Graphic by Janet Ward of NOAA's High Performance Computing
and Communications Program
Representative FY 2002 agency activities
NSF: Support for acquisition and development of equipment,
instrumentation, and distributed systems for advanced research in
engineering and the sciences, including the terascale computing platform
at the Pittsburgh Supercomputing Center; new device and system architectures,
technologies, and tools to assemble nano-size components into functional
IT structures
DARPA: Prototype data-intensive systems and software; processor in
memory for rapid data access; morphable computing micro-architectures
NASA: Extend base of Beowulf clusters of sovereign workstations to
achieve high- performance computing; research to extend the capability
of single-image supercomputing systems
NIH: Planning for evolution to new architectures and algorithms for
advanced biomedical computation
DOE Office of Science: Extend MVICH system of high-performance communications
for cluster computing at the National Energy Research Supercomputer
Center; develop enabling technology centers for libraries of high-performance
software components for science applications and critical computer
science and software issues in terascale computing systems
NSA: Discovery and application of methods to achieve orders of magnitude
improvement in the computational capability needed to derive intelligence
from mathematical and signal-processing problems. Exploration includes
advanced microscopy, micro spray cooling, optoelectronic circuits,
and optical tape
NIST: Prototype software library of mathematical functions based on
client-server transactions for computational problems such as solving
linear systems and eigenvalue problems
NOAA: Innovative technologies and tools for advanced scalable computation
on highly parallel computing systems to provide greater computing
power at substantially lower cost
ODUSD (S&T): Support for universities' acquisition of equipment
for defense-related research in high-end hardware, software, and applications
|
Critical Federal missions and industry needs both call
for new scientific and technical paradigms that significantly raise
high-end computational speeds, provide adaptable and reconfigurable
computing environments, and reduce the size, cost, and power requirements
of high-performance computing and data storage equipment.
For example, the world's fastest computing platform today is DOE/NNSA's
"Option White" system at the Lawrence Livermore National
Laboratory. A massively parallel system made up of 512 IBM multiprocessor
nodes, it requires 13,000 square feet of floor space and more than
3.2 megawatts of electricity for power, cooling, and mechanical equipment.
Option White is capable of 12.3 teraops (trillions of operations per
second) in processing speed. But even such a system is not adequate
for the massive computational requirements of the most complex scientific
problems, whose solutions are critical to the missions of many Federal
agencies as well as to the Nation.
At the same time, the Nation's high-end computing sector - the companies
that produce computing platforms much faster than the standard desktop
computer - is a shrinking fraction of the U.S. marketplace. Business
purchasers of high-end machines prefer mid-range machines that are
less costly and physically demanding. As a result, the technical challenges
of developing technologies that break through today's upper-end barriers
in computing speed, storage capacity, and equipment are left orphaned.
Federal R&D bridges the gap between the products available commercially
and the requirements of critical government missions, to sustain U.S.
capabilities at the highest levels of computational performance.
Currently, the Government supports several dozen high-end computing
platforms at academic computing centers and national laboratories,
along with a number of mid-range machines, that are used by both academic
and Federal researchers. But these are not nearly enough to support
the high-end research and applications needs of university-based and
government scientists. Nor do they offer a viable model for scaling
up to the processing speeds and storage capacity that future advanced
applications will demand. Today's Option White platform, for example,
has 160 terabytes of storage space spread over 7,000 disk drives.
This amount of storage space can hold about six times the contents
of the entire Library of Congress, but it is only a small fraction
of the scientific data that future research will call for.
Finding cost-effective solutions requires fundamental cross-disciplinary
research in disciplines such as physics, chemistry, materials science,
and electrical engineering, as well as innovations in computer science
and applied mathematics. Next-generation supercomputing architectures,
systems software, and middleware must also address interoperability
needs of both Federal agencies and the private sector. These technological
breakthroughs will also aid U.S. competitiveness.
In FY 2002, the NITRD agencies will proceed with research to increase
the delivered performance of computing systems. The goal is to produce,
by the end of this decade, systems that are capable of 1,000 times
or more the speeds of today's fastest systems, while reducing cost,
energy consumption, and footprint, and to develop interoperable systems
software and tools that will:
- Improve sustained application performance, ease of use,
manageability, and high-speed network connectivity of teraops-scale
systems
- Be scalable (expandable) to petaops-scale systems (petaops
systems perform a thousand trillion calculations per second)
- Provide a unifying environment for high-end scientific
computing
|
The demand for substantial increases in computing capability, to a
level many thousands of times beyond today's systems, will continue
to grow in the years ahead. These increases cannot be attained solely
by isolated enhancements in hardware or software, no matter how dramatic
such individual improvements may be. The architecture of future supercomputers
must consist of components carefully developed, assembled, and tuned,
and must be matched by an application development process that allows
close integration with the system architecture. This ubiquitous close
integration will require substantial breakthroughs in every area of
high-end computing research.
- Advanced computing concepts (including nonconventional
architectures, components, and algorithms)
- Systems software technologies (including operating systems,
programming languages, compilers, memory hierarchies, input/output,
and performance tools)
- Systems architectures that integrate device and component
technologies, systems software, and programming environments
(including device technologies, node functionalities, configuration,
software for managing highly parallel computations, and
hierarchical programming), and network connectivity
- Software component technologies for high performance computing
|
|
