We all know this scene from countless movies: The action hero is swerving his car at breakneck speed through the streets of an exotic city while bad guys shoot at him. Cowering in the back seat, the hero’s teammate is desperately downloading the last hundred megabytes of a file that will save the world. The message is clear: Saving the world requires an unlimited mobile data plan.
Meanwhile, back at the lab, Ratul Mahajan and teammates Aruna Balasubramanian and Arun Venkataramani are trying a new strategy for augmenting cellular networks that could make mobile data more affordable for ordinary consumers. Mahajan, a researcher with the Networking Research Group at Microsoft Research Redmond, started the VanLan project, which focuses on enabling cheap, high-throughput wireless connectivity for moving vehicles. Balasubramanian and Venkataramani, researchers at the University of Massachusetts Amherst, have been exploring different wireless-enhancement techniques. In a paper titled Augmenting Mobile 3G Using WiFi, presented in June 2010 during the eighth annual International Conference on Mobile Systems, Applications and Services, the team describes a novel system called Wiffler that leverages the presence of Wi-Fi networks to offload cellular traffic.
Thanks to smartphones, wireless apps, and consumer demand for “anytime anywhere” Internet access, mobile data usage is growing exponentially. Industry analysts predict mobile data usage doubling each year over the next five years, with demand exceeding capacity within the next few years.
“For a good mobile data experience,” Mahajan says, “you need the near-ubiquitous coverage that cellular provides. But cellular data is expensive, and growing demand has forced network operators to conserve capacity by pricing data plans which cap bandwidth usage.”
Balasubramanian and Venkataramani initially collaborated with Mahajan in the summer of 2007 on a project for vehicular Wi-Fi. This scheme, which they called Vi-Fi, used network hopping to switch across short-range wireless networks.
“In the summer of 2008,” Balasubramanian says, “we started another project to improve vehicular access with Wi-Fi but soon realized that using Wi-Fi networks to augment 3G made much more sense than trying to achieve complete connectivity with Wi-Fi networks alone. This was around the time that 3G networks were becoming popular, and we anticipated 3G becoming heavily overloaded because of its limited spectrum—and this was even before the problems in 3G began emerging.”
A moving vehicle sometimes travels through areas where Wi-Fi coverage overlaps with 3G. In these areas, it should be possible to offload traffic from more expensive 3G networks onto Wi-Fi, then switch back when Wi-Fi becomes unavailable. This was the premise behind Wiffler.
But first, the team had to determine whether the overlap of 3G and Wi-Fi networks in urban areas was sufficient to make this scheme viable.
“For the idea to be feasible,” Mahajan explains, “there should be enough Wi-Fi capacity to usefully augment 3G. Since no one before us had done a joint survey of 3G and Wi-Fi, we had to conduct such a study ourselves. Previous work—including our own—had been 3G-only or Wi-Fi-only. Our goals now required a joint survey.”
This led to the first surprise of the project: The survey showed a negative correlation between 3G and Wi-Fi availability. Wi-Fi was more likely to be available in areas without 3G coverage than in areas with 3G coverage. In areas where 3G was unavailable, Wi-Fi was available roughly half the time; thus the combination is more available than the two independently. This implied greater synergy in combining 3G and Wi-Fi than anticipated before the experiment.
“However, the survey also raised a concern,” Mahajan recalls. “We found that overall 3G access is available about 87 percent of the time, while overall Wi-Fi access through open access points, such as free Internet hot spots, is available only 11 percent of the time. Plus, where Wi-Fi is available, it exhibits lower throughput and higher loss rates than 3G.”
At first glance, this suggested the maximum amount of 3G data that could be offloaded was only 11 percent—and that might come at the expense of application performance.
Unlike in the movies, though, not all data is real-time or mission-critical. For example, email and file transfers of non-world-saving urgency can tolerate some threshold of delay, or latency, without significant impact to user experience. This fact informed the team’s approach to designing the Wiffler system, which overcomes Wi-Fi availability and performance challenges by trading off application latency for 3G usage. Its two key ideas are leveraging delay tolerance and fast switching to 3G.
Instead of transmitting data immediately for applications that can tolerate small delays, Wiffler uses prediction-based offloading, waiting until Wi-Fi access is available, then shunting the data onto Wi-Fi networks. The system does this only if 3G savings are expected within the application's delay tolerance. If actual Wi-Fi performance proves unable to transmit the data within a required time window, Wiffler quickly switches back to 3G, minimizing impact to the user.
Devising a way to predict the likelihood of encountering Wi-Fi while driving proved fairly straightforward.
