[quagga-dev 11601] Re: [PATCH] BGP: add aspath_aggregate_mpath that preserves path length

Daniel Walton dwalton at cumulusnetworks.com
Thu Oct 23 13:51:31 BST 2014



For MED churn you can solve it by advertising the bestpath plus the 
bestpath-per-neighbor-AS. 




I think that only solves it if there's one connection to a neighbour AS. 
Otherwise it'd need to be bestpath-per-eBGP-nexthop? 

In all of the MED churn scenarios I have worked on bestpath-per-neighor-AS does the trick. That's not to say there might not be a topolgoy out there where it wouldn't work, I just haven't seen one. For instance in the topologies in RFC3345 there are dual connections to at least one of the ASes send us MEDs. 

There is another scenario for MED churn that we didn't put in RFC3345. It is one that is silenced by preferring the oldest external path when comparing two externals. The funny (sort of) thing is we stumbled across this in the late 90s but didn't fully explore what was causing the churn, we just put in the "prefer oldest external" step in the decision algorithm to fix it. Then a few years later UUNET had tons of churning routes so we started digging and found that it was due to MEDs :( 

Daniel 

At least, MED oscillation arises when different speakers have a view of 
only a subset of the paths out of the AS. The pair-wise preferences of 
paths can be non-transitive: A < B, B < C, C < A can be true. So if a 
speaker has just a subset of these, and another peer a different subset - 
oscillation is possible. 

If bestpath-per-neighbour-as could cause some speakers to have just a 
subset view again, then it allows for MED oscillation again. 

(This could also be fixed by ensuring that the preference of routes in BGP 
increases monotonically with propagation, e.g. considering CLUSTER_LIST 
above MED could fix that). 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quagga.net/pipermail/quagga-dev/attachments/20141023/8e6ba867/attachment-0001.html>


More information about the Quagga-dev mailing list