Petaflops-scale computing faces severe challenges of cost, size, power, complexity, reliability, efficiency, generality, and programmability. Even with the extraordinary progress of complementary metal oxide semiconductor (CMOS) technology and massively parallel processor (MPP) architectures-systems based on conventional technologies that can sustain throughputs beyond a petaflop--petaflops-scale computing may not be feasible until after 2010.

The hybrid technology multithreaded (HTMT) interdisciplinary research project-supported by DARPA, NASA, NSA, and NSF-is attempting to exploit the superior properties of advanced device technologies to overcome current barriers to success. Such technologies include rapid single flux quantum (RSFQ) superconductor logic, optical communications and holographic storage, and processor-in-memory SRAM and DRAM chips. Superconductor logic is capable of 100 times the speed and power efficiency of conventional processors, while fiber optical communications using time division and wave division multiplexing can exceed wire-based channel bandwidth by a factor of 100 or more. Optical storage density and power efficiency using holographic photorefractive techniques may provide an order of magnitude advantage over today's semiconductor memory at comparable bandwidths. But efficient computation requires effective resource management in the presence of extremes in latency and parallelism. The HTMT architecture has been devised to incorporate adaptive latency-tolerant mechanisms based on multithreaded processors and an innovative memory-based proactive task management scheme called "percolation."

An overview of the HTMT memory architecture.

MVICH is a DOE/National Energy Research Supercomputer Center (NERSC)-funded project to provide portable high performance communications for cluster computing. It is an implementation of MPI for the virtual interface architecture (VIA). VIA is an industry-standard interface for system area networks (networks for clusters) that provides protected zero-copy user-space inter-process communication. MVICH provides a high performance MPI available on a wide range of commodity networks. MVICH is being developed at LBNL and is distributed with an open source license. It is part of the same project that is developing M-VIA, an implementation of VIA for Linux. MVICH implements four protocols to maximize performance over a range of message sizes. Over the next year, the project plans to add:

• Support for unreliable VIA. MVICH will handle dropped and out-of-order packets with negligible additional latency.
• Improved support for asynchronous communication using multiple threads.
• Thread safety. This requires a thread-safe VIA implementation as well as additional support for thread safety in MVICH.

We do not think twice about accessing Web pages that are spread across the globe. The goal of Globus R&D is to bring about a similar revolution in computation. Globus-a joint project of DOE's Argonne National Laboratory (ANL) and the University of Southern California's (USC's) Information Sciences Institute funded by DARPA, DOE, and NSF-is developing the fundamental technologies to build computational grids, execution environments that enable an application to integrate geographically distributed instruments, displays, and computational and information resources. Such computations may link tens or hundreds of these resources.

Globus focuses on:

• Research into basic problems in resource management, security, fault tolerance, and algorithms
• Tools, such as prototype software that can run on a range of platforms
• Large-scale testbeds
• Large-scale grid-enabled applications, developed in collaboration with application scientists

Participants in the Globus Ubiquitous Supercomputing Testbed Organization (GUSTO) are testing Globus concepts on a global scale. GUSTO currently spans more than 40 institutions worldwide and includes some of the largest computers in the world. ANL and USC researchers have won the Global Information Infrastructure Next Generation Award for their development of GUSTO's Globus-based prototype for future computational grids. The GII awards recognize and promote best practices and new models in Internet and network technologies. The Next Generation category targets exemplary uses of the information infrastructure that demonstrate its direction and future potential for cutting-edge information and communications technology applications.

This image was generated by Synthetic Forces Express, a large-scale simulation developed at the California Institute of Technology (CalTech) that uses Globus to provide access to multiple supercomputer resources. SF Express conducted a record-breaking 100,298-vehicle simulation-the largest distributed, interactive battlefield simulation to date--executing it on 1,386 processors distributed over 13 computers at nine sites spanning seven time zones.

Legion-an object-based metasystem software project at the University of Virginia funded by the Department of Defense/Naval Oceanographic Office (DoD/NAVO), NSF, and DOE-will build a system of millions of hosts and trillions of objects tied together with high-speed links. The interface will allow users working at home to access data and physical resources-such as cameras, digital libraries, linear accelerators, physical simulations, and video streams-as if these resources resided on their own disk drives. Groups of users will construct shared virtual workspaces to exchange information and collaborate on research. Legion supports this with transparent scheduling, data management, fault tolerance, site autonomy, and security options, achieving high performance by allowing resource selection among available machines and parallel use of many resources. Legion can be used for parallel processing in a variety of applications and can execute a single application across geographically separate hosts or support meta-applications. Legion supports popular parallel libraries such as MPI and PVM; parallel languages such as MPL; wrapped parallel components so that legacy and new software can be aggregated into Legion objects; and exporting the runtime library interface to library, toolkit, and compiler writers. Legion is an open system whose runtime library is available and for which third-party development is encouraged.

Current and planned experimental applications of Legion include Chemistry at Harvard Molecular Mechanics (CHARMM) for molecular dynamics and mechanics such as protein-folding problems; direct simulation-Monte Carlo (DSMC), used at the University of Virginia to study vapor deposition onto surfaces; and coupled applications such as a federated climate model that ties together many lower-level models.

