[arin-ppml] V6 address allocation policy

[Owen DeLong]

On Jan 17, 2010, at 7:46 AM, William Herrin wrote:

> On Sat, Jan 16, 2010 at 7:55 PM, Owen DeLong <owen at delong.com> wrote:
>> On Jan 16, 2010, at 4:28 PM, Michael Richardson wrote:
>>>  If techies need to get their managers to approve a checque to ARIN,
>>> the manager tells them to use IPv4 + NAT.  If the techies do not have to
>>> ask, then they will deploy IPv6 for internal use.
> 
> No joke on that. Three years ago my boss turned me down on deploying
> IPv6 in my spare time at work on the grounds that there were wiser
> ways to spend the $1250.
> 
> 
>>>  (ULA buys you nothing compared to net-10)
>> 1.      ULA buys you a great deal more than RFC-1918.  ULA is statistically,
>>        if not globally unique.
> 
> Not exactly... The analysis in RFC 4193 (ULA addressing) section 3.2.3
> is technically correct but it may be an example of "lies, damn lies
> and statistics."
> 
> First, though de-emphasized in the RFC, the probability of collision
> has a phenomenal growth rate: two orders of magnitude for ever one
> order of magnitude increase in the number of ULA IDs. So you close in
> on a 100% chance of collision not at 2^40 IDs as you'd expect but at
> merely 2^20.
> 
So for every 2 companies merging, you run the risk of a 1:2^20
collision.  Now, let's look at those odds in numbers more meaningful
to people... 2^20 is 1024^2, or, 1,048,576, so, the odds are, literally
not quite as good as 1 in a million of any two companies colliding.

I would argue that the odds of a collision in RFC-1918 are a lot closer
to 1:3 at best since almost everyone uses at least one of 10.0.0.0/24,
172.16.0.0/24 or 192.168.0.0/24 (or some supernet thereof).

So, in order for ULA to buy you nothing, you'd have to be able to argue
that 1:3 and 1:1,048,576 are equivalent risks. If you are willing to make
bets like that, I want to be your bookie.

> Second, consider the way folks tend to behave. Each private network
> built for whatever purpose in a particular company will consume one or
> several ULA IDs. That's each private network in each project at each
> branch of a company. A large company may well have consumed hundreds
> if not thousands of ULA IDs introducing another four to six orders of
> magnitude increase in the probability of collision when two such
> companies want to connect.
> 
Even if this is true (I'm not completely convinced), you're comparing
ULA at 1:100<n<1000 vs. RFC-1918 at 1:0.003.

> Practically speaking, we should start to see anecdotes about ULA
> collisions as folks try to connect 100 to 1000 organizations together,
> still a usefully large number but far fewer than RFC 4193 implies.
> 
Practically speaking, even if you buy into that argument, you're still
quite a bit better off than RFC-1918.

1.	The odds of a collision are still about 300,000 times better.
2.	The percentage of hosts likely to be affected by such a
	collision is orders of magnitude better than RFC-1918.
3.	The above all assumes not using the SIXXS ULA registry
	to keep your ULA addresses unique.

Owen