The parallel route to an ocean model

Journal Article · Fri Apr 02 04:00:00 EST 1993 · Science (Washington, D.C.); (United States)

OSTI ID:6528026

Oceanographers Albert Semtner of the Naval Postgraduate School in Monterey and Robert Chervin of the National Center for Atmospheric Research (NCAR) in Boulder have done a better job of recreating the world ocean in a computer than any modelers before them. One element of their success is teamwork - in the form of a parallel supercomputer that solves many parts of the problem at once. That strategy boosts the speed of computation, but figuring out how to divide up the problem to best advantage can be tricky. For a parallel computer to achieve its full potential, the problem has to be structured so that the processors can work independently much of the time, spending as little time as possible swapping data. But each property of the ocean - temperature, density, velocity - influences every other, so Semtner and Chervin had to look for some other way to split up their problem. Their solution was to assign each processor its own separate chunk of the ocean, so that communications among the processors could be minimized. When Semtner and Chervin parceled out the ocean as best they could and optimized other aspects of the model, its speed on the four-processor Cray XMP jumped 15-fold from 30 megaflops to 450 megaflops. Moved to a cray YMP with eight processors, the model sped up to gigaflop speeds. Even so, it would take 100 hours on the Cray YMP to simulate just 1 year of detailed ocean circulation - and a year is far too short to predict the ocean's behavior in a greenhouse world, which would play out over many decades. If ocean modelers are to contribute to climate prediction, they will have to move beyond machines with a dozen or so processors to massively parallel computers with thousands or even millions of processors. As a start, Robert Malone, Richard Smith, and John Dukowicz of Los Alamos National Laboratory are running the Semtner and Chervin global model on a CM-5 Connection Machine with 1024 processor nodes.

• OSTI ID: 6528026
• Journal Information: Science (Washington, D.C.); (United States), Journal Name: Science (Washington, D.C.); (United States) Vol. 260:5104; ISSN SCIEAS; ISSN 0036-8075
• Country of Publication: United States
• Language: English

Related Subjects

550400* -- Genetics
59 BASIC BIOLOGICAL SCIENCES
ATLANTIC OCEAN
CLIMATE MODELS
COMPUTERIZED SIMULATION
COMPUTERS
CURRENTS
DIGITAL COMPUTERS
GENERAL CIRCULATION MODELS
INDIAN OCEAN
MATHEMATICAL MODELS
OCEANIC CIRCULATION
PACIFIC OCEAN
PARALLEL PROCESSING
PROGRAMMING
SEAS
SIMULATION
SUPERCOMPUTERS
SURFACE WATERS
WATER CURRENTS