BSG: Technical Perspectives on Net Neutrality
Broadband economics, traffic management and "Net neutrality 2.0" are big themes at the upcoming Telco 2.0 events in Los Angeles on 27-28 October and London on 9-10 November. We're delighted to be able to share new analysis and use cases from the Telco 2.0 team, Bain & Co, Ericsson and Analysys Mason at the events, as well as hear from the Group CTO of Deutsche Telekom, the Chairman of Project Canvas, and others.
Our in-going point of view is described in our submission to Ofcom here. As additional context please see below a detailed write up of the recent Net Neutrality conference in London organised by the Broadband Stakeholder Group detailing the experiences of the BBC, 3UK, Ericsson, Cisco, and including valuable data on network costs..
One of the things that struck us about the shindig was that in fact, net neutrality and its opposite weren't really top of the agenda.
The event fell neatly in two halves - geeks before lunch, policy wonks after lunch. The news from the geeks was relatively reassuring - although video demand is storming ahead as predicted, the networks are doing reasonably well in soaking it up. Anthony Rose, the CTO of Project Canvas, reported on the BBC's experience of streaming the World Cup - at the peak, they were pushing out 800,000 concurrent unicast streams, with a peak throughput on the iPlayer CDN of 450Gbps. He predicted that the London Olympics in 2012 might achieve a peak of ten times that.
The good news is that Canvas is taking shape. At the beginning of the project, they had faced "huge" standardisation problems dealing with multiple incompatible interfaces from the TV vendors, but the latest generation of new TVs have Web browsers and support Flash. Notably, that includes the Android-based Google TV.
A key element in Canvas is the widespread use of IP multicast, a technology that the IETF standardised years ago (the first RFC is from 1994, but the original Internet-Draft proposal dates from 1989) but that is still surprisingly under-deployed - it lets client computers subscribe to a group, in which content is shared between them. This means that the content stream is only multiplied between the subscribers once it reaches the multicast group, analogously to a content-delivery network - a major bandwidth saving. The multicast groups, further, will be drawing their content from a mammoth CDN provided by BT.
Rose made a good point in that we need to avoid taking decisions that will constrain our choices further down the line. Specifically, ISPs that decide to be "non-neutral" (or should that be belligerent?) will encourage non-preferred applications both to encrypt their traffic, and to use common ports in order to hide in plain sight. The problem here is that using SSL on port 443, for example, makes it very difficult to cache video or other heavy content locally and therefore makes the problem worse. Similarly, the shift of much file-sharing from P2P protocols to Web-based direct download, driven by efforts by rightsholders to go after BitTorrent, has had the effect of making this content harder to shift in terms of hardcore packet pushing. (Simon Drinkwater of Limelight Networks produced a very nice chart of the shift later in the day.)
Canvas will eventually consist of a set of specifications and software libraries, almost certainly based on Linux and Flash, and a common infrastructure, which is itself made up of the BT-managed CDN and BT-managed private peering with major broadcasters and content providers to get the content into the CDN. Rose expects it to make heavy use of edge-caching - predictively downloading content during the low-usage phases of the day to a large local hard disk, based on your requirements or possibly on your usage profile.
Brian Williamson of Plum Consulting brought in a rare and valuable commodity - actual data on network costs, which revealed a couple of interesting points. First, the performance leap from WCDMA to HSPA+/LTE is impressive - it seems to be a feature of radio network technologies that they disappoint in the beginning and outperform in the long run.
Second, the real difference between access technologies is between the ones with a steep cost curve and the ones with a flat cost curve. UMTS and LTE all have a steeply rising capacity-cost curve - the only difference is that the curves shift out to the right over time. Copper is flatter, and fibre's cost curve is the flattest of all, being dominated by the original trenching budget and thereafter influenced by the abundance of bandwidth, the lack of street-cabinet electronics, and the relatively low cost of upgrades. Analysys Mason estimates the cost of serving a user with fibre at 1GB/month as €46, and the cost of serving a user with fibre at 100GB/month as €49.
The unavoidable conclusion is that although fibre pricing needs to be carefully thought out, in order to both upsell the heavier users and reach the low users, there is little or no case for "non-neutrality" on fibre access networks. However, Williamson pointed out, mobile operators needed to impose "cost-based pricing". That said, it looks like the cost curve for HSPA+-and-beyond is rather kinder than that for DSL. Theoretically, the decision-rule is given by the intercept between the cost curves for different technologies - the advanced wireless curve crosses the DSL curve at a monthly demand of 30GB, and the fibre curve at 45GB.
In practice, of course, serving more than a very few 45GB/month users with LTE would require a lot of spectrum and probably so many Gigabit Ethernet backhaul feeds that you'd need the fibre access network anyway to get the cells deployed. But it's certainly an interesting data set, and one that tends to validate a hybrid fibre-wireless approach to NGA.
More controversially, Williamson also argued that in general, it would be better to adjust pricing than to prioritise traffic. This was something of a theme for the technical participants - there was a general sense that being "non-neutral" might turn out to be more than a bit of a disappointment, that it had important costs in itself, and that there were valid technical solutions for many of the industry's problems that were less controversial.
John Cunliffe, CTO of Ericsson UK, took up the same theme - perhaps not surprisingly from a network vendor - reviewing the short-term technical possibilities. He remarked that there was in practice no limit on capacity once fibre was in place, and that it was therefore a choice on the operator's part which technologies to deploy and quantity/price combinations to offer - once the fibre was in place.
