Anyone looking to make sense of the data tidal wave will find another problem as well: understanding the systems that generate and model the data -- systems that are growing in complexity as fast as the data itself. The Internet is a prime example, points out Professor of Computer Science Rajive Bagrodia.

Twenty years ago, the Internet was nonexistent. Today it is a network of nearly 60 million machines. It's so congested with users that the World Wide Web is jokingly referred to as the "World Wide Wait."

"And people expect it to go up to billions of interconnected devices within a decade," notes Bagrodia, who insists that if individuals and scientists are going to access the data available on the Web, researchers must develop ways to understand the complexities of this kind of exponentially growing system.

"Analyzing the internet is a huge problem, the it really needs to be done," says Bagrodia, "And the sooner the better."

In fact, Bagrodia and his colleagues have been working on doing just that. The India-born scientist got his first education in computer science at the Indian Institute of Technology in Bombay. He came to the States to get his doctorate in computer science at the University of Texas at Austin, where his thesis was on parallel computer and the performance of large-scale systems.

He has been working on these systems ever since. The majority of his research has centered on developing methods to predict the performance of parllel computers when they're given complex problems to solve or model. This will allow computer systems to be designed to account for anticipated problems rather than having to be refined later on to deal with unexpected hitches.

"If you could predict performance," he notes, "you could make design trade-offs in the early stages of a system's development, rather than after. You could figure out where the performance bottlenecks will be and then design the system accordingly."

Among the problems Bagrodia has worked on at UCLA are the performance of large-scale simulations, most importantly, a study for the Department of Energy on testing nuclear weapons. "Once the U.S. government decided ther would be no more physicial testing of nuclear devices, they moved to purely computational wayss to test the technology," explains Bagrodia. The computations, however, are enormous. To model the explosion of a nuclear device with any accuracy requires looking at what happens in milisecond or shorter time scales and over extremely short distances. The model describe the relevant physics in three-dimensional grids, and the smaller hte volume described by each space in the grid -- a variable known as granularity -- the better the accuracy of the simulation. "To run them at the level of granularity they want," says Bagrodia, "even with 20,000 processors, can take more than a few years. So the kind of analysis I do is essential to figuring out whether you can solve the problem by throwing more processing power at it. If not, you have to come up with new fundamental algoritms or techniques."

At present, Bagrodia is working on solving the World Wide Wait problem with a $1.2-million grant from the Defense Advanced Research Project Agency (DARPA). "The idea is to think of the Internet as a complex system and try to understand what causes network congestion and how to deal with it," he explains. "We want to come up with a way in which we can first show under what conditions congestion builds up -- a phenomenon known as hot spots -- and then be able to evaluate whether any particular algoritms or techniques being suggested to solve hot spots can actually ork. No one really understands all the causes of why and when congestion will develop on the Internet, and no one can predict it. Our goal is to really understand how the growth in network size is affecting its performances and then evaluate different ways of solving that problem."

This problem, however, is made ever more complex because the origins of Internet traffic have been changing rapidly. "Just a few years ago, all traffic on the Internet originated on one wired node and terminated at another wired node," Bagrodia notes. "If you view the Internet only as a wired structure, it's already sufficiently complex. But nowww it's getting much worse because people are attatching wireless devices and satellite-based devices. The dramatic growth in the different typoes of devices attatched to the Internet and the traffic they generate is going to make the congestion much, much heavier than it is today."

Bagrodia and his colleagues are working on solving the problem using a technique he calls "hybrid models," a process that involves a combination of simulation and analytical techniques. They're very complex, he says, because the systems they're trying to understand are complex. Bagrodia runs the models on parallel supercomputers so they can be executed in a reasonable amount of time. He uses a language called PARSEC (Parallel Simulation Environment for Complex Systems), which he and his research group designed. "It is one of the very few languages in the world that can run on a large set of parallel computers, and is about the only language used to develop simulation models of such large networks," he says.

The goal of the DARPA contract is to simulate such a network with 1 million nodes and then be able to predict the impact of diffferent types of traffic on its performance. So far, Bagrodia and his colleagues, including Mario Gerla and Lixia Zhang from the computer science department, have been able to simulate Internet-like networks for tens of thousands of nodes. The next challenge will be to learn how to extrapolate from a system of a million nodes to one of a billion, which should more closely approximate the Internet of a year years from now.

"Clearly, we're not going to be able to run a simulation with a billion nodes," admits Bagrodia, "It doesn't matter how fast the computer or what algorithms we use. We know that will be impossible."

The hard, cold reality doesn't seem to daunt Bagrodia, however, as he shrugs, "We'll have to find other ways to solve this problem."

[previous page] [main] [next page]