It is not unusual today for commodity processor chips to perform simple arithmetic operations at speeds ranging from 300 to 600 MHz. These large computational bandwidths, however, are not yet matched by commensurate communication bandwidths between the chips. In fact, the wire interconnections between chips typically operate at only a fraction of the internal computational bandwidth. Optics has the potential to solve the bandwidth interconnect problem between chips, between multichip modules, and at the backplane of boards.

Next generation advanced information processing systems, such as those expected to be used in real-time synthetic aperture radar (SAR) imaging, automatic target recognition, and intensive medical image processing, will require large aggregate computational and communication bandwidths. Since most of these systems are assembled from multiprocessor, memory, and special-purpose digital signal processing chips, the ideal communication bandwidths between them should be on the order of a terabit per second (Tbps). But Tbps bandwidths cannot, at this time, be achieved using conventional wire interconnections. The primary goals of DARPA's VLSI photonics program are to develop and demonstrate the basic technologies that will bring the benefits of optics to chip-scale interconnections.

Theoretical computer science in the 1980s and 1990s has conjectured that a "quantum computer" can solve certain problems asymptotically more rapidly than conventional computers. Such a quantum computer can be thought of as an array of complex two-state systems--for example, the atomic nuclei in molecules with a spin quantum number of 1/2 as is found in hydrogen protons, each of which can store one quantum bit, or qubit, of information. NSA is conducting research into quantum computing using quanta of light as computing elements to demonstrate 1-qubit operations, perform experiments to achieve 2-qubit operations using the optical-lattice method of trapping atoms, and simulate the dynamics of a set of qubits in finer detail than previously achieved. Other NSA-supported projects include experimental research on quantum dots and Josephson junctions as possible qubits and an investigation of individual nuclear spins implanted in a silicon crystal as qubits.

In FY 2000, a consortium of university and Government laboratories is working on a scalable silicon-based nuclear spin quantum computer concept. Other research is under way on the characterization of spin coherence and Rabi oscillations in quantum nanostructures and on the measurement of decoherence times in superconducting qubits.

The goal of this DARPA-funded project at the California Institute of Technology (CalTech), the Massachusetts Institute of Technology (MIT), and USC is to build and operate devices that can store and process information in the quantum states of matter. The objectives are to:

• Develop and explore new ideas for implementing quantum gates in the laboratory
• Broaden the range of problems known to be efficiently solvable using quantum computing
• Design efficient networks of quantum gates to solve computationally difficult problems
• Improve the reliability of a quantum computer that operates under realistic conditions
• Develop a sequential simulator to validate and optimize models and circuits for quantum computers
• Fabricate devices and conduct experimental studies of their performance
• Develop quantum networks for distributed quantum computation and communication
• Explore the applicability of well-tested many-body methods to quantum computation
  
  

• How can the algorithms proposed for quantum computers be implemented efficiently using NMR?
• How can the errors due to imperfect instrumentation and decoherence be prevented, detected, and/or corrected?
• What computational models other than EQC can NMR implement?
• What macroscopic spin orders can be used to practically and efficiently encode information?
• What couplings and feedback among these orders can be most conveniently implemented?
  
  

In FY 2000, NSA-supported smart memory R&D focused on the basic architecture of a computation tile--a building block that must efficiently execute applications exhibiting different types of parallelism. To guide the development, researchers examined how well the machine could emulate existing machine architectures developed for different application spaces.The machines chosen for emulation are the Stanford FLASH, Imagine, hydra, and M-machine.The researchers also investigated architectures for on-chip internconnetion networks, envisioning that future VLSI chips will be constructed with a common interconnection network independent of the function of the chip. The research aims at discovering new interconnection network architectures that best exploit the properties of these emerging on-chip networks.

New initiative

Measurement and calibration for the virtual sciences

The ongoing revolution in computing, communications, and information technologies can transform the foundation for industrial development from physical testing, prototypes, and pilot plants to a new paradigm based on modeling and simulation, where every material or process is conceived, designed, characterized, and optimized using advanced computation and information technology. Yet the methods and tools that allow users to validate, test, and calibrate models of designed materials, complex physical processes, and product performance are missing. Without them, industrial use of these advanced technologies will lag behind technology development.

Starting in FY 2001, NIST will begin a program that will include:

• Development and optimization of broadly applicable methods, algorithms, and associated data to accelerate industrial use of 21st century computing and information technologies for modeling and simulating complex materials and processes
• Development of methods and tools to support the validation, benchmarking, and comparison of models, algorithms, and software used for modeling and simulating materials and processes, including carefully constructed and analyzed open reference implementations, well-characterized test calculations and data, metrics and related tools for evaluating computed results, and testbeds for particular application domains
• Development of consensus standards, definitions, and tools for modeling and simulating information management, software integration, and software interoperability to enhance modeling and simulation in an application environment
• Development of user interfaces for advanced modeling and simulation software and integration with associated data and information resources from diverse sources including Web-based databases and archives