prefix lengths [was Re: stla registry db issue]

[Bob Fink]

Paul,

At 02:43 PM 1/28/00 +1000, Paul Wilson wrote:
>Here's a somewhat belated response on this topic, and on the previous
>thread.
>
>In this discussion it seems to me that the critical question is, what
>defines an ISP?
>
>In the v4 world, BigCo may be a large multinational company with a large
>multihomed infrastructure, which doesn't call itself an ISP but which
>nevertheless qualifies for a Provider Independent (i.e. portable, globally
>routable) allocation. On the other hand, SmallCo may provide ISP services,
>but have a small singly-homed infrastructure and therefore not qualify for a
>PI allocation (needing instead to get address space from its upstream).
>Therefore the allocations that RIRs make are not simply dependent on whether
>the organisation is an "ISP", but rather on a number of other
>technically-based and objective criteria.
>
>We could of course redefine the term "ISP" to mean an organisation that
>actually receives a PI allocation, but that's not really useful or necessary
>since the term is so widely used to refer to a bigger class of
>organisations.
>
>In the IPv6 world, we could likewise define an "ISP" as an organisation that
>qualifies for a /35 allocation and a /29 (subTLA) reservation, and I suspect
>that this definition is behind some of the previous discussions. In this
>case it goes without saying that we would never share a /29 among multiple
>"ISPs", simply because the /29 is reserved for each "ISP".
>
>However we must agree that once a /29 has been reserved/allocated to one
>organisation, there *will* be downstream customer organisations receiving
>assignments from that address block and using those assignments for ISP
>activities. But the critical thing is that whether or not they call
>themselves an ISP (and we don't care) their assignment will be "provider
>aggregatable" and not entitled to be announced globally. Rather, like any
>other downstream customer, their prefixes will be aggregated by the upstream
>provider within its own announcement.
>
>I believe it is this type of organisation (the small, downstream ISP) which
>Kengo Nagahashi was referring to in his original message.
>
>Incidentally, I guess such a downstream organisation (call it an ISP or not)
>could conceivably end up with a /35 or larger assignment, but would still
>aggregate that within its upstream address block, until such time as it
>qualified for its own subTLA allocation. Furthermore, my understanding (and
>correct me if i'm wrong) is that such an organisation could also be
>multi-homed, having an equal-sized assignment from each of several
>upstreams, but again having each of its routes aggregated in the global
>announcements of those upstreams. This may not be the best case, but it
>surely must be a possible outcome for one of the small ISP that grows a
>multi-homed infrastructure without yet qualifying for a subTLA.
>
>Finally, on the question of advertising prefix lengths longer than /29, it
>is in fact the proposal of the RIRs that the prefixes announced by any
>organisation will correspond only to their allocation (anywhere from /16 to
>/35), and not to their reservation (which would be either /16 or /29). The
>justification behind this (as discussed with the IAB in Minneapolis last
>year) is that in a future scenario where many TLA and subTLA blocks were
>released and allocated, and where routing table expansion again becomes a
>concern, it is necessary for ISPs to have an objective means available for
>limiting the size of their routing tables, and the best such means available
>is to filter on prefix length.
>
>The alternative is a huge routing table populated with /16 and /29 prefixes
>only, where no objective means for route filtering is available, and where
>bilateral routing agreements would emerge as the only way for an ISP to
>control its tables.

One opinion (mine at a minimum) is that simply filtering on length of 
prefix does not accomplish any desired result, i.e., that it is not an 
adequate determiner of any interesting metric. ISP's (especially large 
ones) will continue to filter based on whatever criteria are imortant to 
them, and length of prefix alone will not be it.


Bob