The MPLS WG Archive

["HANSEN CHAN" <hchan@newbridge.com>]

Peter,

To be polite (but blunt). I believe you are wrong for the following reasons.

For a centralized solution to have accurate information it must gather information from all nodes at approximately the setup/clearning rate of LSPs in the network. This of course is impractical because the setup/clearing rate may be 1000's per second during failure/resetup events.

Other than initial setup/failure/resetup time, I do not think there will
be a huge amount of activities on the signalling plane. There might be
a few sporadic events in response to hot spots. Then following this reasoning,
the centralized solution should be able to have a pretty accurate information
of the network.

Comment?

Cheers,
Hansen

 

A far more practical approach is to have all nodes aware of general trends, via the normal OSPF/IS-IS floods and significant floods, and to then feedback information along routes that are being used either successfully or unsuccessfully. This allows those nodes that are sourcing LSPs to stay fairly close to in sync with respect to the real network state, at least for the slices through the network they are using. We have done considerable simulation and implementation of this kind of approach and it works quite well.

See the following draft for descriptions of how this works. I also have a presentation that was given at the IETF that I can forward to anybody that is interested.

http://www.ietf.org/internet-drafts/draft-ietf-mpls-te-feed-00.txt

Oh, while I am at it, I am hoping to get a last call on this document ... George, Vijay?

Cheers,

Peter Ashwood-Smith

-----Original Message-----
From:   MPLS [SMTP:mpls@apollo.mctr.umbc.edu]
Sent:   Monday, April 24, 2000 10:46 PM
To:     Chiu, Angela L, ALSVC
Cc:     HANSEN CHAN; Daniel Awduche; mpls@UU.NET
Subject:        RE: TE Extension of IGP

As far as my understanding of path computation goes, in online path
computation scenario, which Dan mentioned is more likely to be the case,
what ensures that the ingress node(s) have a consistent view of the
network, how do we make sure that IGPs have converged and each node has
correct topology as well as TE database ?

An offline tool i think "might" give a more realistic picture of the
network resource allocation state at any time, correct me if i am wrong.

Manish
 
 

On Mon, 24 Apr 2000, Chiu, Angela L, ALSVC wrote:

