Re: [OSPF] draft-kini-ospf-fast-notification-01

[Curtis Villamizar <[e-mail]>]

In message <[e-mail]>
Greg Mirsky writes:
>  
> Hi Curtis,
>  
> thank you for the explanation on IPFRR. I know that things are
> sometimes are not what we call them but I believe that if MPLS FRR
> doesn't work for some reason and LSP rerouted to restore service it's
> not FRR. Would IPFRR still be considered FRR if restoration of
> connectivity relies on flooding and route convergence, not on use of
> pre-calculated RIB?
>  
> Regards,
> Greg


Greg,

Call it what you want.  Multiple failures where MPLS protection
doesn't work is reality and has to be dealt with.  The alternative is
transport style multiple hour outages where single fault protection is
backed up only by manual intervention (an apparent goal of TP).

Curtis


> On Mon, Apr 11, 2011 at 8:16 PM, Curtis Villamizar <[e-mail]> wrote:
>  
> >
> > In message <BANLkTinDb-P=[e-mail]>
> > Greg Mirsky writes:
> > >
> > > Dear Anton,
> > >
> > > I believe that MPLS-TE FRR does not address tasks 2 through 5 in the
> > > same sense as it is required in IPFRR. Certainly, protecting LSP has
> > > to be calculated by SPF, signaled and RIB/FIB/HW properly updated. But
> > > these actions all done prior to when an MPLS-TE deemed protected. Upon
> > > Fault detection the only action required is on the PLR.
> > >
> > > Regards,
> > > Greg
> >
> > Greg,
> >
> > IPFRR does not need tasks 2 through 5 either.  OTOH, IPFRR coverage is
> > often less than full coverage.
> >
> > In both MPLS FRR and IPFRR, if protection works it is handled entirely
> > by the PLR.  In IPFRR, some PLRs have no fast protection and have to
> > rely on flooding.  In IPFRR and MPLS FRR sometimes unexpected multiple
> > failures occur since a previously unknown shared resource is
> > discovered the hard way or an extroidinary event occurs (ie: two
> > fibers on same fault line, etc).  In this case even protection from
> > the MPLS FRR PLR doesn't work.
> >
> > In MPLS if a reroute is required, the CSPF load being N^2 log N (order
> > N CSPF computation have to be run), the LSA flooding has no
> > significant impact at all.  In IPFRR where only one SPF has to be run,
> > flooding is still not the primary contributor to convergence time.  It
> > may be a combination of 4 and 5 below.
> >
> > Curtis
> >
> >
> > > On Mon, Apr 11, 2011 at 1:54 PM, Anton Smirnov <[e-mail]>
> > wrote:
> > >
> > > >   Hi all,
> > > >   even though I put OSPF-FN draft in the subject it is the framework
> > > > approach FN-FRWK which draws more questions. At the very first line it
> > > > reads:
> > > >
> > > >  This document describes an architectural work that competes with the
> > > >> IP Fast Re-Route (IPFRR) solution
> > > >>
> > > >
> > > > Lets compare speed of traffic restoration between the two. So,
> > traditional
> > > > OSPF convergence time consists of the sum of:
> > > >
> > > > 1. Failure detection
> > > > 2. LSA origination
> > > > 3. Per-hop flooding
> > > > 4. SPF (delay and calculation itself)
> > > > 5. RIB/FIB/hardware update
> > > >
> > > > 3, 4 and 5 all can be significant depending on network size, number of
> > > > routes etc.
> > > >
> > > > FRR (both MPLS TE FRR and IPFRR) address 2-5. With good implementation
> > FRR
> > > > should be by order of magnitude as fast as 1.
> > > >
> > > > FN addresses only 3. It doesn't address 4 and 5. As I wrote above in
> > many
> > > > networks they are at least as significant as 3.
> > > >
> > > > So, by the speed of convergence FN doesn't look to come anywhere close
> > to
> > > > FRR.
> > > >
> > > >
> > > >   Now, lets look at FN from another perspective. Router announcing
> > failure
> > > > doesn't benefit from FN. Its immediate neighbors do not benefit from FN
> > > > either - 1 hop traditional flooding should be as fast as 1 hop FN
> > flooding.
> > > > It is only distant routers who benefit from the FN - and the farther is
> > > > router from the failure the bigger is gain.
> > > >   On the other hand, if there exists path alternative to the failed one
> > > > then _typically_ it is not too far (in terms of hops) from the failing
> > one.
> > > > I.e. from point of view of router which is 15 hops away from the point
> > of
> > > > failure it is less likely that routes will change. BTW, ordered FIB
> > approach
> > > > builds on concept that 'old' routes on remote routers do not cause
> > traffic
> > > > blackholing or loops.
> > > >
> > > >   The big problem with FN approach is that routers remote from the
> > point of
> > > > failure benefit most - but at the same time their convergence is the
> > least
> > > > important for end-to-end traffic restoration.
> > > >   The worst case network for FN is fully meshed area. Since each router
> > is
> > > > 1 hop away from every other one FN will give no benefits.
> > > >   The best case network for FN is an area consisting of one big ring.
> > In
> > > > this case alternative path is on diametrically opposite end of the
> > network
> > > > and convergence of remote routers is crucial.
> > > >
> > > >   So yeah, FN will help remote routers to converge faster. But how much
> > > > this will improve end-to-end traffic restoration in real networks? I
> > suspect
> > > > not much. Some degree of meshiness in network topology is the norm.
> > > >
> > > >   FN is an interesting proposal but it is very far from being
> > convincing.
> > > > Pitching FN against FRR is a mistake.
> > > >
> > > > --
> > > > Anton