On 10/09/2012 13:58, Saku Ytti wrote:
On 9 September 2012 22:45, Stewart Bryant <[email protected]> wrote:
The expected use of this technology in the failure case
is in conjunction with IPFRR where following a protected
failure, and in the absence of a convergence control
technology, microloops may form and/or the repair
may be staved.
If failure cannot be IPFRR protected, conventional FIB insertion would be used?
Yes.
Note not only do you need to learn of the failure
and compute the new SPF, you also need to update the
FIB. The FIB update time can be the dominant factor in
re-convergence.
Yes, certain modern chassis based L3 switch can literally take >hour
in real network to update its FIB.
If the platform takes hour you definately need IPFRR.
I once thought of sending commit-to-FIB absolute time in TLV in LSP
(Router NTP can give 500us precision easily). But maybe that wouldn't
be deterministic enough. In oFIB I fear that FIB delay in devices I'd
actually want to use in practical scenarios are less than my
linkdelays, so signalling would just reduce convergence time.
Have you looked at the loop free convergence framework?
This was an idea we considered in the early days we called
it synchronized FIB.
- Stewart
_______________________________________________
rtgwg mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/rtgwg
