Short version: I can't quickly summarise this, but I have a
tentative map of the design space regarding one or
two solutions for "big" and "small" networks.
Eliot's suggestion (as I understand it) to exclude
larger end-user networks from the (presumably)
renumbering based solution for smaller networks
would still leave many "smaller" networks way too
big for "routine" renumbering.
Hi Brian,
While I agree with what you wrote here, responding to Bill Herrin,
who I also agree with in this respect:
>> Disagree. There's a money issue here: large networks can afford and
>> are willing to spend more money on multihoming than small networks.
>> There's no reason that both should be constrained to the same solution
>> at the same cost. The requirement is that whatever the scaling
>> solution, the per-network cost must either be trivial -or- recoverable
>> from the networks who instigate it.
>
> I agree with Bill, and it seems to me that there's a fundamental
> difference between Eliot's o(10^6) big sites and the much greater
> number of small sites. The big ones are very likely to have a DMZ,
> run their own servers, and have multiple points of interconnection
> around the world. In other words, prime candidates for PI based
> addressing and maybe a map/encap style solution.
>
> The small ones are very likely to have a simple firewall/router
> combo, outsource their publicly accessible servers, and have a single
> point of attachment (or at least, several attachments in a relatively
> small geographical area). They are also, I believe, much less likely
> to be significantly disturbed by renumbering than the large sites.
>
> I think there's definitely scope for two solutions.
I want to make the following points:
1 - The observations you make of smaller networks at present are
probably generally true - such as of my little single IPv4
address patch of the Net via an ADSL line, and 8 addresses I use
in a hosting company in Dallas Fort Worth. However, I think it
would be best to avoid a dual solution arrangement which is
predicated on particular patterns of usage which may not persist
indefinitely.
2 - All currently potentially practical (no host changes) solutions
(LISP, APT, Ivip, TRRP and I think Six/One Router) provide a
single solution for all sizes of end-user network.
3 - While I am not completely opposed to there being two different
solutions for "large" and "small" networks, however defined,
I propose Ivip as a solution for networks of all sizes.
Also, I stand by my arguments for a single solution for networks of
all sizes:
Map-encap space only for small end-users? PI space prices
http://psg.com/lists/rrg/2008/msg00454.html 2008-02-22
in order to:
a - Avoid a situation where some small networks either don't adopt
a solution (because there is no solution for big networks, and
they intend to be big one day) or because they adopt an
inappropriate solution (the big one) for the same reason.
b - Provide a simpler, more elegant, overall architecture with no
debates about which solution each network should adopt.
c - To make a solution which is generally not dependent on patterns
of usage, which are likely to change over the next few decades.
I am not saying that a dual solution is unthinkable. Just that
there would need to be very strong arguments for it in order to
justify the extra work of devising the two solutions and coping with
the problems noted above.
There has been some suggestion of "no solution for big end-user
networks", such as by Eliot:
http://www.irtf.org/pipermail/rrg/2008-October/000048.html
who thinks that their numbers - or at least the numbers of separate
BGP advertised prefixes - is within reasonable bounds.
I don't agree with the idea of a million separate prefixes being
within reasonable bounds of the scaling problem.
I guess it is always possible to nominate some criteria for choosing
whether an end-user network is "big", such that the long-term impact
of its BGP advertised prefixes (in the absence of adopting any
scalable routing solution) is deemed to be "acceptable". My sense
is that any reasonable criteria of this nature would leave only a
handful of very large end-user networks, such as a few of the
biggest universities, corporations or whatever, and so relegate
virtually all end-user networks, including those of many
universities, government departments and corporations, to the class
of "small", or at least "not big enough so as to escape the
requirement that they adopt the scalable routing solution".
That would leave many genuinely large networks classified as
so-called "small". It would not be practical to treat those the
same as a DSL-based SOHO network and expect them to do "routine
renumbering" when choosing a new ISP.
The design space in terms of end-user network size seems to be:
1 - Single solution for all sizes. (My preference, and apparently
the assumption behind LISP, APT, TRRP and Six/One Router.)
2 - Two separate solutions, such as:
2a: Core-edge separation, such as map-encap, for "big"
networks, and something like SHIM6 and renumbering
when choosing a different ISP (2 are needed for SHIM6)
for small networks. (Note this precludes any solution
for small IPv4 networks.)
2b: Any others?
3 - No solution for "big" networks, and one solution (I assume)
for small networks. I guess this is Eliot's position: no
solution for some class of "big" networks, with another
solution (map-encap, or SHIM6 and renumbering?) for all other
networks, which are deemed to be "small".
(4 - A solution for big networks but no solution for "small"
networks. No-one is suggesting this - except perhaps
for those who suggest IPv6-only solutions, excluding
IPv4 networks, large and small.)
Since I think Ivip is a good solution - the best solution - for
networks of all sizes, I am happy to propose there be a single
solution for all sizes of networks.
While not wanting to suppress brainstorming, I think that if there
was going to be a serious proposal for two solutions, or for only
one solution for "small" networks and none for "big", then any such
serious proposal would need to involve a pretty robust critique of
Ivip and the other current "single solution" proposals: LISP, APT,
TRRP and Six/One Router (which is IPv6 only).
- Robin
_______________________________________________
rrg mailing list
[email protected]
https://www.irtf.org/mailman/listinfo/rrg
