National Aeronautics and Space Administration Ames Research Center HPCC Program Office Moffett Field, Calif. 94035

For Release:
November 15, 1996

Jarrett Cohen Goddard Space Flight Center, Greenbelt, Md. (301/286-2744, jarrett.cohen@gsfc.nasa.gov)

During Supercomputing '96: Pittsburgh Hilton & Towers (412/391-4600)

NASA-INSPIRED PARALLEL WORKSTATIONS EXCEED SUSTAINED GIGAFLOPS PERFORMANCE AT JPL/CALTECH AND LOS ALAMOS NATIONAL LABORATORY

PITTSBURGH -- Two parallel workstations based on NASA's Beowulf "pile of PCs" concept have each exceeded one gigaFLOPS performance for a price of approximately $50,000.

One system was sponsored by NASA Headquarters' Office of Space Sciences and built by the NASA Jet Propulsion Laboratory (JPL) and the California Institute of Technology. The other was funded and is operated by the Theoretical Division at Los Alamos National Laboratory.

The newest versions of the Beowulf cluster, linking 16 Intel Pentium Pro processors with Fast Ethernet networks, attained over one gigaFLOPS sustained on a cosmological N-body simulation. The architecture was conceived by the Center of Excellence in Space Data and Information Sciences (USRA CESDIS) at NASA Goddard Space Flight Center. The NASA High Performance Computing and Communications (HPCC) Program's Earth and Space Sciences (ESS) Project funds continuing development.

Supercomputing '96 attendees will be able to inspect the systems at the Caltech Center for Advanced Computing Research (#R62) and Los Alamos (#R86) research exhibits November 18­21. Tied together with 16 additional Ethernet channels, the two Beowulfs will work in concert to realize greater than 2 gigaFLOPS on several demonstrations.

The JPL/Caltech Beowulf, which ran the N-body calculation at 1.26 gigaFLOPS, was built in collaboration with CESDIS researchers. It consists of 16 Pentium Pro (200 MHz) processors connected through a 100 megabits-per-second Fast Ethernet switch. The system has 2 gigabytes of distributed memory, a theoretical peak speed of 3.2 gigaFLOPS and 80 gigabytes of disk storage.

Michael Warren of the Theoretical Division at Los Alamos constructed a similar machine that also relies on 16 Pentium Pro processors but contains five Fast Ethernet interfaces per processor. This system achieved 1.17 gigaFLOPS on the N-body code, which was written by Warren and Caltech's John Salmon.

"Using commodity personal computer subsystems allows supercomputer performance at a significantly reduced cost," said Thomas Sterling, a senior scientist at Caltech and JPL who led the original design team. "Any college or university, or laboratory department, can now afford a parallel supercomputer for research and education."

In addition, "Beowulf has a larger memory and much larger disk storage than commercially available workstations in the same price range," Sterling said. "Together with processor speed these qualities provide a robust platform for applications with large datasets, such as in the Earth and space sciences."

The benchmark used by Warren and Salmon to measure the performance is a highly optimized N-body "treecode" simulating the gravitational interactions of 10 million bodies. The fastest overall implementations of the code are on 512 nodes of the Los Alamos Thinking Machines CM-5 (14.06 gigaFLOPS) and the Caltech Intel Paragon (13.70 gigaFLOPS).

Sterling summarized the significance of the Beowulf results: "We have entered a new era in which mass market commercial computing products can be harnessed for large-scale scientific computation, greatly reducing the end user cost and allowing more researchers to do more and better computational science."

Beowulf offers a sophisticated software infrastructure through an enhanced Linux operating system. Linux provides Unix functionality on systems using Intel, Sun and DEC Alpha processors and is widely available at no cost. CESDIS scientist Donald Becker augmented Linux with channel-bonding software to combine the performance of multiple Ethernet network channels efficiently and transparently. This Parallel Linux, distributed on the World Wide Web (wwW), has become the major source of Linux networking software, Sterling said. It also incorporates parallel programming APIs such as MPI, PVM and Bulk Synchronous Parallel.

NASA has instituted a Beowulf University Consortium to assist colleges and universities in building Beowulfs for teaching parallel programming techniques. Sterling pointed out that electrical and computer engineering students also can benefit from the experience of building a parallel computer. Caltech; Drexel, George Washington and Clemson Universities; and the University of Illinois at Urbana-Champaign are current participants. Several other universities, as well as magnet high schools, have expressed interest.

Besides its use in scientific computing, Beowulf is being tested as a mass storage device and as a satellite data processing engine. A $500,000 award from the Defense Advanced Research Projects Agency is supporting a 64-node Beowulf terabyte mass storage system with a gigabyte-per-second I/O rate. This array will serve NASA HPCC's 384-node CRAY T3E being installed at Goddard in 1997. Nine ESS Project Grand Challenge investigation teams will stress the system with computation runs individually producing up to 500 gigabytes of data.

NASA and the U.S. Air Force are collaborating on placement of Beowulf workstations as inexpensive satellite readout stations. They will allow data product generation and model and forecast processing in near real time, which is a ten-fold improvement over the current standard. Planned installation sites include the NASA Regional Data Centers at Clemson, Louisiana State University and the Universities of Hawaii and Maryland, Baltimore County.

For more information, see Beowulf Project wwW sites at the following URLs:

http://cesdis.gsfc.nasa.gov/linux-web/beowulf/beowulf.htm
http://www.cacr.caltech.edu/research/beowulf/
http://loki-www.lanl.gov/

- end -