Just as the Internet’s architects didn’t anticipate the need to defend against armies of hackers, they never foresaw flash crowds. These are throngs of users visiting a Web site simultaneously, overloading the network, the site’s server, or both. (The most famous flash crowd, perhaps, formed during a 1999 Victoria’s Secret lingerie Web broadcast that had been promoted during the Super Bowl. Within hours, viewers made 1.5 million requests to the company’s servers. Most never got through.) Such events-or their more malevolent cousins, denial-of-service attacks-can knock out sites that aren’t protected by a network like Akamai’s, which caches copies of customers’ Web sites on its own, widely scattered private servers. But the question is how many copies to make. Too few, and the overloads persist; too many, and the servers are choked with surplus copies. One solution, described in papers published in 1999 by the researchers who went on to found Akamai, is simply to set a fixed number.
In the not-too-distant future, PlanetLab nodes will adjust the number of cached copies on the fly. Here’s how it works. Each node devotes a slice of its processor time and memory to a program designed by Vivek Pai, a colleague of Peterson’s in the computer science department at Princeton. The software monitors requests for page downloads and, if it detects that a page is in high demand, copies it to the node’s hard drive, which acts like the memory in a typical Web server. As demand grows, the program automatically caches the page on additional nodes to spread out the load, constantly adjusting the number of replicas according to the page’s popularity. Pai says that simulations of a denial-of-service attack on a PlanetLab-like network showed that nodes equipped with the Princeton software absorbed twice as many page requests before failing as those running the algorithms published by the Akamai founders.
This new tool, known as CoDeeN, is already running full time on PlanetLab; anyone can use it, simply by changing his or her Web browser’s settings to connect to a nearby PlanetLab node. It’s a work in progress, so service isn’t yet fully reliable. But Pai believes the software can support a network with thousands of nodes, eventually creating a free “public Akamai.” With this tool, Internet users would be able to get faster and more reliable access to any Web site they chose.
But banishing flash crowds won’t, by itself, solve Internet slowdowns. Other PlanetLab software seeks to attack a subtler problem: the absence of a decent “highway map” of the network. Over the years the Internet has grown into an opaque tangle of routers and backbone links owned by thousands of competing Internet service providers, most of them private businesses. “Packets go in, they come out, and there’s very little visibility or control as to what happens in the middle,” says Thomas Anderson, a computer scientist at the University of Washington in Seattle.
One solution is software known as Scriptroute. Developed by Anderson and his colleagues at the University of Washington, it’s a distributed program that uses smart nodes to launch probes that fan out through particular regions of the Internet and send back data about their travels. The data can be combined into a map of the active links within and between Internet service providers’ networks-along with measurements of the time packets take to traverse each link. It’s like having an aerial view of an urban freeway system. Anderson says operators at Internet service providers such as AOL and Earthlink, as well as universities, could use Scriptroute’s maps to rapidly diagnose and repair network problems in one to three years.