[arin-ppml] Submitted to ARIN Consultation and Suggestion Process

[William Herrin]

On Mon, Aug 22, 2011 at 3:42 AM, Owen DeLong <owen at delong.com> wrote:
> On Aug 21, 2011, at 12:27 PM, William Herrin wrote:
>> On Fri, Aug 19, 2011 at 4:02 AM, Owen DeLong <owen at delong.com> wrote:
>>> I'm talking about the coming day when the transfer markets have
>>> sufficiently exploded the routing table that the DFZ providers are
>>> having to make route acceptance decisions.
>>
>> That's unrealistic. Assuming the /24 barrier remains in place (and
>> there's no good reason to assume otherwise) the terminal size of the
>> IPv4 table will be well under the maximum that can be handled by a
>> cost-effective router.
>>
> By my count there are ~15,000,000 /24s in the usable unicast
> IPv4 space. Please explain to me how the terminal size of the
> 15,000,000 route table will remain well under the 1,000,000
> route maximum in a cost-effective router?

Hi Owen,

1. There's shipping equipment that supports 2M routes and the
manufacturers have been capable of building routers that handle 10M
routes for several years now. They don't because nobody's asking for
them yet. I know this because (A) vendors C and J have each said so
and (B) the math works.

2. The number of unicast /24's in IPv4 is not the worst case terminal
size of the table. For one thing, there are overlaid routes (the /20
goes here a /23 inside it goes there and some /24's go a third place).
That increases the count. For another, even under worst case
assumptions you should expect a huge number of prefixes that nobody
needs to disaggregate to /24 or anywhere close. It's like subdividing
property -- some gets divided into 1/8th acre lots, but only some.
Most is usable as desired at much larger sizes and the next owner
wants all of it, not part. That reduces the count.

The worst case assumptions (the collapse of IPv6 and indefinite
continuation of IPv4) put the terminal size of the IPv4 table at no
worse than the 6M to 8M range... about 20 times larger than it is now
and well inside what the manufacturers were capable of building two
years ago.

While the BGP table size is currently compounding at about 25% per
year, that can't actually continue forever. It levels off somewhere
and that somewhere is well short of the theoretical maximum number of
/24's possible to express in a routing table.


> The reality is that is massively unrealistic and the combination of
> large provider aggregates (2011-3) and other constraints (there
> aren't 2^48 end-sites to give /48s to in reality), etc. will prevent us
> from getting anywhere near [2^48 IPv6 routes in the BGP table].

Indeed. The worst-case demand for global routes in the next 50 years
is only around 2^35 routes.

Back-pressure from the /24 barrier in IPv4 prevents this in IPv4 --
you may want a route but others will want a route more than you do,
driving the price of a /24 out of the range acceptable for your use.
Joe's Pizza, for example, won't spring for the at least $1000 a /24
will take.

In IPv6 the back-pressure is at the /48 barrier. Joe goes to ARIN as
part of a pizza consortium to get a /40 and then they all split it up
with individual /48's at around $50 each and a buck a year.

While a cost effective router can be built to handle 8M routes, no one
has figured out how to build the same router that can handle 4,000
times as many and are not expected to do so with BGP.



>> For that reason, I SUPPORT your suggestion.

> Excellent. The good news is that it's useful for both purposes.


Regards,
Bill Herrin



-- 
William D. Herrin ................ herrin at dirtside.com  bill at herrin.us
3005 Crane Dr. ...................... Web: <http://bill.herrin.us/>
Falls Church, VA 22042-3004