Clusters for a virtual supercomputer.

When a computational problem is simply too big for a modest desktop computer to deal with, it makes sense to get help by tapping into other computers. This is the notion behind cluster computing. All the computers in a cluster are linked in such a way that the "overflow" from one can be handled by the rest. Such a group of computers, spread across many different locations within an institution, can sometimes perform as capably as a supercomputer.

Now, research groups at several universities have embarked on an ambitious scheme aimed at connecting computer clusters at different locations into a cluster of clusters. A user at one computer would have access to as many additional computers as necessary to solve a given problem, no matter where the user is or where the computers are.

Physicist Robert J. Hollebeek of the University of Pennsylvania in Philadelphia, computer scientist Robert L. Grossman of the University of Illinois at Chicago, and physicist Andrew Baden of the University of Maryland at College Park head this collaborative effort, known as the National Scalable Cluster Project.

"With emerging technology in telecommunications, we can link dozens of computers in three cities to act like one super-computer," Grossman says. "It's like putting a virtual super-computer on everybody's desktop."

Such a scheme represents a logical step beyond the Internet -- the network of computer networks -- now widely used by researchers and many others. However, instead of merely exchanging messages and reading files, users end up with access to computational resources far beyond those normally available to their own machines (SN: 5/1/93, p.280).

One crucial element is the use of high-speed, high-capacity telecommunications technology so that geographically distributed clusters can operate as a single computing facility. Another is the development of system software that makes it easy to add new computers to the supergroup to increase its resource pool.

Researchers at each of the three institutions involved in the project already have experience with cluster computing on the local level. For example, Grossman and his collaborators have been developing methods for managing huge quantities of data on such clusters. They recently introduced a technique for approximating the best possible flight path for an airliner traveling from one city to another, taking into account fuel costs, weather conditions, and other factors. Their scheme involves creating a large database made up of short trajectory segments, which can be assembled to match as closely as possible the required conditions for the desired flight path (see illustration). That's typically faster than computing the flight pattern from scratch.

Cluster computing has also aided in the analysis of data from particle physics experiments and on a larger scale may prove useful in medical imaging, robotics, manufacturing, scientific computing, and managing digital libraries. But to get there, the researchers have some sticky issues to resolve, ranging from effective resource sharing to achieving sufficiently high levels of security and reliability.

"Every time you go to a new scale, you discover new problems," Hollebeek says. "It's not a done deal."