What problem should we be working on?

[mccreary at pch.net]

Van's comments in the WG meeting are worth reiterating a bit.  Router
resources such as memory size and CPU speed scale like Moore's Law, and so
growth in the resources required in core routers are really only of concern
if they are growing faster than the available resources.  We have seen the
number of paths grow by ~2x over the past three years, and presumably the
amount of memory required in core routers has grown proportonately.  Moore's
Law suggests the available memory will continue to grow much faster than this,
doubling in capacity every 18-24 months.  This suggests the current rate of
growth is manageable, and in fact we can tolerate a significantly higher
rate of compound growth.

We can make similar measurements of the rate of growth in CPU load on core
routers to see if the growth rate is also slower than Moore's Law.  Does
anyone have data that might allow us to determine this growth rate?

As Geoff pointed out, there are strong economic forces that are pushing us
towards finer grained policy in the core routing table.  No matter what we
do, we will have to cope with very large core routing tables until a
replacement for BGP4 can be deployed.  If the amount of resources consumed in
the core routers is not a problem, then we should be spending more time
identifying the factors that will really cause us problems in the future, and
designing schemes to deal with them.  Since we know from Craig and Abha's
work that convergence times grow with denser interconnectivity, it is likely
that the biggest problem we will face in the near future is increasing
convergence times.

Do we know how proxy aggregation will affect convergence times?  Naively it
seems that proxy aggregation has a chance of helping, since it might hide
some of the finer-grained routing changes from ASes downstream of the
aggregator.  On the other hand, it might also increase route volatility by
coupling transitions in separate routes together.  For example, if the
aggregate must be withdrawn whenever one of the aggregated routes is
withdrawn, then this seems to be the most likely outcome.  We need to try and
determine which of these outcomes is more likely to occur.
--
Sean McCreary                                                 mccreary at pch.net