UD engineers partner with IBM in supercomputing consortium

3:30 p.m., Oct. 16, 2003--A University of Delaware research team is participating in a national initiative to regain preeminence in supercomputing, a position that has been claimed by the Japanese over the course of the last decade.

Guang R. Gao (left), director of UD's high-performance Computer Architectures and Parallel Systems Laboratory, works with doctoral student Juan del Cuvillo on data for the new supercomputing project.

The UD team is led by Guang R. Gao, professor of electrical and computer engineering and director of the high-performance Computer Architectures and Parallel Systems Laboratory. The team is part of a consortium working with IBM on the project, which is funded by the U.S. Defense Advanced Research Projects Agency (DARPA).

In early July, DARPA announced that it had selected three high productivity computing systems projects to continue into the second of three phases. In addition to IBM, the agency selected a joint venture of Cray and New Technology Endeavors and also Sun Microsystems.

The goal of the federal program is to provide a new generation of economically viable, scalable, high productivity computing systems for use in national security and industry by 2010.

“Although the United States pioneered the supercomputer, which is critical to many industries including those involved in national security, Japan has taken the lead in the design and production of supercomputers over the last decade,” Gonzalo Arce, chairperson of UD’s Department of Electrical and Computer Engineering, said. “The U.S. has launched an important initiative to regain the lead and is investing significant resources in industry-academic partnerships to develop a petaflop supercomputer.”

A petaflop is a measure of a computer’s processing speed and can be defined as a thousand trillion flops, or floating point operations per second.

Falling behind the Japanese has been a “wake up call,” Gao said, adding that the result of the DARPA project will be “a supercomputing architecture roadmap” that will provide very high speed, high performance computing to meet a variety of needs.

Gao said UD is joined on the IBM team by a number of other universities, including the Massachusetts Institute of Technology, Cornell University, the University of Illinois at Champaign-Urbana, the University of California at Berkeley and the University of Texas at Austin.

UD was selected for the project “because of our pretty unique long-term work in the area of programming models for communication and data management within large computer systems, some involving hundreds of thousands of individual processors,” Gao said.

In such a computer system, communications between the individual processors is a major problem, Gao said, because so much information is being input, digested, shared and ultimately output.

Through earlier work, Gao’s laboratory established that such systems work best through decentralized rather than centralized control. He explained the finding by comparing it to a political system. In a centralized system, a scientist asked for information must relay the request through many layers of supervisors before being allowed to provide an answer. In a more open system, the scientist can decide on his or her own to respond, thus speeding the process immensely.

“In a large computer system, it is our belief that this has to happen,” Gao said.

To make it happen, the team must write special programs that take into account threading, or the execution of individual tasks, data movement and a consistent way to deal with new information being fed into the system.

Much as manufacturers have gone to just-in-time delivery of parts, Gao said the UD team is seeking to provide for just-in-time delivery of data to improve performance. Information will arrive when needed and so not clog up the system.

According to DARPA, the program will fill a gap in high-end computing that the Department of Defense will experience as it moves from today's high performance computing technology, which dates from the late 1980s, to the future promise of quantum computing.

IBM received $53.3 million for its Productive, Easy-to-use, Reliable Computing Systems (PERCS) approach. IBM is adapting the system layers to application requirements. Adaptability enhances the technical efficiency of the system, its ease of use, and its commercial viability by accommodating a large set of commercial and high performance computing workloads.

During the second phase of the program, IBM will conduct research that will push the state of the art beyond the evolutionary path, with aggressive goals in performance, usability and time to solution.

Cray and New Technology Endeavors received $43.1 million to develop their concept, Cascade, which incorporates hardware and software technology innovations to enable higher productivity for sustained peta-scale computing.

Hardware innovations include new processor architectures to make more effective use of the memory hierarchy; processor-in-memory technologies to provide greater memory bandwidth; and high bandwidth, low-latency, scalable networks. Software innovations include support for scalable shared and distributed memory programming models to facilitate rapid development of new applications while maintaining high performance and portability of existing applications.

Sun Microsystems received $49.7 million to continue work on its integrated system approach known as Hero, which provides a simplified architecture and novel programming tools that boost user productivity, enhance numerical precision, increase system security and support legacy software. The architecture makes quadrillions of calculations per second readily available to programmers.

The Hero project is also providing a “purpose-based benchmarks” approach as a means of measuring and designing systems for the productivity and performance that are the primary goals of the DARPA program.

The DARPA program is in three phases, the first of which was a 12-month, industry-guided concept study that provided critical technology assessments, developed revolutionary concept solutions and generated new productivity metrics.

The second phase of the program is a 36-month research and development stage that will perform focused research and development and risk reduction engineering activities and will culminate in technology risk reduction demonstrations and a preliminary design review for each system.

The third phase is a 48-month, full-scale engineering development effort.

For more information, visit [www.research.ibm.com/resources/news/20030710_darpa.shtml].

Article by Neil Thomas
Photo by Eric Crossan

E-mail this article

To learn how to subscribe to UDaily, click here.