Race Is On To Use Supercomputers In Research

Business

Race Is On To Use Supercomputers In Research

By Robert Sanford
St. Louis Post-Dispatch

January 4, 1988

With machines revved up and running, a race is on to broaden the use of supercomputers in scientific research.

Five supercomputing centers established by the National Science Foundation are principal participants. One of these, the National Center for Supercomputing Applications at the University of Illinois, Urbana-Champaign, announced last month that it had signed a fourth industrial partner, Motorola Inc.

An industrial partnership in the NCSA program is an agreement in which the partner company donates $1 million a year for three years. In return, the company's representatives and researchers receive training in advanced computing techniques and use supercomputers at the center on projects of interest to the company. The other three companies with partnerships at NCSA are Eastman Kodak Inc., Amoco Corp. and Eli Lilly and Co.

A goal at NCSA is to have 10 industrial partners, said Larry Smarr, director. ''Our experience so far indicates we will make that goal,'' he said.

As interest in the use of supercomputers has grown in the last few years, the National Science Foundation h as established communications networks to broaden accessibility to the machines. An example is a network called MIDnet. Washington University here and eight other Midwest universities are members in a leased-wire loop, with Washington U. offering connection to the NCSA at Urbana.

''Time saved by using a supercomputer can be a factor of 100; that is to say that the supercomputer can be 100 times faster compared to usual computer speed,'' said Jonathan Turner, an associate professor at Washington U.

Smarr, at the University of Illinois, is an advocate of supercomputer use. Developments are moving forward in the U.S. effort, he said, and it's about time.

''Time is running out for American industry if it is to compete with the Japanese,'' he said. ''The Japanese government is supporting supercomputer development. They will use it in a strategic sense, targeting industries that they think will be leaders globally in the 1990s.

''I think that supercomputers will become ubiquitous. They will be the machine tools of the 1990s. But years of training are needed. We don't have many industries left to lose to the Japanese.''

Smarr said industries such as biotechnology, electronics and engineering of large structures require supercomputers for timely research into very complex matters. In one instance the investigator may be dealing with creating specific molecules, he said, and in another he may be tracing paths of individual electrons through a semiconductor material such as gallium arsenide.

''Let's consider a project like the aerospace plane, the so-called Orient Express,'' he said. ''It cannot be built without supercomputers. The materials for such a vehicle do not exist. They have to be designed.

''Consider the airflow around it,'' Smarr said. ''You can take the space around the plane and cut it up into 10-foot cubes and calculate the air movement. That sort of measurement could be so coarse that it doesn't produce a very good answer. Or, using a supercomputer, you can cut the space into 1-inch cubes and get a high-resolution solution.

''Just because of its speed, a supercomputer makes many things possible,'' he said. ''You make a model of a problem like tracing electron paths. This amounts to creating a numerical laboratory in which one can investigate the electron transport through a certain material without actually having the material. One can calculate how a transistor turns on and turns off.

''Or one can determine how often a metal stamping machine can stamp out a part without becoming overly hot. That's just one example of how the supercomputer can impact manufacturing.''

At the NCSA at Urbana there are two Cray supercomputers, one of which is upgraded - made faster - each year. The speed is added by adding processor circuits. In 1985 the center had a Cray with two processors. It now has a Cray with four. By 1989 a machine will be upgraded to 16 and by the early 1990s to 64. The technique of adding processors is called parallel processing; the machine breaks a problem into parts and solves the parts simultaneously.

David Kuck, director of a supercomputer research and development center at Illinois U., uses an analogy of shopping in a supermarket to explain parallel processing: If a shopper has 100 items on a grocery list he may take an hour to gather them. But if the shopper brings 99 friends along and assigns one item to each, and if the store has 100 checkout lanes, shopping time can be cut dramatically. Of course, the process must be organized. The shoppers should move about without bumping into each other and the checkout clerks should be able to communicate so that they can arrive at a final bill. Putting such instructions into a machine is not simple.

Parallel processing, pioneered at Illinois, received a boost in the commercial market last month with an announcement by IBM, the world's largest maker of computers, that the company hoped to market a 64-processor machine by the early 1990s. A consultant to IBM in the project is Steve S. Chen, who studied under Kuck at Illinois and who designed some Cray supercomputers.

The National Science Foundation emphasis on developing use of supercomputers began almost three years ago with an announcement that the NSF would give grants of more than $200 million to establish supercomputer centers at Illinois, Cornell, Princeton, Pittsburgh and the University of California at San Diego.

The centers are connected by a leased-wire network called NSFnet. ''With NSFnet in service more than a year, it is being used almost at capacity,'' said Steve Wolff, an NSF spokesman. ''The use is running at 100 million packets per month. A packet is a bundle of characters, or bytes. The capacity at this time is 56,000 bits per second. That is going to be increased 20-fold in the near future.''

NSFnet also is fed by seven regional nets that have connections at more than 150 colleges and universities. MIDnet in the Midwest is one of the regionals. MIDnet members include Washington U., Missouri, Kansas, Kansas State, Nebraska, Iowa State, Oklah oma, Oklahoma State and Arkansas. MIDnet is supported by the NSF at a cost of about $1 million for three years, after which the universities are expected to share costs.

Smarr, at Illinois, says the nets offer more than just connections to supercomputers, they offer communications connections that did not exist before.

''Scientists who have previously communicated annually at national meetings can participate in real-time, multicolleague electronic mail discussions,'' he said. ''I think this development is going to cause a major shift in the sociology of computational science.''

More than 130 machines are in use in the field of supercomputing. Most are owned by manufacturers or oil companies that deal with complex concepts such as modeling the activities of an oil basin. Some owners sell time on their machines to others. The price could be more than $1,000 an hour.

''At the NCSA we are not in the business of selling time commercially,'' Smarr said. ''We have more than 600 scientific projects from universities with access to our supercomputers. On the commercial side we have turned down a number of offers to sell time.

''But we do offer the industrial partnership. It is more than just time on the machine. We give training. The corporate researchers work with our computer experts and ideas are exchanged.

''When Kodak joined it sent 40 researchers. We have more than 200 workstations and microcomputers as connections to the supers,'' he said. ''We have experts at visualizing models. The atmosphere at the center can be stimulating.''

Smarr said the partnership idea has been praised as a new way in which the government can give universities and industry a chance to work together without direct government involvement. The partnership is expensive, he said, so it requires consensus at the corporate level. With this support the supercomputer work can become a strategic part of management.

''Such an approach can contribute to changing the way American industry sees the future,'' Smarr said.