BGP Convergence: Small MPLS WAN

msrohman · ‎11-19-2008

All,

Our company consists of 10 sites connected via an MPLS WAN across the US.

The MPLS WAN is managed by a service provider.

We use BGP for the WAN routing protocol. Each site has its own AS number. The MPLS clound represents one AS (again, see attached diagram).

Because of the small amount of sites and only a few hundred routes (in the routing table), I'm looking to improve MPLS convergence times. The BGP timers for updates and peer connections are currently at the default settings.

Does anyone have any recommendations for timer settings?

I would like to lower convergence times for inserting / withdrawing routes that get redistributed from the LANs of these MPLS sites.

As always, thanks for your responses.

-Mike

Joseph W. Doherty · ‎11-19-2008

I believe the default BGP timers are more in-line with Internet BGP rather than a small private MPLS BGP. I've found using the minimums seems to work well, although you might also inquire with the MPLS vendor. (The vendor I believe you might be using, was surprised when I tuned down BGP timers for faster convergence, although their PE routers didn't seem to have an issue.)

PS:

If you're not already doing so, and if your addressing scheme permits it, you might just advertize a single address block per site. Even minimum EBGP timers are rather slow compared to what we expect from IGPs like OSPF or EIGRP.

msrohman · ‎11-19-2008

Thanks Joseph,

Should I tune down the BGP timers for scanning and advertisement only. Or should I look at keepalives and holdtimes as well?

-Mike

Joseph W. Doherty · ‎11-19-2008

The two I've set are:

bgp scan-time 5

timers bgp 5 15

Which, I think, is yes to both your questions.

msrohman · ‎11-19-2008

Thanks again for the info. I rated your helpful posts.

-Mike

Giuseppe Larosa · ‎11-19-2008

Hello Mike,

you can use:

ebgp external fallover to tear down the BGP session if link goes down.

Then you can implement very low timers for a simple reason: every router in a MPLS VPN model peers only with the directly attached PE node.

I think you can use 1second for keepalive and 3 seconds to tear down the BGP session.

However, it has to be noted that the MPLS VPN provides the advantage of a single eBGP neighbor but your end-to-end convergence will be dependent from the timers and the choices of your ISP.

A network change in a remote site has to travel through the MPLS VPN service provider: at least one level of BGP route-reflector is involved and BGP and MP BGP provides scalability by working with scheduled timers:

the default vpnv4 BGP scanning timer is 15 seconds. Then there are other timers involved.

This is to advise that getting a 45 seconds convergence instead of the current time (more then 180 seconds with default timers for some events) can be a good result.

Another important aspect is if the Service Provider has implemented some additional measures to improve convergence:

if your sites are double attached to the SP as shown in the picture the SP can make all advertisements travel (from both PE-CE pairs) if it uses a different RD.

you cannot see this directly.

But this improves convergence because it saves the propagation time of the least preferred PE-CE pair over the provider network.

The PE near to you has already in its table the backup route for the remote site.

Hope to help

Giuseppe