Load Sharing, Redundancy with multiple links and multiple routers...

[scott.foster]

Here is a question for which my colleagues and I have not found a sound resolution. We want complete redundancy between two locations with routers and wan links. We have two 2611s remotely and two 3662s locally. 2611(A)<-T1->3662(A) and 2611(B)<-T1->3662(B). 2611(A)<-FE->2611(B) and 3662(A)<-FE->3662(B). The traffic relationship is one-to-one meaning that only one host on the remote side of this circuit talks to one host locally, but there is a significant amount of traffic involved here.

We would like to load share traffic between these circuits in a way that also provides redundancy in case a router goes out. We have looked at GLB (Gateway Load Balacing) which looks to be based on the assumption that many hosts are trying to get to a cluster of routers. Multichassis Multilink PPP which looks to be a one-to-many traffic scenario. Multilink PPP only gives us redundancy on a bundle of T1s between to routers. Should a router die, then we are dead as well. EIGRP for load balancing, but the reported distance is equal to the feasible distance therefore the second route is not a successor and can't participate in unequal or equal cost load balacing. So right now we are treating each T1 as a unicast path, one for outgoing traffic and the other for inbound traffic. Another approach we have is to load share by using static routes to the same destinations, but specifying two ways of getting there. I would have to put in some policy based routes to take care of situations when traffic is sent to a router where the serial link is down.

Anyway, we haven't been able to find a good solution to do what we want. Is there any more ideas out there of what I can do?

Scott Foster

scott.foster@merrickbank.com

801-545-6674

[t.baranski]

I've never tried this before but you may be able to use OSPF or RIP and cost manipulation to do what you want. What I'm thinking is that if you manipuate the cost of one of the T1's to be 1 greater than the other T1, then the routers connected to the higher-cost T1 should see routes advertised via both T1's as having equal cost (assuming the cost of your FE interfaces is 1).

For example... Consider traffic from the 3662 network to the network behind the 2611s. Say that the cost of T1-A is 10, the cost of T1-B is 11, and the cost of the FE interfaces is 1. 3662A's cost to Net-2611 via its serial interface will be 11; 3662B's cost to Net-2611 via its serial interface will be 12. When 3662A advertises its Net-2611 route to 3662B, 3662B will have two equal-cost routes to Net-2611 and can load-balance on a per-packet basis if configured to do so. The route will also be advertised in the other direction (3662B -> 3662A), but 3662A will prefer the cost via T1-A, so this won't result in a routing loop.

The same applies to traffic going the other direction as well. One thing to note is that the routers connected to the higher-cost T1 (the "B" routers in this case) need to be the default gateway for their respective networks, as they're the only routers that will have the equal-cost routes in their routing tables. The best way to accomplish this, presumably, is HSRP between both router pairs with the "B" routers having a higher priority and their serial interfaces "tracked" such that the "A" routers take over when these interfaces go down. When a T1 goes down, route advertisements are no longer heard via that T1 so everything should begin flowing over the other circuit.

Again, I haven't tried this so I may be missing something. But I don't see a reason offhand why it wouldn't work. I'd be interested to hear your results if you try it, or any thoughts that anyone else has.

[mark.edwards]

I don't see why you can not configure EIGRP with variance command to allow unequal load balancing. You can change delay/bandwidth to make sure the route is a feasable sucessor then configure variance to enable routing accross the path.

[ruwhite]

You could probably adjust the eigrp metrics to make one of the paths be worse than the other, so you can use variance. Of course, you can only do this in one direction, since you can only have one of a pair of connected routers as the fs of the other router in this sort of situation.

The second part of the problem is going to be getting CEF to switch over both paths, if you're really using just one source to one destination. To do that, you're going to have to use per packet load sharing, which might cause you to have a lot of out of order packets, reducing your performance to the point that it doesn't matter whether or not you can get traffic across that link.

Russ.W