Re: BGP-4 changes

[[e-mail]]

*** Resending note of 08/31/96 15:15
Subject: Re: BGP-4 changes
John,

I now have second thoughts about changing the MED value; stripping the
attribute is of course still OK.  The problem is not route loops,
but non-deterministic routing (i.e. not predictable).

>As far as I'm concerned, being able to strip the attribute at the border
>achieves this goal fine.  Either or both is OK with me.

>In private email Curtis mentioned a case (which in retrospect I should have
>seen) where we should be able to strip MED after route selection but before
>export into IBGP.  The case applies when I have a router facing two
>neighbors across a DMZ.  I want to consider the MED to get the next hop
>"right" across the DMZ, but I want to use shortest-path inside my AS.  I
>believe that this stays loop-free since we only consider MED once we
>have already decided to choose an external route.

Let's assume that a router considers the original EBGP MED in the
route selection process, but sends the modified MED in the
IBGP export.  Further assume a provider modifies the MED to value
0 at two border gateways (A & B), in order to try to draw traffic to
the nearest of these two exits.

Suppose border A gets an external BGP route with MED X first.  With no
competition from B, border A installs the route and exports into IBGP
with MED 0.  Border B will see this same route only with MED 0.

Later, border B receives its route with MED Y (same prefix, same
neighbor AS, same preference, etc.).  So long as Y > 0 (it usually
is), then border B will reject the external in favor of the internal
route from A.

Note that acceptance of the internal route from A over the external
route of B is based totally on timing issues -- first come, first
served.  It only occurs when the MED value is decreased -- that's why
stripping the MED is OK, since missing MEDs have value -1.

To fix this, one might use the modified MED value for both route
selection and IBGP export, but one throws away useful information
as Curtis and John mention above (for optimizing local external
route selection).

Do we need the ability to change the MED attribute value, or is
it sufficient to be able to remove the MED attribute?  Can we get
away with modifying local preference instead of changing the MED
value?

-- Richard Woundy, IBM