Routing System Scaling - Disaster Looming, but Medium-Term Fixes Known

It's very simple: ANYONE can build a multi-Gpps router which can handle
a million or two static routes. Hell, if the Internet were completely
undynamic, alot of things would be much easier.

Likewise, anyone can build a 0 pps router that can handle an order
of magnitude more dynamism in the global routing system than we
observe today.

Finally, anyone can build a multi-Gpps router that can handle
heavy dynamism in a global routing table consisting of a dozen prefixes.

The Internet: bigger, faster, more dynamic. Choose two.

One usually sees "cheaper" as one of the terms in the "choose two" joke.
In this case, it's implicit, since the total price of a large router
is greater than the sum of the (increasingly expensive) cost of providing
the oompf necesary to do each of these three things.

So, there are several Deaths of the Internet which are possible:

  -- it's too expensive to keep up with growth, so utilization
     falls off, since nobody wants to pay the real bills

  -- we blow up on one of the scaling axes
    -- too many routes: poof, we run out of memory
             routers are unhappy when out of memory...
       (ok, more memory is cheap for static storage, but
                    ever-faster memory seems to be a general problem)
    -- too much dynamism: poof, our memory isn't fast enough,
       or the distribution of stuff via BGP/TCP isn't fast
       enough, or we have no CPU cycles to generate tables
       fast enough our our communication among routing-process
       CPUs and forwarding engine isn't fast enough, etc...
       everyone is unhappy when it takes minutes, hours, or
       DAYS to see the end of blackholes/loops/other effects
       of incomplete convergence after a change
       (ok, everyone knows about route damping, yes? )

  -- we blow up because while we can handle some of the
     technical aspects of routing system complexity, we lose
     on human factors

                -- inefficient/broken routing: we don't have the power
                   needed to apply decent policy to the routes we learn
                   from elsewhere, which is already seen in the form
                   of route leakage, inadvertant transit, etc., and we
       don't have the skills to quickly find & fix problems
       that have global effect

The reaction to scale driven by increasing numbers of ever-longer
prefixes should not be a panicky turn to something radical, or
sky-is-falling moaning. It has been demonstrated (by me) that
there is little reachability lost when applying prefix-length
filtering, or when applying progressive flap-damping (e.g., if you
flap something longer than a /24, the route's gone from my tables
until tomorrow). Dennis Ferguson's arguments for maximum-prefix
limits from peers are also very interesting, and have gotten some
testing in the real world.

Ultimately the problem here is that this just makes introducing
ever-longer prefixes harder, but does not expose a real cost to
someone who is thinking of doing this. If various people, or
someone sufficiently large, ever gets around to implementing
the following idea, I think people may be more inclined to renumber
and/or be aggregated, rather than try to introduce "exception routes"
into the global routing system.

What I wish I had finished @ Sprint: a web-based form that lets
one arrange an exception to access-list 112 for some small fee,
like $50/month/change. For your 50 bucks, you see no filtering
on one of your long prefixes for a whole month (or year or whatever).

These days it may be that your $50/month will make your
long prefix less likely to be damped out of the routing table
for a long time, however there is scope for further innovation,
such as charging only $5 to clear your prefix out of the flap-damping
penalty box of one router.

This revenue stream would surely pay for alot of the costs
of upgrading routers with more memory, faster CPUs, more CPUs,
faster memory, better code, etc. to deal with the increase
in the number and dynamicism of the current Internet's global
routing system. It might even fund the switchover to something better
than the IDR architecture we have now.