ ] A740 ) SC|05 Registration is Now Open
To Both Technical and Exhibitor Registrants!
[ ] A366 ) GRIDtoday
DON'T MISS THESE FEATURE STORIES IN THIS WEEK'S ISSUE!
iGrid Chair DeFanti: 'Not at the End of the Rainbow Yet'
DataSynapse Implements GridServer for BNP Paribas
[ ] A785 ) Hurricane Katrina Relief - Make a Contribution.
We've Found More Than a Dozen Agencies
That Are Providing Assistance to the
Full background information on all leading HPC solution providers
[ ] A517 ) Data Direct
[ ] A639 ) Microway
[ ] A351 ) APPRO
[ ] A330 ) Atipa
[ ] A518 ) Cray
[ ] A373 ) Dell
[ ] A332 ) Etnus
[ ] A470 ) HP
[ ] A335 ) IBM
[ ] A331 ) Intel
[ ] A337 ) Linux Networx
[ ] A336 ) Microsoft
[ ] A345 ) Myricom
[ ] A341 ) NEC
[ ] A338 ) PathScale
[ ] A812 ) Penguin
[ ] A343 ) Portland Group
[ ] A346 ) Quadrics
[ ] A339 ) SGI
[ ] A342 ) Sun
[ ] A471 ) Voltaire
For sponsorship information contact: firstname.lastname@example.org
Copyright 1993-2005, HPCwire. All Rights Reserved.
Mirrored with permission.