The goal of achieving sustained TeraFLOPS computing by the end of the decade is the focus of the Federal HPCC program. While many challenges to realizing efficient and cost-effective systems and techniques still exist, the general approach has been prescribed. For example, it is clear that in this time frame, ensembles of between 1,000 to 10,000 superscalar microprocessors will be implemented to deliver peak performance at the TeraFLOPS scale. Software methodologies to support massively parallel processing systems are still in advanced development, but general approaches are well founded although important details have yet to be worked out. The HPC community can anticipate that, given continued investment in the necessary R&D, the HPCC program will achieve its performance goals in the allotted time.
Given the progress of the HPCC program toward its goal, it is time for industry, academia, and government to begin addressing the more daunting challenge of positioning future long-term research programs to develop enabling technologies for PetaFLOPS computing systems. The HPCC program was enacted to confirm U.S. competitiveness in the world's high technology market place; to sustain preeminence will demand that the U.S. HPC community move aggressively beyond the TeraFLOPS milestone to tackle the staggering goal of realizing and harnessing PetaFLOPS capability.
The achievement of GigaFLOPS performance demanded a paradigm shift from previous conventional computing methods in architecture, technology, and software. Vector processing, integrated high-speed devices, vectorizing languages and compilers, and dense packaging techniques were incorporated into a single supercomputer model culminating in such a tour de force as the C-90. Similarly, the achievement of TeraFLOPS performance, which the community is actively pursuing, is demanding a second paradigm shift exploiting VLSI technology, massively parallel processing architectures, as well as messaging-passing and data-parallel programming models. The first of these advances (vector processing) came from industry with support from government in their application by research and applied industrial concerns. The second of these advances (Massively Parallel Processing) is requiring a much larger cooperative initiative harnessing the capabilities and talents of all elements of the HPC community. The first advance occurred largely spontaneously from the creativity and initiative of the community; the second required years of planning prior.
Drawing upon this history to anticipate the nature of the challenge imposed by goals of effective PetaFLOPS computation, it can be assumed that the next revolutionary step from TeraFLOPS to PetaFLOPS capability will be achieved only through yet another multifaceted paradigm shift in architecture, technology, and software methods. Concepts and capabilities not usually associated with mainline HPC community may prove key to the new paradigm. Further, the cooperation and integration of our national resources, even beyond that exemplified throughout the HPCC program, will be critical to success. Coordination and direction of research toward these goals will require substantial planning drawing upon expertise and insights of professionals from all parts of the high-performance computing community. Even as the HPCC program is underway, initial planning for the next step must be undertaken so that upon successful conclusion of the HPCC program the government will be prepared to maintain the momentum and redirect it toward the follow-on goals.
Early recognition of the importance of this issue has come from the NASA Administrator. At his request, a panel was convened to consider the domain of issues expected to significantly impact progress toward PetaFLOPS computing systems. The Supercomputing Special Initiative Team included key personnel from NASA Headquarters and HPCC centers to, ``evaluate whether the U.S. high-performance computing community is aggressively supporting NASA's existing and future computing requirements,'' with the objective of recommending, ``program or policy changes to better support NASA's [high-performance computing] interests.'' The panel explored NASA HPC requirements over the next decade, examined current government supported efforts in HPC, evaluated efforts by the U.S. supercomputing industry and those overseas, and formulated recommendations to the NASA Administrator.
The NASA Supercomputing Special Initiative Team concluded that NASA mission requirements will demand PetaFLOPS capability, but there are major barriers to achieving it. The team also observed that NASA alone could not exert sufficient influence on developments in many of the areas that may contribute to the elimination of these barriers. The team recommended that NASA join with other agencies to sponsor architectural studies of PetaFLOPS computing systems. Cooperation of this type would ensure that mission requirements of all Federal agencies affect the priorities of technology development initiatives toward a PetaFLOPS capability.
This report discusses an important first step in fulfilling the recommendations of the NASA Supercomputing Special Initiative Team to the NASA Administrator. At the first annual comprehensive review of the NASA HPCC Program, it was determined that in cooperation with other agencies NASA should sponsor in the near future a focused in-depth workshop on enabling technologies for PetaFLOPS computing. The purpose of such a workshop would be to determine research directions critical to achieving a PetaFLOPS capability. While this general topic has been considered in other forums such as the Purdue Workshop on Grand Challenges in Computer Architecture for Support of High Performance Computing, it has not been addressed from a perspective that includes Federal agency mission requirements, nor from a view to initiating near-term research programs. Results from the Purdue workshop and various conference panels were first steps in this direction but have not yielded sufficient detailed findings to provide specific research direction. It is important to note that with TeraFLOPS capability still a major scientific and engineering challenge, there was a danger of a PetaFLOPS workshop appearing superfluous to the community. Therefore, it would be crucial that objectives and methodology be well defined so that the workshop could garner the respect of the NASA and HPC communities by providing the first detailed examination of the issues relevant to achieving PetaFLOPS computing.
In determining the scope of the challenge, the workshop was to delineate clearly the limitations of conventional and expected near-term approaches to implementing and applying PetaFLOPS-scaled systems. A number of exotic technologies, architectures, and methodologies have been pursued in academia and industry. These laboratory explorations must be examined for their potential in achieving PetaFLOPS capability. Most promising avenues of pursuit should be characterized in terms of the pivotal technical issues restraining their advance to determine dominant research questions in the field. From this evolved perspective, the workshop would provide a general long-term approach to research in PetaFLOPS computing, including alternative paths to be considered and their interplay with world-wide integrated information and computing resources. Finally, the workshop would deliver specific and detailed recommendations for immediate actions in sponsoring new research initiatives with the primary intent of dramatically narrowing the uncertainty in this field.