New Fields for the Supercomputer
By Lawrence M. Fisher
The New York Times
June 28, 1989
Using a supercomputer to process a payroll or to make hotel reservations may sound a bit like using a sledgehammer to kill ants, but in fact these exotic machines are showing some real advantages for many business tasks.
Once the exclusive domain of scientists and engineers, supercomputing is moving into more conventional commercial applications like the management of rental-car reservations and the development of financial models as well as the processing of transactions like credit card charges.
These are the classic uses of large mainframe computers, which excel in moving and sorting large volumes of data. Supercomputers specialize in performing mathematically intensive processing operations to solve very complex problems rapidly.
But both users and producers are finding that the supercomputer technology called parallel processing, in which tasks are broken up and distributed among multiple computer processors, accelerates some business software as well as it does scientific programs. Parallel processing overcomes the speed limits of individual processors; theoretically, four processors handle a problem almost four times as fast as one.
'Scared a Lot of People'
''It's just become reasonable recently for people to develop data base software on parallel processors,'' said David DeWitt, a professor of computer science at the University of Wisconsin. ''Before, you had to go out and start your own hardware company as Teradata did, and that scared a lot of people off.''
Mr. DeWitt noted that Tandem Computers Inc., a pioneer in the use of multiple processors, has rewritten portions of its Non-Stop SQL data base software to run in parallel, and that the International Business Machines Corporation is taking a similar approach to its mainframe data base product, DB-2.
Some simple parallel systems have been used in business since the early 1980's, but the announcement last week by the Oracle Corporation that it would write a version of its leading data base software for a machine with more than 8,000 processors was a kind of watershed for the technology. Oracle's relational data base program, which can find sets of related records simultaneously, is used by thousands of companies in a range of industries.
''A number of large Oracle customers are extremely interested in the commercial use of supercomputers,'' said Peter Tierney, Oracle's vice president for marketing, explaining why Oracle will produce software for supercomputers made by Ncube, which is based in Beaverton, Ore.
Mr. Tierney said that the first customers would probably be in the petroleum and military industries, which are already big supercomputer users, but that there was also interest from financial service, transportation and hospitality companies.
Jeffrey Canin, an independent supercomputer analyst in San Francisco, said: ''Parallel processing is not a controversial issue anymore. Companies with the credibility of Oracle blessing this architecture will make it safe for other kinds of applications.''
There is ample precedent for supercomputers moving into the commercial arena. Nearly every new computer architecture has first been adopted by the scientific community, which is more willing to take risks to gain performance, and then by business. Perhaps the best example is the Digital Equipment Corporation's Vax minicomputer, which started as an engineer's tool but became a successful general-purpose computer.
Indeed, Gordon Bell, the chief designer of the Vax and now a vice president with the Ardent Computer Corporation, said parallel processing would gradually take over commercial computing. ''It's happening in the technical areas and there's no reason it won't happen in commercial areas,'' he said. Only the conservatism of commercial users and their investment in existing software would slow the change, he said.
But some analysts are skeptical, noting that converting software to run on a parallel system is not easy, even for scientific users. Within the broad field of parallel processing, machines run from four to eight processors to as many as 64,000 in ''massively parallel'' systems, and some may be better suited to business than others.
''It is going to be a number of years before true parallel processing machines are really considered mainstream architecture,'' said Gary Smaby, an analyst with Needham & Company in Minneapolis. ''There are certain types of problems that can be broken up very cleanly across multiple processors, but most common programs don't lend themselves to that.''
Most analysts say the machines most likely to gain a foothold in the commercial market are the so-called mini-supercomputers, typically priced from $300,000 to $1 million, rather than true supercomputers like those of Cray Research Inc., which range in price from $5 million to $25 million. A Cray spokesman said that while the company's machines were sometimes used to run commercial software, they were not bought for that purpose. He added that Cray was not aiming for that market.
One maker of parallel systems that has focused on the commercial market is Sequent Computer Systems Inc. of Beaverton, Ore. Sequent's machines fall below the supercomputer level, but offer faster performance than comparably priced minicomputers by linking up to 30 microprocessors of the type commonly used singly in personal computers.
