CRA Testimony on NSF Advanced Computational Infrastructure


given by Mary Vernon, Professor, Department of Computer Science,
University of Wisconsin–Madison

before the Subcommittee on Basic Research,
Committee on Science, U.S. House of Representatives

March 19, 1996


Mr. Chairman and members of the Subcommittee:

Thank you for the opportunity to testify on the National Science Foundation program, Partnership for Advanced Computational Infrastructure.

My name is Mary Vernon and I am a Professor of Computer Science with a joint appointment in Industrial Engineering at the University of Wisconsin at Madison.

I am the Project Director of the Computer Science Department's Laboratory for Research in Next-Generation High-Performance Computing Systems. My research involves developing models to estimate the performance impact of proposed design changes for high-performance computers.

I also am a member of the Board of Directors of the Computing Research Association (CRA), a scientific society whose membership consists of nearly all the graduate departments of computer science and computer engineering in the United States; affiliated professional societies; and several U.S. industrial research laboratories that perform basic research in computing.

My knowledge of the NSF Supercomputer Centers comes from collaborations related to my research, as well as from my participation on two recent panels that have evaluated the program:

1) The 1993 NSF Blue Ribbon Panel on High-Performance Computing, chaired by Lewis Branscomb.

2) The 1995 NSF Task Force, chaired by Edward Hayes.

Today I would like to explain why I strongly endorse both the restructuring of the program as proposed in the Task Force report, and the planned recompetition process as described in the NSF Program Solicitation. I will make six key points.

1. The accomplishments of the NSF Supercomputing Centers over the past 10 years have been tremendous

The Centers have provided computing capabilities significantly beyond what's possible at individual research universities. For example, my Laboratory at Wisconsin has a 64-node CM-5 system. This is very useful for code development and testing, and for small-scale experimentation with new algorithms and system design issues. However, we rely on the 512-node CM-5 at the National Center for Supercomputing Applications (NCSA) for larger-scale experimentation.

Similarly, another group in my Department has recently acquired an 8-node IBM SP/2 system and will be relying on the 512-node SP/2 at the Cornell Theory Center for large-scale experimentation. These are just two examples of how the Centers are providing leading-edge capabilities that are critical to science and engineering research.

The Centers' capabilities have enabled fundamental advances in science and engineering, as documented in Appendix B of the Task Force report. I was unaware of the extent of the accomplishments in other fields prior to my participation in the Branscomb panel, but I quickly learned that the program is serving the nation and the broader science and engineering community extremely well. The vision that created the program was excellent.

A survey of the science and engineering research community in Appendix G of the Task Force report shows that for large numbers of researchers in all disciplines, the Centers have been indispensable.

For example, the Centers provide an indispensable experimental test bed for computer scientists designing the next generation of high-performance programming languages and systems.

The Centers provide technical support that is beyond what an individual university can provide. At Wisconsin, we have only one staff member assigned to our CM-5. User support is largely nonexistent. This is acceptable for researchers who are investigating next-generation computing systems. However, it is inadequate for researchers whose goals are advances in other areas of science and engineering.

2. The NSF Centers Program has been thoroughly studied by a broad spectrum of science and engineering researchers—both those who depend greatly on Center resources and those who might envision other productive uses for the funding. There is widespread agreement concerning the need for a future program as recommended in the Task Force report.

The scientific goals of the Centers Program are paramount. The 1993 NSF Blue Ribbon Panel, chaired by Lewis Branscomb, had representation from applied mathematics; computer science; physics; chemistry; biology; and civil, chemical and environmental engineering.

The 1995 Task Force had similarly broad representation, as explained in Edward Hayes' testimony.

Both reports came to similar conclusions about the success of the Centers program over the past 10 years, and the importance of continuing to provide access to leading-edge capabilities for scientists and engineers in the next decade.

3. The Program will be strengthened by the proposed restructuring to include Leading-Edge sites that are strongly partnered with experimental facilities and research centers at other universities, as well as other national or regional high-performance computing centers.

In the past, Leading-Edge Centers have learned about new high-performance computing technology primarily from vendors and informal relationships with NSF research centers.

Formal partnering with experimental facilities and research centers at other universities will expand the Center capabilities for earlier and broader experimentation with new high-performance technologies.