He mentioned bonded DSL but warned that relatively few customers' copper was in sufficiently good condition to make use of it, and said that wireless would anyway overtake copper very soon. He also presented data on the coverage boost from using CPE devices with external antennas, which is significant and another point in favour of fixed-wireless hybridity.
He remarked that deep-packet inspection is a CPU intensive task, and therefore costly.
Router design and operation is fundamentally about processing as many packets as possible in the dedicated hardware, without troubling the much slower CPU and software logic - some operations engineers use the number of cache misses (events when the hardware cache contains no matching entry and a packet has to be processed in software) as a performance metric. This should be a very serious consideration.
Cunliffe predicted that capacity would easily keep up with demand, but that prices for end users might rise in the short term.
Simon Drinkwater, who we mentioned earlier on, gave some details of LLNW's work and estimated that CDN traffic was growing 18% year on year. He referred to the problems of handling "planned and unplanned crowds", and said that he expected content providers to pay operators for preferential service, just as they did CDNs today. Interestingly, he also gave the impression that LLNW was increasingly interested in content management, analytics/reporting, transcoding and the like rather than pure CDNing - which might make sense if they expect their customers to be paying operators out of the budget their existing business comes from.
Philip Sheppard, director of network strategy at 3UK, introduced a data-heavy presentation which began by defining the constraints on coverage, capacity, and speed. He showed that the demand for mobile data had hit an inflection point in December, 2007, and had since been growing rapidly until it was now 20 times greater than voice in capacity terms.
3UK had made selling mobile broadband a priority, which had resulted in quality being a major challenge. Customer satisfaction ratings had fallen sharply in the second half of 2008. Interestingly, satisfaction with 3UK and T-Mobile UK starts to vary together very soon after the creation of their network-sharing joint venture, although it's not been long enough to tell if this is also true of Orange.
They had addressed the challenge through a combination of network-sharing, capital investment, pricing, and traffic management. Between 2008 and 2010, the network had gone from 7,500 Node-Bs to 12,500, and from 90% to 98% population coverage, while upgrading the air interface from HSDPA 3.6Mbps to HSPA 7.2Mbps.
He stated that one of the biggest differences between traffic classes was sensitivity to latency, and that one of the main motivations for investing in LTE would be the promise of getting typical round-trip delay from 120ms down to 30. There was a trade-off between responsiveness for Web browsing and raw bandwidth.
Because they "had to say something about traffic management", he described the breakdown of typical data traffic in the busy hour. 41% was streaming, 38% Web browsing, 7.6% SS-7 signalling and other network-infrastructure activity, and 6.4% P2P. In the busy hour in the busiest cells, 3UK is rate-limiting the P2P - and its own signalling traffic! Significant volumes of the P2P consisted of Microsoft Windows Update.
He introduced some data on the time profile of various device classes. Dongles peak at 9 p.m, while the iPad has a significantly more even distribution over the day with a faster ramp-up in the morning. However, like the iPhone, it was a significant source of extra signalling traffic. He stated that the data network was idle 60% of the time, and that this could be used to deliver content to Project Canvas or similar STBs as part of a two-sided business model.
Summing up, he stated that the cost of technology upgrades was relatively trivial (being largely software), and that since the network-sharing deals with T-Mobile UK and Orange, sites and backhaul were much less of a problem. Incremental costs were now primarily determined by spectrum.
When Telco 2.0 spoke to him later, we asked how much he thought 3UK would gain given a completely free hand to prioritise, block, and rate-limit traffic. He reckons that they might succeed in extending their upgrade cycle by 10% - a bit more than three months in the three-year accounting life of the equipment. He also stated that he didn't consider voice to meaningfully cross-subsidise data - the additional costs of switching, numbering, interconnection, billing, and 999 made it a relatively expensive service to deliver.
Next up was Dominic Elliott, a consulting systems architect at Cisco Systems, who discussed some of the engineering challenges involved in delivering TV-scale video over the Internet. He made the point that video has the curious property that errors are both very obvious - the picture freezes or pixellates - and also recondite and difficult to debug. Broadcasters typically used a few dedicated, private TDM circuits to link their studios with the transmission towers - problems on one of them were immediately obvious, but this wouldn't be the case on the Internet.
He remarked that he was personally very dissatisfied with one well-known IPTV provider's service, although he thought that the CPE was part of the problem. In general, advanced CPE for TV was a difficult problem - so-called "set top boxes" tend to actually be "under the telly boxes" with restricted cooling, exposed to dust, dogs, and children. However, it was important to provide the end user with options, especially if prioritisation was considered. The technology existed for the user to make their own decisions on this.
He remarked that Cisco's main priorities here were quality metrics and reporting, distributed systems, integrating CDN and broadcast, and traffic management. Interestingly, he suggested that traffic management included a large element of operations management - coping with disasters, as he put it - rather than prioritisation, DPI, etc. It would be very important to get real-time performance data. He pointed to a boom in direct peering with content providers and a tendency for CDNs to become more distributed and closer to the user.
He finished by remarking that capacity was solved, but that both the core and the edge of the network needed to be more intelligent. He warned that users expected TV to be very reliable, and that the cost of support would be a real issue.