Data caps have become the first line of defense of mobile operators against network congestion. But it is becoming increasingly obvious that imposing limitations on traffic allowances is not an effective tool to manage traffic, and in fact it may have unintended negative effects on congestion (more on this later).
As we argued here, sufficiently generous data caps are necessary to protect networks from abuse, but they usually affect only a very tiny percentage of subscribers. In the US, for instance, 5 percent of users account for 68 percent of traffic, according to Sandvine. Still, stringent data caps that apply to all subscribers risk to alienate many of them, because most do not know that they do not fall in the 5 percent high-usage group.
Most importantly, operators want to drive up subscriber adoption and satisfaction, and that inexorably means more traffic. Other things being equal, more traffic means more base stations to increase capacity, and therefore more capital and operating costs, and, eventually, lower profitability. Clearly this is an unsustainable model in the current situation where traffic is increasing very fast, but ARPUs are at best stable.
Uneven traffic distribution
But does it have to be this way? In our analysis of traffic growth in the report mentioned above, we identified multiple drivers to traffic growth that bring uneven contributions to the overall traffic profile. To manage traffic effectively, we argued that addressing different traffic growth drivers independently, with the tools that are more appropriate to it, allows the operators to decompose the traffic load issue and make it more manageable. For instance, with regards to location, traffic is increasing both in high-density public areas and from home, but the increase in capacity density that is required is likely to come from different topology solutions--e.g., small cells will help in public areas but not in residential areas, where femto cells are better suited.
Luckily for mobile operators--and for subscribers--there is quite a lot of room to improve the utilization of mobile networks without any costly equipment upgrade or addition. This is not to say that mobile operators can avoid investing in network expansion--which is clearly necessary--but simply that that are way to use the existing infrastructure more efficiently that may reduce the scope of the expansion needed (or at least let subscribers have more fun).
Inevitably, no network, either fixed or mobile, is used at 100 percent capacity (see a previous column I wrote). Networks are dimensioned to cope with peak traffic demand, so during non-peak hours they typically carry less traffic than they have capacity for. In mobile networks, the time-of-day distribution of data traffic load shows a clear peak in usage around 9:00 p.m. and 10:00 p.m., with the lowest level of usage in the middle of the night as expected, where traffic load can be one third or one fourth of peak traffic (Figure 1). This distribution is fairly constant across countries, and very similar to wireline time-of-day traffic distribution.
What if we could flatten the time-of-day distribution, so that the idle network resources in the middle of the night can be used more extensively? This will have the obvious effect of increasing the overall network throughput at a almost-zero marginal cost to the operator. Capex and opex depend on peak traffic, and traffic during non-peak hours is virtually free.
The benefits are obvious, but so are the challenges to any attempt at flattening the time-of-day distribution. Subscribers want to use their data plans when they want, and they will not stay up all night simply because their data connection may be faster. That is why we have peak hour in the first place: people tend to like to congregate in few places, and do similar things at similar times. So is it hopeless to try increase traffic load in non-peak times?
The BitTorrent approach
Enter the scavenger app.
Eric Klinker, CEO of BitTorrent, very aptly uses the "scavenger class" term to refer to applications like file sharing which have low priority requirements--so low, in fact, that they can fit below the best-efforts class. File sharing is one of the major culprits of network congestion accounting for 30 percent of mobile traffic (Figure 2) according to Allot Communications--although video streaming is becoming the worst offender with 37 percent of traffic--and BitTorrent is a major contributor, accounting for 21 percent of mobile data traffic in North America according to Sandvine. In wireline networks the percentage of file-sharing traffic is even higher. BitTorrent realized that it needed to act from its end to reduce the impact of the massive amount of traffic its clients generated. It developed a transmission protocol that enables BitTorrent traffic to be carried only when the traffic load on the network is sufficiently low that the additional traffic does not slow down other active applications.
Beyond best efforts
In both fixed and mobile networks, transmission control protocol (TCP) is the dominant protocol to transmit data, as it has error correction which ensures that packets are correctly received, and its own congestion control to provide reliable transmission in high-traffic environments. All TCP traffic competes for the same resources and, although QoS can be used to prioritize traffic, it still manages all traffic within the same framework, with the lowest class, best-efforts, including most data applications, with the exception of real-time ones, like voice or video streaming. BitTorrent's micro transport protocol (uTP, or μTP) creates the "scavenger class" to transmit data below the best-efforts level, using user datagram protocol (UDP), which is simpler than TCP and lacks its own congestion control. uTP continuously measures the delay in packet transmission to estimate congestion. If there is no congestion, it transmits data; if the traffic level is too high, it graciously waits until congestion clears.
The BitTorrent approach to congestion is a very compelling model that can be extended beyond file-sharing applications, to include software upgrades or content downloads that do not have to happen in real time (e.g., downloads of video content that can be downloaded overnight, and stored to be viewed during the day). Applications that can send low-priority traffic during non-peak time can substantially spread out network utilization and effectively increase network throughput by reducing peak traffic in a way that is cost-effective (free to mobile operators, and relatively cheap to application providers), and does not have a negative impact on subscriber's experience. In fact, the lower traffic load during peak hours may result in an improved subscriber experience as the impact of congestion is reduced.
A low hanging fruit
If this is such an effective and low-cost approach that keeps subscribers happy, why is it not more widely used? There are many reasons, among which the fact that there is no standardized solution that can be adopted by myriad of application developers in the fragmented mobile market. But perhaps even more importantly there is little incentive for subscribers to use (if available) or demand (if not) scavenger apps--all their traffic equally counts towards their monthly allowances, so why bother being good citizens? And how would they need to play nice in the first place, when operators do not openly acknowledge that congestion is an issue they are facing and instead prefer to argue about who has the fastest 4G (sic) network?
Unintended effects of traffic caps
This is exactly where traffic caps fail. By making all traffic equal, regardless of time, place, or network traffic load, operators encourage data thriftiness among subscribers. Instead of using their network efficiently, by for instance scheduling traffic downloads at off-peak hours, they try to limit their use to the more important applications. And the data usage that is more important to subscribers tends to be during peak times and, as a result, data caps may encourage a time-of-day distribution that is even spikier than it is in an unlimited data plan regime.
Plans without punitive data caps, or data caps that exclude peak hours may instead reduce peak-time traffic, as they encourage subscribers to download what they may want ahead of time, without worrying to go beyond their monthly allowance. For the operator, the small reduction in peak time traffic is well worth the increase in traffic during off-peak hours. For the application developer, having to keep network congestion into account is additional work, but also an opportunity to differentiate apps, and add value. And the subscriber no longer has to worry about maxing out the data allowance before the end of the month.
Monica Paolini, PhD, is the founder and president of Senza Fili Consulting and can be contacted at [email protected]. Senza Fili Consulting provides expert advisory services on wireless data technologies and services.
Figure 1. Network aggregate traffic profile. Source: Sandvine
Figure 2. Mobile data usage split by application type. Source: Allot Communications