AT&T touts tool to map IP traffic

FRAMINGHAM (10/14/2003) - Researchers at AT&T Labs have applied leading-edge statistical techniques to create what they say are the first real-time traffic reports for IP networks. This breakthrough is the final piece of a six-year effort at AT&T Labs to find a way to provide IP network management that is on par with that of the extremely reliable telephone network.

AT&T Corp. has coined the term "tomo-gravity" to describe its new IP traffic measurement technique. The technique combines statistical gravity modeling with computer-automated tomography, which is the CAT in a medical CAT scan. The result is better network planning, provisioning and troubleshooting, researchers say.

"Without the traffic-matrix computation, network engineering for IP networks was more an art than a science," says Albert Greenberg, division manager for network measurement and engineering research at AT&T Labs. "We've been knocking off different pieces of (this problem). . . . This is the missing piece that we recently discovered."

AT&T's tomo-gravity software plugs SNMP data into equations that run in seconds to create an accurate picture of AT&T's IP traffic. The reports provide a detailed look at the traffic flowing across AT&T's IP network, which carries more than 1 petabyte - or 1,000 terabytes - of data each day.

Similar reports on traffic volumes and patterns are used widely in capacity planning and management of telephone and ATM networks. These traffic reports were unavailable for IP networks until now because scientists had not discovered the equations needed for the calculations.

Estimating traffic demand "is an incredibly important problem, with uses in network planning and provisioning," says James Kurose, a professor in the department of computer science at the University of Massachusetts Amherst. "Telephone networks have used such information extensively, but this information has been very hard to get for IP networks."

Kurose calls AT&T's tomo-gravity approach "ingenious" and says it estimates traffic "in an extremely computationally efficient way. What's also impressive are the validation results that they've shown in an operational network."

AT&T says it will run daily IP traffic reports as part of its standard network operations by year-end.

For AT&T's corporate customers, the IP traffic reports should translate into better reliability for AT&T's IP services, company officials say. These reports are designed to help AT&T route around congestion, failures and scheduled maintenance using real-time traffic data instead of estimations that often proved inaccurate.

With an accurate picture of its IP traffic, AT&T will be able "to engineer the network to ensure that there's no impact when there is any type of preventative maintenance or failure," Greenberg says. "No matter what hits the network (our corporate customers) will see a continuous service."

AT&T also is considering using IP traffic reports as part of a managed service for enterprise customers.

"If you look at any enterprise network, they all have a migration plan. . . . It's just the nature of networks and the dynamics of the economy," Greenberg says. "This is one of the tools that can help (a company) plan and run their networks and migrate from A to B." It can help network managers make good decisions about migrations, and it can help during the switchover process.

Corporate IT executives are keen on the idea of being able to buy IP services with better underlying network engineering.

"If AT&T is saying that they can make their IP network more intelligent . . . that would be a potential benefit," says Jay Woloszynski, vice president of shared technology services and CTO of Oxford Health Plans Inc., a Connecticut healthcare company. "If they can do it without serious overhead, it sounds very interesting."

The tomo-gravity software produces what's called a traffic matrix for an IP network.

A traffic matrix tells network engineers not only the volume of traffic on major backbone routes but also the source and destination of all the packets on the network. Traffic matrices are available for the public switched telephone network and ATM networks, which are both connection-oriented networks. However, IP networks are connection-less, so it has been impossible to gather this information until now.

"A basic traffic matrix would be able to look at the roads and not only tell you all the people who are driving right now between New York and Orlando but how many of them are between New York and Philadelphia," says Greenberg, who leads the team of researchers that invented tomo-gravity. "It also tells you all the entrances and exits for all the people on the road."

The tomo-gravity software uses information that is readily available for IP networks - the number of packets across particular links and the configuration data on the routers - to extrapolate the traffic matrix.

Routers in AT&T's IP backbone count packets across tens of thousands of links, with measurements being collected every 5 minutes using SNMP. Network operators also collect configuration data on all the routers in AT&T's IP backbone.

"We already collected these two pieces of data . . . but we didn't know how to put them together until recently," Greenberg says.

To arrive at the tomo-gravity approach, AT&T researchers first used gravity modeling, which is a standard statistical approach for analyzing data. They refined the gravity modeling and combined it with tomography - also used in brain and Sun imaging - which let them use all the data they needed to calculate an accurate traffic matrix.

The AT&T researchers then automated this process by creating software that crunches through huge volumes of SNMP data and produces a spreadsheet that lists the amount of data flowing between each of the routers on the network. AT&T's tomo-gravity software runs on a standard laptop and takes only seconds to produce an IP traffic matrix.

Researchers at the lab are training the company's network operations staff to use the tomo-gravity software to output regular IP traffic matrices. AT&T's network operations staff will use IP traffic matrices daily to predict traffic under planned or unexpected router or link failures, to forecast future network requirements, to optimize routes and to minimize congestion.

The researchers invented tomo-gravity a year ago, but they first presented it at two network engineering conferences this summer. The researchers have tested the software for a year, getting it ready to move out of the labs and into production use.

AT&T funded the tomo-gravity research internally, and a team of four engineers worked on it.

The labs' six-year research effort was designed to create automated, scientific tools for conducting network engineering on IP networks. The team tackled three problems: identifying the topology of IP networks, optimizing routing on IP networks and creating a traffic matrix for IP networks. The tomo-gravity approach solved the last of these three problems.

AT&T's tomo-gravity software is "very good at giving conservative and reliable estimates of IP traffic," says David Donoho, a professor of statistics at Stanford University in Palo Alto.

Donoho helped AT&T researchers understand and refine the statistical techniques that the tomo-gravity software uses to generate its speedy and accurate results. The challenge of creating an IP traffic matrix is similar to other problems Donoho has studied, such as brain imaging, where relatively little data is available to answer questions.

"I heard about what AT&T was doing (with tomo-gravity) and I said: 'Gee, did you recognize that there is an information theory way to view this problem?'" Donoho says.

Next for the tomo-gravity researchers is developing ways to gather better network traffic data from routers and network probes.

"If we could redo what the line cards and routers are able to measure, then we can do better than SNMP data," Greenberg says. "We'd get more refined measurements, and we could engineer the IP network better."

Join the newsletter!

Error: Please check your email address.

More about AT&TFuture NetworkStanford University

Show Comments
[]