Netflix Performance on Top ISP Networks

I have a 50 megabit plan with Rogers and all my SSL traffic throttles the entire line with an exponential dropoff to 25 Kilobytes a second. The packet loss is severe and persists for all packets even HTTP. It takes a long time sometimes an hour for the connection to recover. You can imagine how many applications this would impact (syncs with large files, P2P, Dropbox make the Internet virtually unusable).

I don't speculate on the matter, but here is some more information.

http://www.christopher-parsons.com/blog/isps/rogers-network-...

http://torrentfreak.com/rogers-bittorrent-throttling-experim...

There are a few elements to it that one could use to dispute its value — for example, Netflix has no way of knowing what level of service a customer has paid for. They also don't appear to differentiate between different types of service from a single ISP (e.g., they lump together AT&T's old ADSL customers in with ADSL2/U-Verse, and, presumably Verizon's DSL with FiOS). I would hope that they are also filtering for sessions where a higher bitrate is requested, which Netflix's Silverlight app doesn't do unless the window/screen is large enough to benefit from the extra resolution.

Of course, one could argue that those differentiators don't ultimately matter, and that if its customers are opting for older or slower service levels, that can tell you something about the ISP's pricing competitiveness.

I gotta nitpick the graph though, it's incredibly hard to figure out which line is which ISP (too many lines and similar colors). A regular bar chart with the ISP names under each one would have worked fine (since the avg. bandwidth didn't change much over the course of the study).

Yeah, we are getting raped either way. It's pretty much impossible to win this argument as a consumer, even if Netflix wins their with the ISPs.

I used to be with Teksavvy where 200 gb/ month was the norm. I just might have to convince my g/f and get her to switch.

Just to compare, Rogers is $59 for 60 gb/ month, Teksavvy is $29.99 for 200 gb.

As a baseline, all of the ISPs on this list that I've used (ftth, coax and dsl providers) have been able to push to 90%-100% of quoted line speed 99% of the time.

What this chart shows is netflix performance, which is quite different than general network performance. What goes into netflix performance? Total netflix service load, CDN pop load, CDN circuit size, SLA/QOS terms, the choices the netflix client makes about adaptive bitrates and shaping the provider applies to the traffic.

Take a look at the US chart and note the correlations in spikes and dips across all providers at certain times - 10/21 peak, 10/23 valley, 11/13 valley etc. These are indications of aggregate stream load netflix wide and are an indication that netflix manages their peak throughput to be pretty close to their max throughput (as well they should). But it's also a strong indication that if 10% of netflix customers switched to a competitor overnight average stream performance for netflix would jump noticeably, perhaps even 10%.

How their adaptive streaming works is another huge factor. First they measure your current performance to the CDN POP and then explicitly chooses to use only 60% of that[1]. So add 40% to all the ISP numbers out of the gate. TCP+HTTP overhead is another ~4%. Downloading something else while you stream? If the streaming sees more than a few headroom faults (ie 1000ms worth of data takes more than 600ms) the system will quickly adapt down in bitrate trying to do its best to never hard fault (buffering). It likely is considerable more conservative in moving the stream speed back up. Customer using 802.11b? Max throughput there is ~5mbps, so netflix will use a max of 3mbps. In an urban area on a congested channel? You may only see 1-2mbps at times. Microwave on? etc.

Lastly, with the news that netflix is 20% of ISP traffic at night the network providers have every incentive to manage that traffic down. Look at the best US performer at ~2700kbits. The fastest HD stream is 4800kbits - Nearly twice the speed. Worst case if an ISP allowed full rate streaming they might increase their peak network load by 10% or more.

Who was it that announced netflix is now 20% of night time traffic? Our old friend sandvine, provider of the DPI gear made famous by comcast using it to throttle bittorrent. This tells us two things - 1) sandvine is on net at at least a few major broadband providers and 2) by announcing that number they are implicitly telling network managers that they can help reduce that traffic.

Netflix makes it very easy to do fine grained control on bitrates. If you are watching the HTTP headers you know exactly which customers are on which stream rate without having to count the throughput. Most users won't notice a QOE change from 4800kbits to 3200kbits. Want to cut your network traffic? Use DPI to identify the higher rates and label them bulk-toofast. Then either explicitly rate limit them, give them a low queue priority or simply route the acks through a big buffer under performing router or indirect route. Once the adaptive bitrate slows down to a rate you like switch the label. The more subtle of these methods are much, much harder to detect than the comcast RST's were, and wouldn't technically violate any net neutrality rules in place.

[1] http://www.breakingpointsystems.com/community/blog/emulating...

edit: Ah well, I guess that will do, can't make them stand up in a column.

I hasten to add that part of this must be a function of competition here. I can choose from Comcast, DSL or AT&T's U-verse thing, so they have some incentive to provide a compelling bandwidth offering. Still – at least where data service is concerned, Comcast delivers.

By showing that they care ALOT about what their users care about, content delivery performance, they demonstrate credibility, a key component of trust.

Consistent trust building exercises such as these does more than a temporary sale ever could for long term customer loyalty, even if it can’t be perfectly measured.

I would like to see more companies try the same, even if they can’t prove the ROI on a spreadsheet before hand.

Why use a spreadsheet when you have a perfectly good gut you can check to see if it is the right thing to do?

p.s., in case you missed it -- 21 Dec FCC ruling: http://www.thedailybeast.com/blogs-and-stories/2010-12-21/ne...

If we were then to rank ISP's on throughput, even though a rural ISP is serving a critical and under-served part of our community, they would be rewarded with negative perception.

On an absolute scale, it's so-so customer service, but on a relative scale it's amaizing how many things they don't do wrong compared to other ISPs I've worked with (Comcast in particular)

Almost anytime of the day, I can saturate my DSL line's download bandwidth from fast sites such as Amazon S3. But Netflix doesn't always stream in HD even though my link is faster than what's required (4800 Kbps).

I suspect the bottleneck is at Netflix's end.

So the 15 Mbps is way above anything in the Netflix graph.

So, the ISP can't always be the bottleneck keeping data rates below 3 Mbps.