“Our prediction of Wi-Fi encounters is based on the observation that access-point meetings occur in bursts.” Mahajan says. “If the mobile node meets a Wi-Fi access point frequently because it is in a dense urban area with lots of hot spots, then the node is likely to meet the next access point within a short time interval. Similarly, if the mobile node has not met an access point for a long time because it is on a highway, the node is unlikely to meet an access point anytime soon.”
Armed with this observation, the researchers calculated the frequency of access-point encounters using a simple predictor that computes the average time between access-point encounters, based on the time between recent past encounters.
“Prediction-based offloading provided a way out of the problem of having only 11 percent Wi-Fi availability.” Mahajan explains. “As we implemented and evaluated this idea, we were excited to discover that with maximum delays of less than one minute, Wiffler could offload half the 3G data despite low Wi-Fi availability. This was a really significant finding.”
Before the team could be certain of its results, however, it needed to answer an important question: Were the findings representative? In other words, were the results specific to a particular urban environment, or was it a valid general finding? To examine this issue, the team set up test beds in three cities: Seattle, San Francisco, and Amherst, Mass.
The most challenging aspect of conducting experiments in three locations was ensuring that the data collected was consistent. The measurement study required shipping computers across the country and dealing with the inevitable problems that arise when setting up remote experiments. They faced equipment damage, as well as difficulties with remote setup that required sending equipment back to the Redmond lab for more testing and configuration—an onerous but vital process, Balasubramanian recalls.
“We had to verify that the data we collected was not biased in any way due to a faulty measurement setup in the vehicles and logging software,” she says. “This was especially difficult because we had a lag between when measurements were conducted and when we received the results. We were constantly refining the setup and software. We also spent nearly two months running alternate experiments and conducting hypothesis tests to ensure the validity of our results.”
But the effort proved worthwhile, and the researchers were in for a pleasant surprise. Findings from each city were amazingly consistent, despite key differences in their traffic characteristics. Amherst is a college town, and while both Seattle and San Francisco are large cities, San Francisco is much denser than Seattle. Furthermore, in Amherst, the experiments were based on buses with preset routes, while in Seattle and San Francisco, the experiments were based on personal cars. Finally, the experiments in Seattle included long periods of freeway travel.
“Under such a diversity of conditions,” Mahajan recalls, “we expected the experiment’s results across cities to be very different. But, in fact, we recorded similar results from all three, which gave us confidence in the generality of our findings.”
When asked about the amount of time team members spent in the back of moving vehicles, Mahajan laughs.
“Yes, for part of the experiment, we were at the back of the Microsoft Connector bus to and from Redmond, learning about the characteristics of real-life workloads generated by real people on the move. At other times, I was driving around Seattle with the gear plugged into my car. I also arm-twisted friends in San Francisco to do the same. The Amherst experiments were more systematic. Leveraging the University of Massachusetts’ DieselNet project, we set up several buses with gear we controlled remotely.”
The team sees direct benefits to network operators from Wiffler because the system reduces the stress on cellular networks. Users also can benefit, because Wiffler reduces cellular-bandwidth usage—an important consumer issue now that network operators are structuring usage-based mobile data plans. Mahajan also notes that while their experiments took advantage of free Wi-Fi hot spots, in a real-world deployment, Wiffler could make use of paid hot spots, which increases Wi-Fi availability and also offers better Wi-Fi performance.
When the Wiffler team started the project in the summer of 2008, many in the research world did not believe that cellular-spectrum shortage was an issue. This was partly because cellular companies were offering unlimited data plans to consumers. But looking at the trends, the VanLan researchers were convinced that network operators would soon need to make more efficient use of existing spectra. By the time they published their results in June 2010, most major wireless operators were announcing changes to mobile data plans—which made Wiffler a timely piece of research.
Mahajan sees value in bringing other techniques, such as caching and pre-fetching, to further reduce the amount of data that needs to be transmitted over 3G in the first place. But he feels that Wiffler represents a culmination of the VanLan project.
“We have looked at numerous ways of connecting moving vehicles, and in each case, we pushed the envelope of what is possible. I feel it’s time to take a break from vehicular networking and focus on other things. But I love cars and motorcycles too much and suspect that I’ll be back to this again soon.”
Flash forward: rush hour on the I-5 freeway in Seattle. A car swerves into the HOV lane. In the back seat are networking equipment and a slightly carsick Microsoft Research scientist. He is downloading a file that will be analyzed for ways to make the world safe for low-cost, unlimited cellular data.