Objective

Develop detailed architectural strategies for achieving petaFLOPS capable systems as a means to guide future research programs.

Approach

Sponsor a set of point design studies in coordination with NSF and DARPA for a duration of six months. These studies are to propose and examine innovative approaches to achieving petaFLOPS systems in the 2007 time frame and to consider their implications for system software. Two or more applications are to be studied with respect to each architecture approach.

Accomplishments

NASA has co-sponsored eight point design studies with NSF and DARPA. These were formally described at the PetaFLOPS Architecture Workshop (PAWS '96) in April, which was chaired by T. Sterling, then of USRA CESDIS at Goddard Space Flight Center. One of the point design study concepts was developed by Sterling at CESDIS/GSFC and submitted in collaboration with Messina of Caltech, Likharev of SUNY, and Gao of Delaware. The concept was to combine advanced superconductor and optical technologies with multi-threaded architectural flow control mechanisms. A second proposal to do a study of a Processor-in-Memory architecture by P. Kogge of Notre Dame was also accepted. This work had early sponsorship by GSFC under its petaFLOPS project in FY96. The eight funded proposals include:

• MORPH: A Flexible Architecture for Executing Component Software at 100 TeraOPS
  (by Andrew A. Chien, Rajesh K. Gupta, University of Illinois at Urbana-Champaign)
• Pursuing a Petaflop: Point Designs for 100 TF Computers Using PIM Technologies
  (by Peter M. Kogge, Steven C. Bass, Jay B. Brockman, Danny Z. Chen, Edwin Hsing-Mean Sha, University of Notre Dame)
• Architecture, Algorithms and Applications for Future Generation Supercomputers
  (by Vipin Kumar, Ahmed Sameh, University of Minnesota)
• Design Studies on Petaflops Special-Purpose Hardware for Astrophysical Particle Simulations
  (by Stephen L.W. McMillan, Drexel University; Piet Hut, Institute for Advanced Study, Princeton; Junichiro Makino, University of Tokyo; Michael L. Norman, University of Illinois at Urbana-Champaign; Frank J. Summers, Princeton University)
• Hierarchical Processors-and-Memory Architecture for High Performance Computing
  (by Jose A.B. Fortes, Rudolph Eigenmann, Purdue University; Valerie Taylor, Northwestern University)
• The Illinois Aggressive Cache-Only Memory Architecture Multiprocessor - (I-ACOMA)
  (by Josep Torrellas, David Padua, University of Illinois at Urbana-Champaign)
• A Scalable-Feasible Parallel Computer Implementing Electronic and Optical Interconnections for 156 TeraOPS Minimum Performance
  (by Sotirios G. Ziavras, Haim Grebel, New Jersey Institute of Technology; Anthony T. Chronopoulos, Wayne State University)
• Hybrid Technology Multi-Threaded Architecture
  (by G. Gao, K. Likharev, P. Messina, T. Sterling)
  • Proposal Abstract
  • Extended Abstract of Research Study Proposal

Significance

The point design studies effort is the first substantive attempt by the community to provide in-depth understanding of the potential opportunities for developing petaFLOPS computers. At PAWS they were examined from the perspectives of architecture, applications, and system software. NASA's contribution, including that of GSFC, has been instrumental in organizing, participating in, and documenting the workshop and its findings, including those derived from the point design study analyses. As a result of this work, the community addressing the petaFLOPS question has identified four core archetypal architecture classes: PIM, hybrid technology, clustered COTS-based systems, and special-purpose devices. Of these, as noted above, the first two have been directly influenced by the GSFC petaFLOPS project.

Status/Plans

The final results of the point design studies are to be presented at the Frontiers '96 conference in Annapolis, MD, in October. This conference series, now in its 10th year, was initiated by GSFC and is in part sponsored by the GSFC HPCC program. Representitives of the studies will participate in The 2nd PetaFLOPS Frontier Workshop (TPF-2), which was also begun by GSFC and CESDIS. Future activities by GSFC researchers are planned to contribute to petaFLOPS in the area of applications studies to enhance the analysis of various architecture approaches.

Point of Contact

Thomas Sterling
Jet Propulsion Laboratory
tron@cacr.caltech.edu