[ipv6-wg] Youtube over IPv6!

On 12 feb 2010, at 11:03, Peter van Eijk wrote:

> Interesting discussion, and some new hypotheses.
> I am doing some research on the side to figure out how to best measure IPv6 progress.
> 
> From the quantitative angle, the size of your network hardly matters. A six fold increase (sic) is still a six fold increase. At CSBnet is goes from 1 Mbit/sec to 6 Mbit (roughly), at DE-CIX it goes from 0.3 Gbit to 1.8 Gbit.
> 
> More interesting is the IPv6 fraction. At DE-CIX this has risen six fold to 0.16% of total traffic (roughly), while at the AMS-IX it is steady at 0.17%.
> 
> As the total traffic levels at these exchanges are comparable (around 0.6 Terabit/sec avg) I have the idea that we need two hypotheses.
> As Marco suggests the AMS-IX may have traditionally carried a lot of IPv6 netnews, which explains its early lead. My other hypothesis is that the youtube traffic is privately peered outside of the amsix statistic, because Google has datacenters in the Netherlands (I have other research to support that). 


Hypothesis 3: 2002::/16 is routed as one block, now this is anycast but it's highly likely _all_ traffic for this block coming from one datacenter ends up at the same path, given the distance between AMS and FRA I don't think much people would care where it ends up.

And yes a sixfold is a sixfold but it greatly depends on user behavior. Youtube is interactive, so it causes traffic burst, normally these would flatten out against other users, but with low numbers of users these spikes all of a sudden do show up. Next to that youtube only needs a very limited amount of bandwidth compared to other protocols such as usenet and p2p who are happy to eat away whatever they can.

Martin reports he has around 1000 ~ 1500 customers, our tunnelbox shows similair numbers. His traffic is a couple of megabits, our box rarely drops below 100 mbit/s. Now I'm not making any judgement who is better or has a bigger one, but it clearly shows that it is not as straighforward as one may think. I happen to have a usenet box in my network which is on IPv6, thousands of residential users will take care of the rest. I guess the youtube traffic is there, it must be there, but I haven't seen any. Even on my own link at home it won't show up because those little peaks on the graphs get blown away by other traffic and since I'm the only user I know which is which :)

If you want measurements, measure at the end points that is the only reliable method. Make sure everybody measures the same, using the same set of counters (either byte counters or flow sampling) or find a way to correct your data between the 2. Make sure averages, peaks and other oddities are represented in the same way and based on the same time interval or make sure to correct for these as well.
Keep in mind not everybody has the same view of the world, especially since one of the bigger IPv6 content sources only returns AAAA to selected parties, others like ourselves don't run dual stack at all, the published address for our usenet box is A only, the v6 hostname is published in other places aimed at people who know what they are doing or looking for.

And the list goes on and on, for instance due to OS weirdness (or shall I say brokenness) probably half of the traffic I could get over IPv6 comes in over IPv4, this solely relies on whatever DNS query returns first. Accoording to reports on nanog, DECNIC just stopped responding to queries from 6to4 hosts, this might show a drop again in the German graphs.

Now don't get me wrong, I like statistics but please make sure they are correct and even more important make sure to draw the correct conclusions from them and don't have people shopping for the results what they want to see. Google just made another great leap forward by dual stacking youtube, but it's still a very small step for mankind.

Maybe, trying again to avoid opening the can labeled 'sample interval', we as a group can try and come up with a model on how to measure IPv6 deployment:

Should we look at the public graphs and if so, how can we correct these figures to actually be meaningfull ? Or are we better of measuring at the end points and how are we going to convince people to share their numbers ? What makes a better representation of the situation, the amount if traffic or the number of hosts actually being dual stacked ? Should we consider all traffic or look specific protocols ?

I personally think we should look at numbers of hosts, how big a percentage of the users have IPv6 (native/6rd/tunnel/6to4) and how man hosts are out there who have both A and AAAA records asociated (and you might want to test for *.ipv6 and www6 as well). Traffic figures are nice but don't mean anything, if you take a close look at the AMSIX history you will find at least 2 occasions where you see a significant drop in v6 traffic, one only lasted about a day and was caused by 2 usenet providers having trouble at the same time. The other one lasted for months and was due to 1(!) usenet feed being moved back from v6 to v4 causing traffic to drop from 0.1 % to 0.0 % again. I've seen that graph show up in a number of publications.

In the end we can expect the v4 world to keep on working, for me the big question to answer is "how far are we in making sure the any to any model will survive after we ran out", if by that date I still have 80% of my traffic on IPv4 I won't care, as long as my users can access every and each service they wish for and at the same time somebody who only has an IPv6 address is still able to communicate with my users and can access all of my and my customers' services.

MarcoH