At 7:17 PM +0000 7/10/02, Phillip Heller wrote: >On Wed, Jul 10, 2002 at 04:42:25PM +0000, Lupi, Guy wrote: > Let me preface this by saying that I am trying to learn more about large > scale BGP design and operation. This question is on route reflectors when > you have multiple POPs in seperate IGP domains. If you currently have one > POP and are going to move to 2 within the same AS, you can either run full > mesh (doesn't scale), reflectors, or confederations. Assuming you don't > currently have a central core that the POPs connect back to, how well does > reflection scale? I was reading Building Service Provider Networks > [Berkowitz], and it states that iBGP doesn't scale well once you go above > 15-20 sessions per router. It also states that most ISPs run reflectors > instead of confederations, but I believe that statement is being made under > the assumption that the ISP will have a central core to which the POPs will > connect. This would indicate to me that assuming you don't have a central > core, one could only connect 6 or 7 POPs (dual reflectors for redundancy) > together using reflection before you would have to either create a central > core to reduce the amount of iBGP sessions, or turn to confederations. > Perhaps the best way to accomplish this would be to establish a "core" in > one of the POPs and run reflection from there, which is also presented as a > solution in the book? Any opinions? I have made an attempt at ASCII > drawing below, to me the central core solution makes more sense. > >In my experience, a core ibgp mesh will scale to at least 70 sessions >per device. I would suggest that route reflection certainly be done >between core and aggregation devices per pop.
With more modern routers, I certainly can see that working, especially depending on how well you dampen flap and generally control the churn rate. One also has to be careful about all the issues evolving in the permanent oscillation drafts and discussion in the IDR WG. 15-20 is an old Cisco recommendation that, like the number of areas per OSPF ABR, certainly can be exceeded when you understand the issues. > >Central reflection may be prone to failure depending on the design of >the network. I certainly didn't mean to imply that there is a single reflector. My book examples typically show dual reflectors in an important cluster. One of the questions, of course, is the extent to which you use (G)MPLS in the core. There are traffic engineering, shared risk group management, and other issues that can make a non-traditional, largely BGP-free core attractive. The BGP intelligence lies in the edge and aggregation routers, which inherently distribute control plane workload. > >Also, when you mention seperate igp domains, are you referring to >areas/levels, or instances? Both OSPF and ISIS scale quite well using >area or level hierarchy, which mostly mitigates the necessity for >seperate igp instances. Agreed. Indeed, there are many operational carriers that have 1000+ routers in a single area, albeit with some tuning and stable facilities. Message Posted at: http://www.groupstudy.com/form/read.php?f=7&i=48529&t=48509 -------------------------------------------------- FAQ, list archives, and subscription info: http://www.groupstudy.com/list/cisco.html Report misconduct and Nondisclosure violations to [EMAIL PROTECTED]
