Oak Ridge National Laboratory

 

News Release

Media Contact: Fred Strohl (strohlhf@ornl.gov)
Communications and External Relations
865.574.4165

 

Three linked supercomputing centers earn gold medal for high performance

OAK RIDGE, Tenn., Feb. 12, 1997 — The Department of Energy's (DOE) Oak Ridge National Laboratory (ORNL) is part of a three-center consortium that is linking supercomputers from sites across the country to solve scientific problems that are too large to solve on a single computer.

Their success in this approach to problem solving netted them a gold medal in the category for fastest linked computers of the High Performance Computing Challenge.

The award was presented during the Supercomputing '96 Conference in Pittsburgh to ORNL, DOE's Sandia National Laboratories and the Pittsburgh Supercomputing Center. The competition, in which participants from around the globe demonstrate leading-edge projects, is held annually at the Supercomputing Conference.

Four massively parallel processing (MPP) computers, in various combinations, are linked to solve scientific problems that are too large to solve on a single computer. These include two Intel Paragons at ORNL, one Intel Paragon at Sandia and a Cray T3D at Pittsburgh. An MPP consists of hundreds to thousands of processing nodes, each of which is analogous to a powerful desktop computer. Nodes are connected to one another via a high-speed network within the machine. MPPs solve problems using a divide and conquer strategy in which each node works on a small part of the overall problem. Nodes exchange data over the network, when necessary.

Each MPP consists of between 512 and 1,824 nodes. Eventually, 3,872 nodes will be able to be brought together to solve a single problem. What would take 10 years to solve on a personal computer could be solved in one day on the linked system.

While the linkage concept is simple, there are obstacles to overcome. A computer that is waiting for data from another machine cannot do its work. Robust, high-speed networks provided by DOE's ESnet and the National Science Foundation's vBNS are solving this problem in order for the machines to get the data they need quickly and continue computing.

The four computers are from two different manufacturers and run a total of three different operating systems. The programs must be ported to each machine separately and then made to run on them as if they were one machine. This is like having to translate a book for readers who speak three different languages and then lead a discussion about it. Parallel Virtual Machine (PVM) software, which was developed at ORNL, serves the function of a simultaneous translator. PVM lets a user connect different machines and presents the user with the image of the connected machines as a single MPP.

Initially, three problems are being solved using this system. The first is a model of the alloy nickel-copper, which exhibits magnetic behavior when it is composed of predominantly nickel, but does not when it is predominantly copper.

Using a computer program developed at ORNL, scientists are beginning to uncover the physical mechanism responsible for complex magnetic behavior first observed 25 years ago. This research into the fundamental nature of magnetic alloys paves the way for studies of a variety of magnetic materials. Application of this research could improve computer storage devices such as hard disk drives, magnetic motors used in the power generation industry and shadow masks to sharpen images on computer monitors and televisions.

The second problem involves predicting the response of a nuclear weapon to the effects of a hypothetical nearby explosion. The blast and fragmentation environment from the nearby blast would present the possibility of a sympathetic detonation of the weapon. The calculations will help assess the safety performance of the warhead in such a scenario without the need for an extensive and costly full-scale test program.

The final problem links atmosphere, ocean and sea ice computer models to study the Earth's climate system. Physically realistic climate models enable scientists to assess the consequences of natural and man-made environmental changes on the Earth's climate.

"This project has linked more than computers," explained Tim Sheehan, director of special projects at ORNL's Center for Computational Sciences. "Physicists, mathematicians, computer scientists, network engineers and applications programmers from all three centers have worked together on this project for more than one year."

Driven by the desire to solve massive scientific problems, they continue to work together to push the frontiers of science by exploiting the power of linked massively parallel supercomputers.

ORNL, one of the Department of Energy's multiprogram national research and development facilities, is managed by Lockheed Martin Energy Research Corp.

.