Monday, 11 March 2013

Netflix - bandwidth piglet

Netflix is notorious (at least amongst ISPs) as a bandwidth hog, consuming 1/3 of all North American internet backbone bandwidth. However, what's fascinating is that when you look at their usage from the perspective of an individual user's bandwidth, rather than the network as a whole, it's actually not overwhelming.

Some ISPs have an incentive to degrade Netflix traffic, since they have their own TV offerings they'd rather consumers bought instead. As a way to stop this happening, Netflix reports the performance it measures on different ISPs This data tells us an number of fascinating things.

1. Low bandwidth per stream
The bandwidth used per stream is low. In developed markets, the average stream is generally in the range or 2-2.5 Mbps. This isn't too much of a surprise, given that this is about typical for an SD TV picture. Whatever might fill up 100 Mbps bandwidth, it isn't likely to be standard TV

2. Performance driven by factors other than access speed
Most ISPs experienced a performance drop of approximately 5% in December (see Finland for example). Clearly this isn't because the access networks were downgraded. Rather it's likely because things got busy for the network (and Netflix) in that period, leading to congestion. This is important because it underlines that a poor experience for an end-user is often not caused by a problem in their internet access. There are many other parts of the system that can get congested

3. Access technology has relatively limited impact
While the numbers show a discernible difference between technologies, it isn't huge. In the US Verizon's FTTH-based FiOS service delivers Netflix streams at 2.10 Mbps. AT&T's U-Verse, based on FTTN, runs at 1.91 Mbps, not far behind. (Even the fabled Google fibre is only running at 3.35 Mbps for Netflix). Clearly the FTTH networks are capable of greater speed, and Netflix adapts to the available bandwidth, but in practice FiOS and U-Verse customers were having very similar experiences. This could be because there was a choke point somewhere else in the system, or the user had capped speeds (to husband data allowances) or the content was already playing at maximum resolution. No matter, the extra bandwidth was giving very limited benefit, even with a bandwidth intensive service like Netflix.

4. DSL compares well
Fourthly, the FiOS speed is all the more striking when compared to UK DSL providers. Of these only one (Everything Everywhere) was materially below FiOS' 2.10 Mbps. Of the other three, one exceeded it and two matched it (if we give the benefit of the doubt to Talk Talk, a whisker behind at 2.04 Mbps).

5 Stream speeds low relative to access network capabilities
These levels of usage (in the houses that actually use Netflix) are relatively low compared to the capability of the access network. Akamai measures average peak connection speeds at 29.6 Mbps for the US and 28.1 Mbps for the UK. While this isn't a strict apples-to-apples comparison to the Netflix figures, which are averages and will peak higher, it does suggest there's ample headroom even in today's networks.

Maybe massively higher resolution video and 3D TV will change this picture, and suddenly Netflix will be using much more? Perhaps. But even 3D TV will only require 12 Mbps, and this won't be widely used for a long time to come, not least because there just isn't that much 3D content yet.

Netflix may be a hog when it comes to the internet backbone. However, as we've seen, when it comes to the access network it's just a piglet.


The above post is as originally written, but MikeyB has pointed out something important I'd missed:

The Akamai average peak speeds are not averaged across lines, but but across IP addresses. Because IP addresses are dynamically allocated, one IP address might be applied to several lines in the course of a quarter (Akamai's measurement period), and Akamai's average-peak measure will only count the fastest of these lines. I'm not sure of the magnitude of this issue - it will depend on how frequently IP addresses are reallocated (particularly for fast lines), but there's no question that it means that Akamai's figures are a less good guide to the capacity of the average line than I thought, overstating it to some extent.

The other figure Akamai publishes is the average connection speed, which is undistorted by the 'peak' issues described above. However, it has its own issues. The average will be lowered by congestion at peering points, contention in the backhaul, server overload, interference with the end-user's wifi signal and various other things - none of these have anything to do with the speed of the access network, and thus this metric understates the capacity of the typical last mile connection.

Akamai's average speed will also be affected by simultaneous use in the home - if more than one person is sharing the connection, then the bandwidth 'perceived' by Akamai will also be reduced. Again, this will cause it to understate the speed of the access network, though may accurately reflect the speed available to a given person on a household.

Thus the capacity of the access network is likely lower than the peak speeds 29.6 and 28.1 Mbps for the US and UK respectively, and higher the average measured by Akamai, which is 7.2 and 6.3 Mbps.

Thus, I think the wider point that Netflix's bandwidth needs (at roughly 2 Mbps) are a long way below the capacity of the access network for the typical user stands. But my thanks to MikeyB for pointing out my error.


  1. Surely the Akamai "average connection speed metric" makes more sense when discussing long-lived isochronous streams like Netflix than the "average peak connection speed" metric?

  2. Hi Rob, broadly agree with the point you are making, but why did you use peak access speed? You are comparing an average with an outlier, a crystal clear logic error. If the average and peak were close it might make sense, but they clearly aren't.

  3. Jason. MikeyB -

    Thanks very much for taking the time to comment.

    I think whether the peak or the average speed is the one to use depends very much on why you're asking the question.

    If you're Netflix designing your product, then the average speed is the one that matters certainly, but if you're deciding whether an upgrade of the access network will help (my question), then I think peak is more relevant.

    My logic is as follows. There are many things that explain the difference between Akamai's peak and average speed. The average will be lowered by congestion at peering points, contention in the backhaul, server overload, interference with the end-user's wifi signal and various other things. One thing that won't explain the difference (generally) is a variation in the capability of the last mile. Fibre or copper in the last mile has a relatively fixed capacity, unlikely to vary much over the course of a day or week. (I generalise - there will be exceptions).

    Thus Akamai's peak speed is, if you like, a measure of the last mile capability, when it is the weakest link in the chain.

    So, if you want to compare Netflix to the access network, it's the peak speed you need. Ideally (as I mentioned) you'd want to know what Netflix's peak demand was, not just their average, but they don't publish that.

    I'm open to being convinced on tbis, but that was my logic.

    1. Hi Rob,

      The way Akamai calculate peak speed is per IP not per line (as there is no real way to measure that). Taking the Australian market, the top 2 ISP's only offer dynamic IP's and the third one charges extra for a static IP. This means over a quarter there is quite a high chance the same IP will be linked to different lines. Because the peak is simply the fastest speed per quarter only the fastest line that IP is linked to will be recorded. I imagine the US and EU markets would be similar (as they all seem to have broadly less Unique IP's than subscribers.

      If there is any skew in that data it is likely to be towards the fastest lines, as in DSL there is a strong link between a slow line and dropouts (which is what causes a new IP to be assigned).

      I think in your list of things that effect the average you have left out an important one, the number of people in the house. If a connection normally has 3 people using it a peak speed when only one person is home isn't really representative.

      I think average peak represents an outlier situation, the best line the IP is attached to that quarter, when the least people are using that line that quarter.

    2. +MikeyB

      Sorry to be slow responding, I've been travelling.

      You make a really important point - I had not picked up that the Akamai numbers are based on IP addresses. (Why do they not do it via cookies, I wonder?) As you say, this may well distort the picture - if a particular IP address happens to spend just a few hours in the month allocated to (say) a 100 Mbps line, that will be what sets the peak speed for that IP address. I will edit the post accordingly.

      On the point about several people using the connection simultaneously, I agree with the technical point, but I suspect its impact on overall averages may not be massive - not least because (roughly) two-thirds of households only have one or two people in them.

  4. This comment has been removed by a blog administrator.

  5. This comment has been removed by a blog administrator.