Re: [secdir] secdir review of draft-ietf-alto-protocol

[Richard Alimi <[e-mail]>]

On Sun, Feb 2, 2014 at 7:08 PM, Dan Harkins <[e-mail]> wrote:

>
> On Sun, February 2, 2014 11:33 am, Jeffrey Hutzelman wrote:
> > On Sat, 2014-02-01 at 10:54 -0800, Dan Harkins wrote:
> >
> >>  Also, given those
> >>      restrictions and the fact that a tag just has to be less than
> >>      or equal to 64 octets, the probability of identical tags being
> >>      used is not zero. I think the probability of the tag from
> >>      example 11.3.1.7 is 0.5 to collide with one of just 460
> >>      other Network Maps.
> >>
> >>      I suggest requiring a tag to be 64 octets. That will make
> >>      even money probability of collision among nearly 3000
> >>      other Network Maps, which is safer.
> >
> > OK, maybe I'm confused and reading out of context here.  But I once had
> > someone tell me I needed to change my 5-character username because they
> > were requiring all usernames to be at least 6 characters, _in order to
> > increase the number of possible usernames_.  That is, they claimed they
> > were increasing the size of a namespace by eliminating possible names.
>
>   Well that's a hair brained policy, but username selection is not a good
> analogy. I was at a company that had no strict requirements on a username
> so there should have been a near infinite size of the namespace. But we had
> a collision when the company had less than 10 employees because there
> was another "dan" at the company.
>
> > The point is, if a tag is required to be exactly 64 octets, you get
> > 0x5e^64 possible tags.  But if it is required to be up to 64 octets, you
> > get Sum(i=0..64) 0x5e^i possible tags, which is strictly greater than
> > 0x5e^64.  So, requiring a tag to be 64 octets _reduces_ the number of
> > possible tags, thereby increasing the chance of collision.
>
>   That would be the case if all tags in the Sum(i=1..64) 0x5e^i tagspace
> were equally probable of being chosen. Which implies implementations
> choosing a random tag length for each tag generated in addition to a
> random tag selection scheme for the randomly chosen length. I suspect,
> though, that in practice the tag length will be fixed for a particular
> implementation and the tag selection scheme will not necessarily be
> random. So the herd mentality, plus the proliferation of one or two
> companies' ALTO servers, will result in a severely reduced size of the
> effective tagspace and the increased possibility of collisions.
>
>   A tag generated as SHA256(NetworkMap) represented in 64 hex
> characters would basically guarantee you'd never have a collision.
> Saying, "it can be anything you want as long as it's less than 64
> octets" would not.
>

Should I interpret your comment to say that we should to require particular
mechanisms for generating version tags, or be more explicit about
suggesting mechanisms that have a low collision probability?

To help steer readers towards better implementation practices, we'll change
the examples to use hashes in the version tags.

Thank you again for the review!


>
>   Dan.
>
>
>
>