[v6test] Concerns about network impact

[Kevin Day]

I've had a couple of emails so far from network operators concerned  
about the impact a sudden rush of v6 traffic might have on their  
network. Some of the points brought up are:

1) Only having enough v6 capacity for current near-zero usage, not  
wanting to upgrade because users are playing with something
2) Routers that handle v6 routing purely in software, afraid that a  
lot of v6 usage will kill their router
3) We might expose weak spots in the v6 backbone, causing wide-scale  
outages.
4) We might give users bad instructions for enabling v6, causing  
support headaches for ISPs

My take on these..,

1) Networks might not have enough capacity

TCP should handle quite a bit of this, if their v6 connectivity is  
slow we should scale back to handle it like any other congestion. I  
understand that some networks are going to have far less v6 capacity  
than they do v4, I don't see that this is really unique to v6.

2) Routers that deal with v6 in software

I'm really having trouble imaging that this is going to be a problem.  
A bit of digging seems to say that any router which is old enough that  
it's doing v6 purely in software is probably not doing much v4 traffic  
by today's standards. I can't see how everyone has their network  
engineered, but is anyone out there aware of networks enabling v6 on  
production routers that can't handle it actually being used?

3) We might break the internet

I'm really hoping we don't do this. :) I don't believe it's possible -  
but if we actually did break something I'd hope the lesson(s) learned  
from it would be more valuable than not doing it.

4) We're going to encourage people to install v6 connectivity which  
won't work, breaking things for end users.

Right now there are loads of places telling users how to play with v6  
(tunnels, 6to4, etc). This problem does exist already - as part of the  
data that prompted me to actually run the experiment, we ran a test a  
while back.

On a very busy site(that doesn't target technically-competent users)  
we added two 1x1 .gif images. The order in which the images were  
placed on the page was reversed randomly on every page load, to try to  
reduce bias from a page load being aborted.

ipv4.gif is on a hostname that has only a v4 A record.
4or6.gif is on a hostname that has both AAAA and A records.

Over a short test period:

ipv4.gif was loaded 278821 times.
4or6.gif was loaded 278704 times. (0.042% lower than the v4 only image)
191 hits were on the v6 IP.

Some extrapolations from this data would seem to indicate:

308 users have v6 connectivity turned on in some way. (0.11% of our  
total viewers)
191 have working v6 (62%)
117 have broken v6 (38%)

So, this seems to say that there's about a 40% failure rate over time  
in end-users turning on v6 (either automatically or intentionally).  
This really does concern me, because we experienced (like many others)  
that if you add AAAA records to a site that targets non-technical  
users, you start getting complaints that it won't load anymore.


My suggested mitigation strategies:

For #1 and #2, we're going to suggest to networks that feel their  
network is going to break to preemptively block or rate-limit access  
to the IPs used for this experiment. Since we all know how long it  
takes to get people to actually remove filters later, I'm very  
hesitant to use our own address space for this. It sounds like we'd  
qualify for an "experimental allocation" from ARIN to run this from,  
I'm just waiting to see if any of our sponsors will pony up for the  
fees for the allocations.

For #3, we're publishing a 24/7 NOC number before beginning, and will  
ask that anyone who sees a problem as the result of what we're doing  
to contact us right away.

For #4, we're going to try to come up with fool-proof instructions on  
how to enable v6 if an end-user is interested, including uninstall  
instructions if it doesn't work. Stressing the importance of  
uninstalling if it doesn't work. Working with vendors before-hand to  
ensure the instructions are as universal as possible.


Thoughts? Comments?

-- Kevin