• For example, there are a number of computer science groups—at the University of California at Berkeley, Wisconsin, the University of Washington and other places—that are experimenting with Network of Workstation (NOW) technologies that could eventually support production of high-performance computing applications. Partnerships with such groups can enable the NSF Leading Edge Centers to experiment with, anticipate and prepare for next-generation technologies more thoroughly than has previously been possible.

• For the same reason, formal partnerships with research groups developing next-generation computational algorithms and techniques will also be important.

Other partnerships that include coordinated planning of resource allocation with university, regional or national high-performance computing centers will enable more cost-effective use of the nation's computing infrastructure.

4. The requirement that Leading-Edge sites provide balanced computing capabilities that are one or two orders of magnitude beyond what is available at leading research universities, together with recent trends in technology and the computer industry, make it both possible and necessary to reduce the number of Leading-Edge sites.

In the past, vector supercomputers were a reasonably stable technology. A single system could be considered “leading edge,” in terms of computing power and computing technology over a period of five years.

Due to introduction of parallel supercomputers, which have a much greater pace of evolution, leading-edge systems change significantly on a three-year time frame.

The current Centers Program has required an influx of funding from other sources in order to keep up with recent advances in high-performance systems. Even so, Centers are financially limited in how quickly they can purchase next-generation systems. Science and engineering advances would benefit if a more rapid upgrade path at the Leading-Edge sites could be achieved.

In addition to keeping pace with advancing high-performance computer technology, it is important that the Leading-Edge Centers provide balanced systems. As processors become faster, this means purchasing larger amounts of memory and more expensive disk subsystems for each machine. Furthermore, there are significant advantages to aggregating both computing power and memory at fewer sites, as this enables the solution of even larger and more complex physical system models. The aggregation of computing power and memory is becoming more important for leading-edge computational science and engineering than the ability to access machines from different vendors.

Thus, trends in computer architecture and the need for the largest-scale balanced systems each argue compellingly for a smaller number of centers, each funded at a higher level. This point is well understood by NSF and the computational science and engineering community, but is perhaps not thoroughly explained in the Task Force report.

5. Recompeting the Centers is the best approach to achieving the most effective Leading-Edge Sites and Partnerships.

As the Task Force began its work, the group as a whole, and I in particular, started out with a healthy level of concern about the potential risks associated with recompetition of the existing centers. These concerns were expressed in the 1993 NSF Blue Ribbon Panel Report. Two of the Task Force members had served on the Panel that produced that report.

There are strong arguments against recompeting the Centers; most importantly the invaluable expertise at the existing Centers and the need for stability in order to maintain that expertise. At least one former Center staff member commented in response to the survey in Appendix G of the Task Force report that “a primary reason for my leaving the centers program was an unease about NSF's commitment to the program.”

However, from 1993 to 1995 we noted several key developments that changed the balance in favor of a new competition. In addition to the usual advantages of recompetition such as the degree of creativity and renewed commitment that a Program Solicitation invites, the Task Force concluded that there had been important changes in the technology that called for a new structure for the centers—a structure based on partnerships between leading-edge sites and other partnership centers with significant technical expertise and, importantly, midlevel computing capability effectively integrated into the “partnership center.” As a result, the Task Force concluded that a whole new competition was needed. Such a change is best accomplished by an open competition and merit review. In the end, the likely increased overall strength of the responses to the competitive Solicitation, outweighs the significant stability issues.

6. Members of the broad science and engineering community, following the merit review process and criteria outlined in the new Program Solicitation, will make wise choices for the nation.

It is likely that such a broad spectrum of expertise has never been brought to bear on a decision. The review panel members will fully understand the intricacies of the new directions and will be best able to judge the number and combination of proposed Partnership sites that will best serve computational science and engineering, and the national interest, in the next decade.

Summary

In summary, Mr. Chairman, I strongly support the restructuring of the NSF Supercomputer Centers Program to create Partnerships for Advanced Computational Infrastructure. These will provide leading-edge computational facilities as well as coordinated resource allocation and experimentation with new technologies on midrange systems. The partnerships should be funded at a level that allows the leading-edge systems to keep pace with changes in high-performance computing technologies and to provide balanced processor, memory, and input/output capabilities that are one to two orders of magnitude beyond what will be available at major research universities. The planned NSF proposal review process is the appropriate mechanism for determining the specific number and types of partnerships that will best meet the needs of computational science and engineering in the next decade.

Thank you for this opportunity to testify.

Document last modified on .