C. Scott Gibson, Sequent's president and chief operating officer, said the company's computers offer ''practical parallel.''
''We wanted our parallel machines to be out there living and breathing in real applications,'' he said. Sequent's customers include the Internal Revenue Service, which uses Oracle software to respond to taxpayer questions, and the Radisson Hotel Company, which coordinates reservations for 200 hotels using an Action Software program and a four-processor Sequent system.
Parallel systems have also created new types of applications. The Teradata Corporation of Los Angeles produces a specialized data base supercomputer using up to 268 processors. Teradata's machines can be linked to the mainframes or minicomputers of other companies and are used primarily for decision support, which Richard Van Hoesen, a spokesman, defines as ''the ability to take large amounts of information and manipulate and aggregate it in such a way that it's usable.''
Teradata counts many large companies among its customer base, including Citibank, K Mart and A.T.&T. ''This is mainstream big business,'' Mr. Van Hoesen said. ''The reason we're in there is we're enabling them to do things they know they need to do and can't find any other way to do.''
Mr. Van Hoesen uses a deck of cards as an analogy for a computer data base to explain why parallel processing works so well. If one person is asked, for example, to find all the sixes in a deck, shuffling through all 52 cards would take a minute or more. Four people each shuffling through 13 cards would reduce the task to seconds, and 52 people could accomplish the task in an instant. The same is true of computer processors searching a data base.
Thinking Machines Inc. of Cambridge, Mass., has also moved beyond scientific problems to business applications. The company recently sold two of its 64,000-processor computers, dubbed the Connection Machine, to Dow Jones & Company for a new electronic information retrieval service called Dow Quest. This service lets an unskilled user browse rapidly through a vast library of information from business newpapers and magazines.
''What goes on behind that is a huge computation that just could not be done on a sequential machine in a reasonable amount of time,'' said Danny Hillis, Thinking Machines' chief scientist. In such applications the issue of converting existing software to run in parallel is not relevant, ''because you didn't have old software to do that,'' he said.
''We ourselves look more toward new applications,'' he added.
Other companies have found that the distinction between science and business applications is not black and white. Business tasks may rely on algorithms similar to engineering programs, and many scientific tasks require just as much data base management as commercial jobs do.
''Many problems you find in business can be adapted to high-performance 'scientific' machines like ours,'' said Frank Vince, vice president for marketing at the Convex Computer Corporation in Richardson, Tex., the leading maker of mini-supercomputers. The advantage is the ''ability to do much more work in a given period,'' he said. Many business applications, like financial modeling, are actually scientific in nature, he added.
Justin Rattner, director of technology for Intel Scientific Computers, agreed. ''Even though the problems are different in business, the math is very similar to problems we've already tackled in science,'' he said. But the value of parallel processing in business remains very application-dependent, he added. ''Parallel machines really shine when you're doing very complex queries, the kind of question that ties up an I.B.M. for hours,'' he said.
Wider Uses for the Big Machines
For the past decade supercomputers have been used almost exlusively for scientific applications. Now new types of software that exploit a supercomputer technology known as parallel processing are introducing the machines to commercial applications in a wide range of industries.
Some of those applications:
Banks: Customer transactins, mortgage and loan analysis.
Brokerage firms: Trade execution, program trading, tracking os stocks and assets.
Hospitals: Patient records, insurance reporting, purchasing decisions.
Mail-order houses: Order fulfillment, customer demography.
Manufacturers: Adjusting just-in-time productin, inventory analysis, quality control.
News Services: Electronic Information Retrieval.
The Difference in Parallel Processing
Computers bases on parallel processing are making it easier to perform complicated tasks more rapidly and creating new capabilities for a wide variety of businesses. Under the old single processing system, one program is handled by one processor and tasks are processed in sequence. In a multiprocessor system, each processor handles one program. In parallel processing each program is broken into different tasks that are distributed among more than one processor.
Diagram (pg. D4)
Copyright 1989 The New York Times Company