MPLS allows you to do the best of both worlds. You can set your TE metrics to the latency/route-miles and then build bandwidth constrained LSPs from place to place and they will take the least latent, uncongested path that they fit on. Similar things can be done with ATM routing costs, and contracts on spvcs. Both of these are considerably more work, but it gets traffic flows on good paths very well. There is at least one large ISP doing metrics in that manner. -----Original Message----- From: Daniel Golding [mailto:dgolding@sockeye.com] Sent: Fri 7/19/2002 4:43 PM To: Joe Abley; Me Cc: Sush Bhattarai; nanog@merit.edu; Tom Holbrook Subject: RE: IGP metrics on WAN links I suspect the approach you take depends on how your network looks. If you have many pipes of a variety of sizes, doing IGP metrics based on pipe size makes a good deal of sense, then adding twists for things like ckt latency. However, folks with uniform sized networks, and uniform traffic between coasts probably tend to set IGP metrics for latency, with pipe size being the exception that they bias for afterwards. The latter is probably more prevelent in an established network, the former in a network undergoing a large fiber build. - Dan

-----Original Message----- From: owner-nanog@merit.edu [mailto:owner-nanog@merit.edu]On Behalf Of Joe Abley Sent: Friday, July 19, 2002 4:25 PM To: Me Cc: Sush Bhattarai; nanog@merit.edu; Tom Holbrook Subject: Re: IGP metrics on WAN links

On Fri, Jul 19, 2002 at 02:11:29PM -0600, Me wrote:

...

I think you missed part of his comment: " of course there are always some "twinking" done regularly to give higher priorities to the higher bandwidth, link condition etc"

so fiber mileage is just the base, with modifications to make it work correctly, based on bandwidth, etc.

Yeah, my (limited) experience is the opposite. At the previous large operator at which I had enable, the IGP metrics were chosen primarily according to circuit size, and were subsequently tweaked for other issues (such as circuit latency, or the requirement to balance cross- US traffic across non-parallel circuits).

In my experience, congestion is a much more effecive killer of service than latency due to optical distance. Hence attracting traffic to circuits where there is more likely to be headroom seems a more reasonable first-order approach for choosing metrics.

That experience is all in networks where intra-AS traffic engineering was done at the IP layer, however; in networks where there is a lower layer of soft traffic engineering maybe other approaches would be more appropriate.

Joe