[c-nsp] using BGP bestpath MED missing-as-worst

Adam Chappell adam.chappell at gmail.com
Mon May 8 06:57:44 EDT 2006


Hi all.

Does anyone on this list use the BGP bestpath modifier
"missing-as-worst" for MEDs?

I have this running consistently within a significant-sized AS based
on a decision a long time ago that routes without metric should be
considered worst for best interop with other BGP implementations. This
is in contrast to the IOS default and RFC4271 now says that routes
without MED should be considered similar to lowest possible MED as
well, so I think it's probably time to revisie this decision.

Another factor in reconsidering the use of this feature is a recent
Cisco IOS bugfix, which has cuased me some complications with using
"missing-as-worst".

CSCef34800 fixes a bug in IOS that means that the highest possible
metric value, 4294967295, is treated as an unreachable metric, eg. the
route is not selected as part of the best path selection algorithm and
is never advertised to any other peers.

In line with this, using "missing-as-worst" results in routes with no
MED having a value of 4294967294 assigned rather than 4294967295 as
might be expected.

The standards don't seem to document that 4294967295 should be an
inaccessible route metric, just the worst possible value - so the
bugfix addresses this point and makes 4294967295 a valid metric that
BGP path selection will consider.

Unfortunately, it also changes the missing-as-worst behaviour so that
routes without MED receive the newly-redeemed 4294967295 metric as
well.

This causes potential problems for anyone attempting a transition from
code without the bugfix to code with it since a route with the
4294967295 metric can be learned at an EBGP border and passed to an
internal router that doesnt have the bugfix and therefore doesnt
select the route.

The only solutions I've been able to come up with are:
- turn off missing-as-worst everywhere; do soft clears/route refresh etc.
- rewrite all customer EBGP sessions to ensure that they do have a valid metric.

The first seems the easiest, esp. if this feature isnt well used by
others. The second is quite intrusive becuase of the labour involved,
but also because it could compromise the ability of multihomed
customers to use MED legitimately.

Does anyone else have any similar experiences on this...?    Thanks.

-- Adam.



More information about the cisco-nsp mailing list