Science Investigators Achieve 10 GigaFLOPS on Goddard CRAY T3D

by Jarrett Cohen

NASA-funded researchers recently passed a performance milestone of 10 billion floating-point operations per second (gigaFLOPS) on supercomputing applications across the spectrum of the Earth and space sciences.

Speeds ranging from 10.02 to nearly 18 gigaFLOPS were enabled by a 512-processor CRAY T3D supercomputer placed at Goddard Space Flight Center in October 1996. The nine Earth and Space Sciences (ESS) Project Grand Challenge Investigation teams were assisted by Cray Research, Inc. scientists in achieving the advances. The computer codes will be available on the World Wide Web's National High Performance Computing and Communications (HPCC) Software Exchange to benefit the wider research community.

The primary goal of the ESS Project is to employ scalable parallel computers to reach a new understanding of Grand Challenges, fundamental science problems with broad impact whose solution can be advanced by high-performance computing. As application speed (along with memory) increases, so do the size, timescale, and complexity of the problems that can be addressed.

The 10 gigaFLOPS performance milestones were the first of three that will continue with 50 and 100 gigaFLOPS, to be met on CRAY T3E systems as large as 1,024 processors. The first 256 processors of a 512-processor CRAY T3E were placed at Goddard in late March. The Investigations conclude in 1999. Progress on the full set of milestones could be monitored on the ESS Project's Web site.

Simulations of the Earth's Core and Mantle Dynamics
Peter Olson, Johns Hopkins University
The goal of this investigation is to simulate the chaotic processes that drive the evolution of the planet's interior, and in turn shape its surface, over timescales ranging from hundreds to millions of years. Olson's team achieved 11.58 gigaFLOPS with TERRA, a 3-D spherical finite element mantle dynamics code developed at Los Alamos National Laboratory.

SAR Interferometry and Imaging Science
David Curkendall, Jet Propulsion Laboratory
Curkendall's team aims to use multiple supercomputers to process and visualize satellite-collected synthetic aperture radar (SAR) data to allow close monitoring of regional changes in alpine glaciers, plate tectonics, and rain forests. Their Scalable SAR Software Suite (S4) processed two images simultaneously at a speed of nearly 18 gigaFLOPS, a nine-fold improvement over NASA's current Thinking Machines CM-5 SAR data-processing system.

Four Dimensional Data Assimilation
Peter Lyster, University of Maryland
Melding observations and climate model prediction into a robust data analysis scheme for NASA's Earth Observing System, this effort will provide the most accurate possible picture of the atmosphere through space and time. The Physical-space Statistical Analysis System (PSAS), which is the core of the group's data assimilation approach, attained 10.4 gigaFLOPS. Assimilating 79,938 observations took 82.9 seconds on the 512-processor CRAY T3D, a tiny fraction of the 18,240 seconds required on a single CRAY C90 processor.

An Earth System Model: Atmosphere/Ocean Dynamics and Tracers
Chemistry
Roberto Mechoso, UCLA
Mechoso's team is focused on realistic portrayal of Earth's climate through building four coupled, highly complex models with high spatial resolutions: atmospheric general circulation, oceanic general circulation, atmospheric chemistry, and oceanic chemistry. They achieved 10.1 gigaFLOPS running the atmospheric and oceanic general circulation models uncoupled but "side-by-side."

Rayleigh-Benard-Marangoni Problems in a Microgravity Environment
Graham Carey, University of Texas at Austin
This group models fluid flows in low gravity environments to test the effectiveness of manufacturing higher quality thin films and coating processes in space and the functioning of the space station's life support and safety systems. Their <MGFLO> computational fluid dynamics code for analyzing instabilities in free surface phenomena reached 16.5 gigaFLOPS.

Turbulent Convection and Dynamos in Stars
Andrea Malagoli, University of Chicago
Malagoli's team studies some of the most fundamental and least understood turbulent processes in the interior of stars like the Sun, whose dynamics are only beginning to be inferred from new space probe and Earth-based observations. Their milestone involves two magnetohydrodynamics (MHD) codes, for modeling fluid dynamics in the presence of a magnetic field. A speed of 10.4 gigaFLOPS was achieved using Piecewise-Parabolic Method Convection (PPMC), which solves Euler's equations for compressible fluid flow with essentially no viscosity. Their MHD Pseudo-Spectral (MPS) code performed similarly at 10.1 gigaFLOPS. MPS solves the MHD equations in terms of the primitive velocity and magnetic field intensity variables. Both PPMC and MPS use 3-D Fast Fourier Transforms.

Solar Activity and Heliospheric Dynamics
John Gardner, Naval Research Laboratory
The goal of the Gardner group is to model the tangled three-dimensional structures that develop in the magnetic field of the Sun's corona (outer atmosphere), which NASA observations show to have a key role in the physics of solar activity. They also have two MHD codes. CRUNCH3D, a high-order spectral method, ran at 10.52 gigaFLOPS; it is useful for studying processes that require accurate resolution over a broad range of spatial scales. Their flux-corrected transport (FCT) code, computing at 10.02 gigaFLOPS, takes the finite-volume approach, optimal for modeling systems with flow properties that undergo large changes over short distances.

Graphical representation of AMR/Tamas Gombosi, et al., Univ. of Michigan