[c-nsp] Equal cost load balancing between geographically dispersed sites

[Dale Shaw]

Hi,

[This message is quite long.]

I am trying to accomplish a routing configuration with equal cost load
balancing that is breaking my mind.

The topology can be viewed here:
http://i9.tinypic.com/6c4m1aw.jpg

Some facts:

- The routing protocol is EIGRP
- The two WAN links are both 64Mbps
- "wan1" and "wan2" have the same view of the WAN routes
- No weird and/or wonderful routing protocol knobs have been turned
- The Etherchannel between "core1" and "core2" is layer 2 (dot1q) trunk
- All other links shown are layer 3 ("no switchport")
- All switches depicted are Cat6500s. WAN head-end routers are
7200-G1s and the branch office WAN routers are 2800s.

At each distribution switch ("dist1" and "dist2"), I need to be able
to load balance traffic out to the WAN. At this stage I'm only
concerned with outbound (LAN to WAN) load balancing. This is the only
key requirement - effective use of both (expensive) WAN links is my
goal.

There is no requirement to do traffic load balancing in other areas of
the network.

Irrespective of the source of the traffic destined for the WAN, it
should be load balanced between the two WAN routers. This means
traffic could be sourced from "MAN office A" or "B" or "C" or the
"servers" cloud or from any of a number of other sites in the MAN.
This means I may need to send traffic arriving at the distribution
switches back out the interface it came in on.

I suppose this means I need to do something with VLANs, or tunneling,
or some other feature I'm not familiar with. In other words, I need to
form routing adjacencies over interfaces other than the physical
interfaces.

My first attempt at a design was to change the links between the core
and distribution from L3 to L2 trunks and then:

1. Create 4 "dedicated", "point to point" VLANs (core1/dist1,
core1/dist2, core2/dist1 and core2/dist2) and then form EIGRP
adjacencies over SVIs. Permit each VLAN to traverse the 4 respective
dot1q trunks.

2. Create a new VLAN and allow it to exist in the core and
distribution layers. Configure L3 SVIs for this VLAN on the two
distribution switches and place the WAN router LAN-side interfaces in
this VLAN as access ports.

This allows "dist1" and "dist2" and "wan1" and "wan2" to be L2
adjacent. Routing adjacencies can be formed and each distribution
switch will see two equal cost paths to WAN destinations (via the two
WAN routers). The problem is that "dist1" and "dist2" will also learn
these routes via their adjacencies with the core switches.

At this point dead flies and sawdust exploded from my head and I
figured there must be a better way.

I want to end up with something manageable and I want to avoid having
to muck around too much with EIGRP metrics, variance etc.

This is an existing, production network and while I can make changes
as required, re-designing the whole thing is not really viable.

I've probably missed some key points, but that's my braindump for now.

Chances are I'll look at this again tomorrow and it'll become obvious
but in the meantime (and in case it doesn't), thanks for any insights!

cheers,
Dale