> Hansen,
>
> See comments in line.
>
> Angela
>
> AT&T Labs
> Room C4-3A22
> 200 Laurel Ave.
> Middletown, NJ 07748
> Tel. (732) 420-2290
> Fax (732) 368-1746
> Email: alchiu@att.com
>
> > -----Original Message-----
> > From:       HANSEN CHAN [SMTP:hchan@newbridge.com]
> > Sent:       Monday, April 24, 2000 6:17 PM
> > To: Chiu, Angela L, ALSVC
> > Cc: Daniel Awduche; mpls@UU.NET
> > Subject:    Re: TE Extension of IGP
> >
> > Angela,
> >
> > > I just want to add one other point. When a network operator wants to use
> > a
> > > small set of LSPs to remove a few hot spots in the network assuming all
> > > other links having satisfactory performance, an offline tool may be the
> > > preferred way to identify which traffic trunks that contribute to those
> > > congestion points should be re-routed away from those hot spots, and
> > then
> > > figure out what the new non-shortest path should be, and leave all the
> > rest
> > > of the traffic on its normal shortest path.
> >
> > If I understand the picture correctly, that offline tool should do the
> > following:
> >
> > 1. collect statistics from the network and analyze network trunk bandwidth
> > utilization to understand where the hot spot is
> > 2. compute the LSP paths (offline) with certain algorithm for the optimal
> > use of
> > trunk bandwidth in the network
>       [Chiu, Angela L, NNAD]  Yes. As I pointed out before, between your
> steps 1 and 2, the offline tool needs to identify, for each hot spot, a set
> of traffic trunks that contribute to the congestion and need to be rerouted.
>
>
> > But as Dan said in his message, LSPs are more likely to be computed
> > online. I
> > cannot see how a node can compute all LSPs for the whole network. It might
> > be
> > able to do a good job for LSPs originating from itself. But for LSPs
> > originating
> > from other edge LSR? I would think a offline tool is a better place to
> > plan for
> > all LSPs.
>       [Chiu, Angela L, NNAD]  Actually, it is the ingress node (the node
> where an LSP originates) that is responsible for the path computation based
> on the link state information propagated via IGP extension.
>
>       Angela
>
> > Cheers,
> > Hansen
> >
> > >
> > >
> > > As far as I know, all the online LSP path computation algorithms
> > provided by
> > > the vendors are not capable of taking the normal IP traffic into account
> > > since the algorithms only see the bandwidth requirements from traffic on
> > > LSPs. If one has to use the online LSP path computation algorithm,
> > he/she
> > > needs to use some offline tool to figure out the aggregate bandwidth
> > > requirement by the normal IP traffic on each link (excluding the traffic
> > > trunks that will be routed via LSPs), substract that from the total
> > > available bandwidth for that link, and make that the new available
> > bandwidth
> > > for the online LSP path computation algorithm to use.
> > >
> > > Regards,
> > > Angela
> > >
> > > AT&T Labs
> > > Room C4-3A22
> > > 200 Laurel Ave.
> > > Middletown, NJ 07748
> > > Tel. (732) 420-2290
> > > Fax (732) 368-1746
> > > Email: alchiu@att.com
> > >
> > > > -----Original Message-----
> > > > From: Daniel Awduche [SMTP:awduche@UU.NET]
> > > > Sent: Monday, April 24, 2000 1:59 PM
> > > > To:   HANSEN CHAN
> > > > Cc:   Anoop Ghanwani; mpls@UU.NET; Daniel Awduche
> > > > Subject:      Re: TE Extension of IGP
> > > >
> > > > Hansen,
> > > >
> > > > Online LSP path computation is the preferred operational model in many
> > > > contexts -- for many good reasons.  Obviously, service providers
> > > > have the option to activate these aspects according to their
> > > > circumstance and needs...
> > > >
> > > > As a simple rule of thumb, in networks with adequate capacity, online
> > > > constraint-based routing should suffice for LSP path placement. In
> > > > relatively under-capacitated networks, however, significant
> > > > offline effort may be required to squeeze additional utility from the
> > > > infrastructure.
> > > >
> > > > Cheers,
> > > > /Dan
> > > >
> > > > On Fri, Apr 21, 2000 at 11:06:08PM -0400, HANSEN CHAN wrote:
> > > > > Dan,
> > > > >
> > > > > I agree that most MPLS implementations perform LSP path computations
> > > > online. But I
> > > > > always thought the working LSPs deployed in the networks are still
> > > > computed
> > > > > offline. You only use online computations when you're
> > > > re-routing/repairing the LSPs
> > > > > around some failure. Is my understanding correct?
> > > > >
> > > > > Cheers,
> > > > > Hansen
> > > > >
> > > > > Daniel Awduche wrote:
> > > > >
> > > > > > Hansen,
> > > > > >
> > > > > > Yes, many (perhaps most) contemporary implementations perform
> > > > > > LSP path computations online. This is a mandatory requirement
> > > > > > in some operational contexts. It's also possible to augment
> > > > > > the online system with offline software tools.
> > > > > >
> > > > > > Cheers,
> > > > > > /Dan
> > > > > >
> > > > > > On Thu, Apr 20, 2000 at 06:04:13PM -0400, HANSEN CHAN wrote:
> > > > > > > Dan,
> > > > > > >
> > > > > > > To make sure I understand. Do you mean the path of LSPs is
> > computed
> > > > on the
> > > > > > > node, not by software tools?
> > > > > > >
> > > > > > > Thanks,
> > > > > > > Hansen
> > > > > > >
> > > > > > > Daniel Awduche wrote:
> > > > > > >
> > > > > > > > Actually, the original assertion/generalization is false
> > > > > > > > (i.e. that "LSPs in today's MPLS network are usually computed
> > > > > > > > off-node").
> > > > > > > >
> > > > > > > > Cheers,
> > > > > > > > /Dan
> > > > > > > >
> > > > > > > > On Thu, Apr 20, 2000 at 02:08:22PM -0400, Anoop Ghanwani
> > wrote:
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > > I am trying to understanding the use of TE extension of
> > IGP in
> > > > a MPLS
> > > > > > > > > > network. From my understanding, you need TE extension when
> > > > you're doing
> > > > > > > > > > on-node path computation. However, since LSPs in today's
> > MPLS
> > > > network
> > > > > > > > >
> > > > ^^^^^^^^^^^^^^^^^^^^
> > > > > > > > >
> > > > > > > > > We're hoping it won't stay that way forever because it's
> > > > limiting
> > > > > > > > > to have to rely on offline tools for all traffic engineering
> > :-)
> > > > > > > > >
> > > > > > > > > That means that traffic engineering would need to be more
> > > > dynamic,
> > > > > > > > > and the routers would play a more active role in determining
> > > > paths
> > > > > > > > > and possibly doing network optimization.  Hence the IGP
> > > > extensions.
> > > > > > > > >
> > > > > > > > > > are usually computed off-node (in software tool), why
> > would
> > > > the use of
> > > > > > > > > > TE extension be critical?
> > > > > > > > > >
> > > > > > > > > > Appreciate if someone can shed some light on this
> > question.
> > > > > > > > >
>