[c-nsp] Fast IGP on 6500 & gigE

[Phil Mayers]

Marcus.Gerdon wrote:
> Hi,
> 
> you've written most routers are dual-attached, so the concern mostly
> is failure detection and not re-establishment of a neighbor I think.

Correct

> If you go into debounce or carrier-delay you'll raise the convergence
> time as a link failure will be ignored for a short time before
> processes are notified.

Further reading indicates that carrier/debounce are by default as low as 
you can safely get them on a 6500, so I think I can disregard these.

> 
> OSPF should immediately react on an link-down event, so I'd try to
> speed it up this way. If you use 2 separate SVI for the 2 connections

We're already doing this.

> and each VLAN has only 1 port it is allowed in (either a single
> access port or exactly 1 trunk port) the SVI should go down along
> with that single port.

It does.

Interestingly info I've read indicates that routed interfaces signal 
upper-layer protocols much faster than SVI interfaces, which is 
something I'll have to investigate.

> 
> Playing around the timers I keep for last resort - as there's always
> the risk to de-stabilize the network seriously (I've seen people
> trying to get the last second out of a protocol resulting in
> occasional burn-downs far too often).

Hmm.

Further digging has some pretty concrete recommendations from Cisco in 
presentations and such suggesting:

timers throttle spf 10 100 5000
timers throttle lsa all 10 100 5000
timers lsa arrival 80

e.g.

http://www.ciscoexpo.sk/slides/41-vsettey_fast_convergence.pdf

The default SPF initial delay is 5 *whopping* seconds; which means that, 
no matter how fast your link detection and SPF propagation is, it'll be 
at least 5 seconds before you even *start* trying to converge.

Having read the docs, I have a hard time seeing how changing these can 
"burn down" the network - the spf and lsa timers have exponential 
backoff. Would you care to elaborate on the failure modes you have seen?