On Mon, 14 Dec 2015, Martin Winter wrote:
If getting through means getting to RFC then yes. But partially on a
good side as well (i.e., not constant changes on protocols).
Yeah, there's definitely a good side to it. BGP doesn't lack for all kinds
of extensions these days...
But my idea is to have at least have a formal draft submitted and some
round(s) of discussion (and maybe a 2nd version draft based on feedback)
to address the obvious issues and concerns which may come up.
Sure.
Don’t forget that there is quite a bit of influence on the best path
based on filtering and route-maps and (I think) none of this can or
should be negotiated.
Filtering out a path is fine. Changing attributes in a way that can only
increase the 'cost' of a route is also definitely fine - says the math.
Changing attributes in a way that effectively decreases the cost would
risk causing loops with a simple distance-vector protocol (a decreasing
cost == a negative cost, and that risks creating a negative cost cycle in
the graph, which shortest-path algorithms will keep following around and
around as every 'trip' around the cycle is /decreasing/ the cost of the
route - so it's always cheaper to follow the cycle than break out; a
0-cost cycle could also induce loops, but would depend on implementation
details).
However BGP is path-vector, and has AS_PATH and CLUSTER_LIST to detect and
break such cycles in eBGP and iBGP reflection paths. So, should be safe.
I.e. any given /single/ path in BGP must be simple (can't loop on itself),
because of those attributes.
So, filtering and tweaking attributes is generally safe - least, it
doesn't make the not-so-safe parts of BGP any worse, AFAICT.
I see the internal decision process of order (and potential
enable/disable some steps) as just one part.
Changing the decision process is very different to changing the existing
metrics on a path though. Changing the decision process inconsistently
across routes can easily lead to problematic cycles in preferences across
routers - which can lead to instability.
The path-vector checks can't stop this (they only catch loops in any /one/
path), as the BGP system as a whole would be chasing a cycle in the
preferences between /different/ paths.
regards,
--
Paul Jakma, HPE Networking, Advanced Technology Group
Fortune:
crop circles in the corn shell
_______________________________________________
Quagga-dev mailing list
Quagga-dev@lists.quagga.net
https://lists.quagga.net/mailman/listinfo/quagga-dev
