Re: Non routable IPv6 registry proposal

[Brian E Carpenter <[e-mail]>]

On 21-Jan-21 14:41, George Michaelson wrote:
> MAC addresses are on a shared medium.
> 
> ULA are disjoint. The risk is that 1 in 220 networks in a field of
> 100,000 things pick the same value of NON ROUTABLE

The probability of a collision among 100,000 ULAs is 4.54*10^-03, according
to RFC4193. So most likely there is no collision at all, but let's
continue...
 
> and then.. decide to join. Now, tell me the risk of 1 in 220 things,
> both picking the same ULA, *AND* seeking to join their private
> networks.

It's unknowable, because there is no calculus for how likely network A is
to interconnect with network Q, where A and Q are arbitrary choices.
What is knowable, and is worked out for you in RFC4193, is that the
probability of A and Q both having picked the same pseudo-random 40 bits
is 1.81*10^-12.

Of course, if we assume that 10 billion sites around the world each pick
a pseudo-random 40 bit number, the probability that collisions exist is
effectively 1. But that doesn't matter; for any two sites that happen to
interconnect, the probability remains at 1.81*10^-12. Even for a set of
10 sites that happen to interconnect, the probability of collision is
only 4.54*10^-11.

Somewhere above 1000 sites, the probability of a collision gets as high
as one per million. So if you run a very large corporate network, assign a
separate ULA /48 prefix to each site, and merge two sites, you are looking
at a 1 per million chance of a problem, unless you make a trivial check
before doing the merge. (However, why a corporate network would do that
is a mystery. They could simply use one ULA /48 for the whole network.)

I have never lost any sleep over this issue, which has been well understood
and documented for 15 years.

    Brian
 
> -G
> 
> On Thu, Jan 21, 2021 at 11:39 AM Christopher Morrow
> <[e-mail]> wrote:
>>
>> (I'm not a ULA fan, it's going to cause problems.. but)
>>
>> On Wed, Jan 20, 2021 at 7:16 PM John R Levine <[e-mail]> wrote:
>>>
>>>> No. Collissions are actually almost guaranteed (birthday paradox).
>>>
>>> When I do the birthday paradox calculation, I find that the chances that
>>> 100,000 random numbers each 40 bits long are all different is about 95%.
>>> I'll take those odds.
>>>
>>
>> Warren made this website/application/etc:
>>   http://mac-collision-probability.appspot.com/calculate
>>
>> which was used for privacy address problems/discussions, but... if you
>> make it 40 bits and 100,000 stations
>> apparently you'll get a collision 1 out of 220 times. I think that
>> means that ULA network selections COULD overlap at about the same
>> rate.
>>
>> There are ~1.4m small businesses in the US, if they all chose ULA
>> that's more than a few collisions.
>> Collisions matter because when 2 networks that collide come together
>> it's messy :( to untangle and decide whom is going to do what :(
>> it's also not always obvious that it's happening :( until something crashes :(
>>> Regards,
>>> John Levine, [e-mail], Taughannock Networks, Trumansburg NY
>>> Please consider the environment before reading this e-mail. https://jl.ly
>>>
>>
> 
>