Supercomputers

Gregory Chudnovsky, 39, with the help of his older brother, David, built a supercomputer in the living room of his New York City apartment. Seven feet tall and 8 feet across, the supercomputer was put together from mail­ order parts and cost only $70,000 [compared to $30 million for some com­mercial supercomputers).

The brothers, both mathematicians and former citizens of the Soviet Union, have found some drawbacks to their homemade machine, which they named "m zero." They must keep the computer, along with 25 fans, running day and night. They must make sure the apartment's lights are turned off as much as possible, to prevent blowing the wiring in the living unit. "The building superintendent doesn't know that the Chudnovsky brothers have been using a supercomputer in Gregory's apartment," reports journalist Richard Preston, "and the brothers haven't expressed an eagerness to tell him." Still, the machine makes their lives more convenient. The "m zero" performs computations that make up the basis of many of the scholarly papers and books they write on number theory and mathematical physics.

Most supercomputer users aren't as resourceful as the Chudnovskys and must buy their equipment from manufacturers. Typically priced from $225,000 to over $30 million, supercomputers are high-capacity machines that require special air-conditioned rooms and are the fastest calculating devices ever invented.

Supercomputer users are those who need to model complex phenomena. Examples arc automotive engineers who simulate cars crashing into walls and airplane designers who simulate air flowing over an airplane wing. "Supers," as they are called, are also used for oil exploration and weather forecasting. They can also help managers in department-store chains decide what to buy and where to stock it. Finally, they have been used to help redesign parachutes, which are surprisingly complex from the standpoint of aerodynamics. The supercomputer simulates the flow of air in and around the parachute during its descent. In 1995 Intel announced plans to build a supercomputer that would enable scientists to simulate the explosion of a nuclear bomb. Supercomputers are designed in two ways:

The traditional design, now 20 years old, is vector processing. In vector processing, a relatively few (1-16) large, highly spe­cialized processors run calculations at high speeds. The drawback is that tasks are accomplished by a single large processor (or handful of proces­sors] one by one, creating potential bottlenecks. In addition, the proces­sors are costly to build, and they run so hot that they need elaborate cool­ing systems.

Massively parallel processors! The newer design is called massively parallel processing (MPP), which spreads calculations over hundreds or even thousands of standard, inexpensive microprocessors of the type used in PCs. Tasks are parceled out to a great many processors, which work simultaneously. The reason the Chudnovsky brothers were able to build their super so cheaply was that they used standard microprocessors avail­able for PCs in an MPP design.

A difficulty is that MPP machines are notoriously difficult to program, which has slowed their adoption. Still, with the right software, 100 small processors can often run a large program in far less time than the largest supercomputer running it in serial fashion, one instruction at a time.

Massively parallel processing might seem as powerful as one could expect. However, fiber-optic communications lines have made possible supercomputing power that is truly awesome. In 1995 the National Science Foundation and MCI Communications, the nation's No. 2 long-distance provider, established a Very-high-speed Backbone Network Service (VBNS) that links the five most important concentrations of supercomputers into what they call a nationwide "metacenter." Each of these locations has more than one supercomputer (Cornell and Champaign-Urbana have six each). With this arrangement a scientist sitting at a terminal or workstation any­where in the country could have access to all the power of these fast machines simultaneously.

Servers

The word "server" does not describe a size of computer but rather a partic­ular way in which a computer is used. Nevertheless, because of the princi­pal concerns of this book—the union of computers and communications— servers deserve separate discussion here.

A server, or network server, is a central computer that holds databases and programs for many PCs, workstations, or terminals, which are called clients. These clients are linked by a wired or wireless network. The entire network is called a client/server network. In small organizations, servers can store files and transmit electronic mail between departments. In large orga­nizations, servers can house enormous libraries of financial, sales, and prod­uct information. The surge in popularity of the World Wide Web has also led to an increased demand for servers at tens of thousands of Web sites.

Network servers appear in a variety of sizes. As one writer points out, they may consist of "everything from souped-up PCs selling for $10,000 to main­frames and supercomputer-class systems costing millions." For more than a decade, he points out, the computer industry was driven by a rush to put stand-alone microcomputers in offices and homes. Now that we are far along in putting a PC on every desktop, the spotlight is shifting to computers that can do work for many different people at once

On the one hand, then, this puts "big iron"—minis, mainframes, supers— back in the picture. Recognizing this, IBM has combined formerly separate personal, midrange, and mainframe computer units into an umbrella organi­zation called the Server Group. On the other hand, the demand for servers based on microcomputers has made souped-up PCs and Macintoshes a growth industry—and is bringing these machines "close to the power of more expensive minicomputers and mainframes," according to some PC makers.

Now let's move on to look inside computers to see how they work.

Дата добавления: 2015-11-14