[c-nsp] question about service provider network design

[Adam Armstrong]

Nathan wrote:
> On Tue, Oct 21, 2008, Adam Armstrong <lists at memetic.org 
> <mailto:lists at memetic.org>> wrote:
> > Nathan wrote:
> >> - Is running OSPF on a switch at all useful when the switch is
> >> connecting routers that are running MPLS, MP-BGP, and OSPF? Can it
> >> provide faster detection of link loss?
> >
> > The routers can see each other directly at L2? Then no. It might make it
> > easier to keep the switch's management loopback connected though.
>
> Well I don't see how the LDP would keep running if the switch cut off 
> L2. The switch would need to speak LDP . . . which would make it an 
> MPLS P router, which would be cool but I'm quite sure neither 2960s or 
> even 3550s can do that :-) P router with eight gigabit ethernet ports 
> running at line speed for the price of a 2960 anyone? Seriously, what 
> kind of beast does that? A 7600 or 6500 I suppose, anything smaller?
Umm. I've no idea what you're talking about now... The switch doesn't 
speak LDP. It can merely participate in your IGP for its loopback address.

Just give the switches an IP in the subnet that exists on their layer 2 
domain and point their default route at one of the PEs (or do hsrp 
between a couple of them).
> > Consider switching to IS-IS, assuming your kit can do it.
>
> The switches can't, but I do think the routers can. What would the 
> benefits be? If I change to IS-IS, now's the time.
Well, the switches aren't important here, so if you plan to do ipv6 in 
the future and aren't a huge ospf fan, have a look at isis now and 
switch if you like it. It's definitely a lot easier to manage and 
troubleshoot. Not to mention not having to run two versions of ospf when 
you want to do ipv6!
> > Do you have a diagram?
>
> I'm not sure that ASCII art will cut it, but I'll try . . .
>
> First option:
>
>  /----------SW----------WAN---------SW-----------\
>  |  |  |     |                       |     |     |
> PE PE PE     |                       |    PE PE PE
>  |     |     |                       |     |  |  |
>  \----------SW----------WAN---------SW-----------/
>
> This way I don't have to have each PE connected to both switches in 
> order to communicate directly, it's only when a switch goes down that 
> PEs only connected to that single switch will have a problem. I'll 
> have to place different VLANs on top and bottom and use MST so that 
> both links are used. If I lose the ethernet link on a WAN link, MST 
> notices immediately and reroutes traffic.
>
> Second option:
>
>  /----------SW----------WAN---------SW-----------\
>  |  |  |                                   |  |  |
> PE PE PE                                  PE PE PE
>  |  |  |                                   |  |  |
>  \----------SW----------WAN---------SW-----------/
>
Second option is the sensible one. Think of it as building 2 core layer 
2 domains across witch all of the PEs can talk to eachother. During 
normal operation, they balance across the two domains, when a switch or 
link dies, the traffic goes across the other. It's a relatively standard 
design.

http://alpha.memetic.org/basic.jpg is how i would draw it.
>
>
> > When you say WAN, what do you mean? A long distance ethernet circuit? 
> Or a
> > Serial/Pos/etc?
>
> Thay are seen as gigabit ethernet (copper or fiber), but they run over 
> the national backbone of bigger fish than I.  They are probably AToM 
> pseudowires. Unfortunately that means that when one goes down (not 
> often, maybe once or at most twice a year) I don't always lose the 
> ethernet link (and I suppose I might get one-way communication only).
Well, tune your IGP so that it notices as quickly as possible and pulls 
down the link.

You want as few routes as possible in IGP (so just links and loopbacks), 
but i guess you already knew that! :)

adam.