[ppml] Policy to help the little guys

[William Herrin]

On Wed, Mar 19, 2008 at 4:43 PM, Danny McPherson <danny at tcb.net> wrote:
>  I'm not saying it's not a worthwhile exercise, I'm simply
>  saying that for anyone that has attempted to compute the
>  cost of bit carriage on an IP packet network based on
>  distance carried, or destination network class (e.g., peer v.
>  customer), or class of service, or router platform redundancy
>  requirements, or rack space, or power consumption, or AC,
>  or amount of required resilience (e.g., SRLGs), access
>  platform redundancy, network management infrastructure,
>  much less OPEX, etc..., or been through similar pricing
>  models, it's just not that simple.

Hi Danny,

The nice thing about all of those factors is that they intersect at
two easily measured junctions: the price you'd pay to run with devices
that do all of it and the price you'd pay to run with devices that do
all of it except the required large routing table.

The art to cost analysis is finding those junctions so that you can
eliminate the subfactors from the equation.


Personally, I'm more interested in looking for creative solutions to
the BGP cost problem that would moot the whole PI/PA issue than I am
rehashing the semi-annual "PI folks are treated unfairly" debate.
Suppose you did something like this:

1. ARIN takes a /16 and lets it be known that only /25 or longer PI
prefixes will be assigned from this /16 (nothing /24 or shorter).

2. Each registrant tells ARIN exactly what routes they intend to
announce out of their PI prefixes.

3. Transit ASes which want to participate can register the number of
DFZ routers they have with ARIN. ARIN will pay them a monthly amount
that works out to $0.04 per router in the AS per registered route in
the /16 per year. So, if there are 4096 /28's coming from the /16 and
Verizon has 400 DFZ routers, ARIN will pay Verizon (4096*400*0.04)/12
=$5500/mo to carry those routes. In return, participating ASes agree
to either filter no routes they receive or transmit within the /16 or
filter only those routes which don't match ARIN's list.

4. ARIN then collects that same fee plus a handling surcharge from
each of the PI registrants who registered a route announcement within
the /16. So, if transit ASes with a total of 50,000 routers have
signed up and you've registered to announce one /26 route for your /26
assignment then you pay ARIN (($0.04*50,000)/12)*1.1 = $183/mo for the
BGP announcements, thereby guaranteeing that they're carried by 50,000
of the DFZ routers.

5. Each year, ARIN solicits bids for an ISP to serve as a "catchall"
tunnel broker for the whole /16. The tunnel broker announces the whole
/16 and (if you pay him) then tunnels any traffic to you via GRE in
order to cover whatever sections of the 'net decided to filter the /25
and longer routes instead of accepting payments from ARIN. Each year
the registrants within the /16 get to vote on which tunnel broker bid
to accept.

Note that if enough of the DFZ routers are carrying the small routes
(including your local ISPs) there is a very good chance that packets
for you will navigate via the /16 route in non-participating ASes and
then latch on to your /25 or longer route once they reach a
participating AS.


Now, that could let folks announcing routes pay for those
announcements so that ARIN could get out of the business of setting
minimum PI assignment length. Workable? Maybe, maybe not. But at least
it's trying to solve the problem which led to PA/PI instead of
complaining about how unfair it is.


Here's an entirely different tack:

1. ARIN takes a /16 and lets it be known that only /25 or longer PI
prefixes will be assigned from this /16.

2. Each year, ARIN solicits bids for a tier-1 backbone to announce a
/16 route which covers the prefix. The registrants within the /16 vote
on which tier-1 bid to accept.

3. Each registrant makes private arrangements (or hires an agent to do
so) between himself and a reasonable set of ISPs that lie between him
and the tier-1 announcing the /16 so that they carry his route. For a
dual-homed network, this will typically be less than 10 companies
including the tier-1 itself.



On Wed, Mar 19, 2008 at 4:54 PM, David Williamson <dlw+arin at tellme.com> wrote:
>  I'm entirely baffled by this assertion.  [...]
>  I don't doubt your analysis shows that inserting a route in the DFZ
>  causes a bit of expense for everyone else in the DFZ, but I don't see
>  how PA versus PI has anything to do with it.

David,

The answer, of course, is that PI and PA shouldn't have anything to do
with it. Everyone announcing a route should pay for announcing that
route directly. Unfortunately, no one has figured out how to make that
work so PA/PI serves as a poorly fitting proxy that's been good enough
to keep the system stable.

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