The MPLS WG Archive

[Curtis Villamizar <curtis@fictitious.org>]

In message <20030312231923.GP23300@gblx.net>, Matthew Meyer writes:
> 
> Thus spake Curtis Villamizar (curtis@fictitious.org):
> 
>  |
>  |I welcome corrections from anyone that knows of an LSR that implements
>  |OSPF/TE or ISIS/TE that doesn't use the flooded information as a hint
>  |that some of the LSPs are down.
>  |
> 
> I know of one, though that may have been on old code.  Regardless,
> for the sake of interoperability with less intelligent code, I'd
> think folks would agree one would not want to skip sending the 
> tears on link-down, at least by default.  

Absolutely.  Always send the tears becasue this does the resource
cleanup.  The IGP hint just makes convergence go faster.  The ingress
teardown can't replace the resv tears past the point of failure.  Also
multiple failures (due to SRLG) can create isolated middle segments of
the former path.  Getting the midpoints to clean up quickly frees
resources that may be needed.

> Another point to consider is that IGP convergence is a function of 
> network size and IGP tuning. It wasn't too long ago that a number
> vendors were less aggressive in their IGP tuning due to processor
> constraints and other factors.  As a result it has been my 
> observation that many networks have resisted the urge to 'jump in'
> and retune their IGP (memories of meltdowns die hard).  Knowing 
> that IGP arrival beats RSVP in a lab or on certain code + certain
> vendor is one thing, assuming these features are deployed optimally
> and on all platforms is quite another.

I'm quite sure you know this but for the benefit of others.

Two things changed.  One is that vendors used to start the SPF and
then flood.  Flooding was really slow and this issue was discussed as
operational experience at NANOG.  The other is that vendors decided
that the origination delay after quiescence should be zero.  Dave Katz
(Juniper) has argued for very fast reflood essentially eliminating any
protocol imposed flooding delay.

The result has been convergence has been getting faster for all of the
major routers who all seem to be paying attention to and participating
in this discussion at NANOG.  (Yes Scotty, there is intellegent life
at NANOG, and it seems friendly enough).

[Discussing operator experience as a means to improve protocols and
protocol implementations - what a concept!]

> Because of the number of variables relating to customer and vendor 
> implementation choices, a skip-sending-tear-on-link-down feature
> would IMO get implemented (if at all) as a vendor specific knob.
> Translate: this isn't standards material. :-/

The skip-sending-tear-on-link-down would be a real bad idea and I
don't think it should even be implemented as an option.

Sounds like you misinterpreted my suggestion or I missed something in
the suggestion that was originally posted.

The IGP and RSVP tears are both hints that the ingress can use to
determine that a link is down.  The former can be made more efficient
and faster but that doesn't mean the latter should be eliminated
because it still serves a purpose including among other things
interoperability.

> Matthew

In practice both arrive before all but a few CSPFs are redone so for
reroute its somewhat of a moot point.  For standby LSP the resoration
can be accomplished prior to processing a bombardment of PATH tears
and RESV tears that result from a failed link.  For subsecond standby
LSP resoration with lots of LSPs to restore this sort of optimization
is beneficial.

Curtis

ps - now back to out regularly scheduled arguments about
call/connection separation.  :(