[ppml] Effects of explosive routing table growth on ISP behavior
sleibrand at internap.com
Wed Oct 31 18:01:11 EDT 2007
Jason Schiller wrote:
> When routing slots become scarse we all will have the difficult decesion
> of which routes to install.
> It is my suspicion that ISPs will discard non-customer routes first before
> customer routes. The first set of routes likely to be discarded are
> non-customer routes that are a more specific of an aggregate. This will
> break TE.
Yes, it will *partially* break TE. I maintain that the vast majority of
the TE benefit remains, however, via more-specific announcements to your
upstreams, localpref communities, and AS path prepends.
> The next set of routes likely to be discarded are non-customer
> At this point, a paying customer may complain that it cannot reach some
> non-customer destination. The ISP will explain that the destination in
> question is not a customer, and does not pay to have its route in the
> table. The ISP will then offer the option for his or her customer to pay
> to have the route in the table, or the destination can pay to thave their
> route in the ISp's table. Talk about net neutrality issues.
I think there's a less messy alternative to going this far, which I have
attempted to outline in other messages & threads.
> On Wed, 31 Oct 2007, Scott Leibrand wrote:
>> I don't think that's true. Today, anything you advertise in IPv4 down
>> to a /24 will be accepted more or less by everyone. However, that's
>> only a convenience for optimal TE. In order to just maintain
>> reachability, all you really need is for your upstreams (all your
>> transit providers and all their transit providers) to accept your
>> deaggregated routes from you and from each other. The rest of the world
>> need only accept the RIR-assignment-size covering aggregate, route the
>> traffic toward one of your transit providers, and then let the
>> more-specifics take over when they hand the traffic off to them.
> For large ISPs with widespread peeing, and lots of mult-homed customer
> overlap, this basiclly means they need to carry nearly all of the routes.
Perhaps you could help us understand the magnitude of the difference a
little better with some actual statistics, if you could share them. Of
the /23 and /24 routes in your table (excluding class C swamp space),
how many (and what percentage) are non-customer routes? My intuition /
hypothesis would be that the non-customer routes outnumber the customer
routes, so you would still get significant benefit from filtering the
non-customer more-specifics if we come to a routing-table-size crunch.
More information about the ARIN-PPML