Skip to Content

Forecasting Network Traffic to Avoid Meltdowns

Understanding where wireless bandwidth demand will come from can keep communication systems from overloading.
December 16, 2010

In September, when the University of Alabama football team took on San Jose State, more than 100,000 fans packed the stadium and watched the Crimson Tide wash over the Spartans, 48 to 3. After the game, cell-phone customers flooded AT&T and the Tuscaloosa News with complaints of dropped calls, missing text messages, and an inability to connect to the network, not only at the arena but across the city.

Mapping the net: Modeling software sketches out a communications network, with lines color-coded according to how much traffic they’re carrying.

Bandwidth logjams have suddenly made it more urgent to come up with better ways of modeling and dealing with future traffic growth so that networks don’t experience total meltdowns. “When it was just voice, planning models were easy to design; engineers were able to do it on spreadsheets,” says Nick Shanker, CEO of Cerion, a company in Frisco, Texas, that models upcoming traffic demand for cellular networks. “You just don’t do linear planning anymore. You need help.”

Big games, political events, and high-tech conventions are notorious for overwhelming cell-phone networks with surges of traffic. Even on regular days in Manhattan and San Francisco, networks frequently exhibit the same symptoms. The problem occurs when large numbers of people use their smart phones at the same time to make calls, send photos, download video, or post updates to Facebook. But it’s not just one-time events or Times Square crowds that can swamp network capacity; bandwidth demand is surging just about everywhere thanks to the leap in smart-phone subscriptions and the introduction of new applications such as videoconferencing.

Predictive modelers anticipate future traffic demands and help telecommunications carriers plan accordingly. They start by collecting a carrier’s data to understand what has gone on in a network and what it looks like now—how much traffic is transmitted, what percentage is voice or video or text, what path it takes through the network. Then they run simulations to assess the impact if, for instance, a carrier starts selling the iPhone, or changes its marketing plan, or moves from 3G to 4G services. Shanker says his company’s models have a horizon of about 18 months. They’re most accurate for the first six months, and the company tends to update them on a monthly basis.

But now there’s worry that the models aren’t looking far enough into the future. “The traffic forecast is becoming the most important factor in providing the quality of experience to the end user,” says Stefano Savioli, head of network optimization and assurance for Nokia Siemens Networks. “What maybe was true the month before may not be true the month after.”

The models don’t simply look at how much traffic a network carries. They also examine whether the traffic is being routed efficiently. They might find, for instance, that one equipment node is operating at only 30 percent capacity, while another is at 90 percent. Too little capacity and you drop calls and alienate customers. Too much and you’re wasting money. Shanker says one European company planned to add five base station controllers to its network, at $1.5 million each, to increase its capacity. Cerion modeled the same network and found that by rearranging how traffic was handled, it could add just one, saving $6 million.

But sheer growth is a problem too, especially when it comes to the demand to carry more movies and music, which together soared from 25 percent of traffic on mobile devices in January 2010 to 41 percent by September, according to Sandvine, a Canadian maker of traffic management tools. On fixed networks, audio and video is now at 43 percent, up from 30 percent a year earlier. Netflix alone accounts for as much as a fifth of all bytes downloaded in North America during peak hours, Sandvine says. In eight months, mobile traffic from social-networking sites like Facebook increased by a third in North America and almost doubled in Latin America. And Morgan Stanley projects that mobile data traffic will more than double each year, rising to nearly 3.5 million terabytes by 2014.

Creating accurate forecasts depends on understanding how different phones and apps interact and how those interactions affect the network. Phone manufacturers make a variety of design decisions about, for instance, how the device will maintain its connection to a network or load an app. As a result, one phone may signal the network 60 percent more often than another running the same app.

Nokia Siemens runs tests to figure out the demands imposed by every combination of phone and app, then feeds those numbers into its predictive models. It also analyzes the packets of data being sent across networks to find out how many users with a particular phone run a particular app. All that helps project how sales of new phones and new apps will affect traffic.

Sheer demand isn’t the only source of traffic problems. Modelers must also account for failures in the network backbone—problems resulting from software glitches, fiber-optic cables that get cut by construction projects, or hackers launching denial-of-service attacks. Sometimes unusual things happen at the worst times. For several hours heading into so-called “Cyber Monday,” on November 29, many Comcast customers in the Northeast lost Internet access. The problem turned out to lie with Comcast’s domain name servers, which translate the website names that people type into their browsers into the strings of numbers that computers read.

Some fixes, like providing extra equipment, might be more reliable but more expensive, whereas adding certain software might be cheaper but might not protect against as many failures. Gordon Bolt, associate vice president of engineering at OPNET in Bethesda, Maryland, says his company’s predictive models can figure out which types of failure are more likely and suggest the most cost-effective combinations of protective measures.

Savioli believes that as more carriers start to understand the impact of smart phones, issues like dropped calls at football games will become a thing of the past. But he says predictive modeling will still be needed when the next new device or an as-yet-unknown application places new demands on the network. “We still have operators all around the world that have not changed what they’re doing, that haven’t accepted the paradigm of the smart phone,” he says. “I do believe that this problem we see now will be taken into account in the future. But what kinds of new problems we will see in the future, I cannot tell you.”

Keep Reading

Most Popular

Large language models can do jaw-dropping things. But nobody knows exactly why.

And that's a problem. Figuring it out is one of the biggest scientific puzzles of our time and a crucial step towards controlling more powerful future models.

OpenAI teases an amazing new generative video model called Sora

The firm is sharing Sora with a small group of safety testers but the rest of us will have to wait to learn more.

Google’s Gemini is now in everything. Here’s how you can try it out.

Gmail, Docs, and more will now come with Gemini baked in. But Europeans will have to wait before they can download the app.

This baby with a head camera helped teach an AI how kids learn language

A neural network trained on the experiences of a single young child managed to learn one of the core components of language: how to match words to the objects they represent.

Stay connected

Illustration by Rose Wong

Get the latest updates from
MIT Technology Review

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

Explore more newsletters

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at customer-service@technologyreview.com with a list of newsletters you’d like to receive.