North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

RE: multi-homing fixes

  • From: Vadim Antonov
  • Date: Tue Aug 28 03:38:43 2001

On Mon, 27 Aug 2001, Roeland Meyer wrote:

> Please detail the exact costs of a, BGP inserted, routing table entry. Is
> it, maybe, 50 cents?

What is the cost of the prefix inserted with causes somebody other's boxes
to crash and burn?  The "last" one?  We don't know at which point the
network simply passes the threshold of being able to converge faster than
the next update comes.  That does not make the cost of every update
pushing the network closer to that point immaterial.

Unlike traffic overload, routing flap overload is _not_ self-correcting.
In fact, given the present technology it is very likely to be
self-amplifying (i suspect that most current BGP implementations in case
of severe overload would simply delay keepalives until peers start
resetting sessions; having BGP to run over strictly serializing transport
(TCP) does not help, either).

The only known non-capital intensive fixes are route aggregation and
intentionally degrading routing system responsiveness (aka flap damping).
Both have severe limitations.  When they run out of gas, it's forklift
upgrade to the new generation of routers.  Keeping up with Moore's law is
not free.

So far, the current backbone upgrade cycle kept up.  Assuming that the
capital cost of the backbone routing equipment installed globally is about
$5bln (this is an out-of-the-blue figure), and it currently works at
design capacity with approx. 100k prefixes, the per-prefix cost of forced
upgrade is about $50k, not including labour costs, and indirect costs of
decreasing network stablility causing customer dissatiscfaction and
resulting in expensive customer churn.

Obviously, not all cost may be attributed to maintaining routing
infrastructure ("traditionally" the upgrades are justified by the need to
maintain competitive backbone speed).  Times have changed, though, and
upgrading switching capacity no longer has to be a wholesale box
replacement.  Unfortunately, this is not the case with routing update
processing capacity (having parallelized routing stack implementations
helps, but not all that much).

Therefore, the cost of extra prefix is definitely not $0.50; it is _much_