I'd like to ask an advice about this situation:
We're planning to connect four different datacenters (two in each campus using two uplinks. Each datacenter has one 4506 and we'd like to connect datacenter A in campus A to datacenter A in campus B. The same strategy will be used to connect datacenter B in campus A to datacenter B in campus B. Considering that we already has uplinks between datacenters A and B in the same campus, is it possible to connect both campus using different uplink in 'load balance' mode?
Yes, your 4506 when configured as l3 switches using any routing protocol will always send traffic using the best path, and if tow or more paths are all equally good, will use them in load balancing mode. That is all automatic with a simple and basic configuration.
Thank you Paolo. I just forgot to mention that the uplinks use different transmission technology: one is a 300 Mbits radio and the other is a 1Gbit FO. What do you mean for basic configuration? What technology you think is fair to support this load balance architecture?
Routing protocols (IGRMP, EIGRMP) that support unequal cost paths, could be used. (The could be configured to try to send traffic in proportion to the bandwith of the paths.)
Hi Joseph, That was discussed many times in the past already. The thing is that so far there is no know production network using eigrp unequal load balancing, if you have one send the details and some traffic shapshots, until then it remains a paper design in my opinion.
So to address the original question, if you send a diagram of the site and how they are connected, plus an estimate of the traffic patterns, one could design in theory how is desirable to route traffic.
Paolo, I've missed those past discussions. Assuming you're correct there's no known production network doing it (which might also be since you need both different bandwidth paths and IGRMP/EIGRP -- I've always only seemed to have just one or the other), is there a reason you feel this "paper design" wouldn't work or isn't suitable? (Beyond, perhaps, being cautious about being first?)
Hi, first of all to put it bluntly I don't think EIGRP is any better that the standard OSPF, nor I have seen any situation where it solves anything that can't be done with OSPF. If you want to consider that a personal opinion, so be it, but also consider it comes backed from 15 years IP networking experience of yours truly, and some of the reasons why I think like that are better suited for private discussion than exposed in a public forum.
Then when it comes to the often mentioned but never proven "unequal cost load sharing" my objection was that knowing a little router architecture, what ultimately decides how traffic is sent over links, is CEF. I was then then told that EIGRP has "special hooks" to outsmart CEF and decide by itself how much traffic goes and where. Fine with me, but the thing still is that I haven't ever see that working in lab, much less in production.
I think that a sane approach to network engineering is to be give proven solutions that enables business to run with certainty (you may have heard that already). Then if we want to tinker with "traditionally set" networks in search of improvement, fine with me, but I my gut feeling is that the "feature" we're talking about, is not going to gain anybody of us a place is the "hall of fame", if there's such a thing in our industry.
EIGRP unequal cost load sharing works under one condition - process switching required. I had tested this in the past in my lab gear. Process switching isn't something that many would want enable because of the burden it would put on the CPU. If enabling process switching isn't a problem for you unequal cost load sharing works as Cisco documents describe.
Assuming the impact of using unequal cost routing is unacceptable, you're left with the question whether to treat the two paths as equal or primary and backup.
Since the lesser path offers less than half the bandwidth of the higher, I would recommend you route as primary and backup.
Something to keep in mind, if unequal cost routing delivers unacceptable performance in your environment today, that could change with newer software releases.
Also, perhaps not clear, I only had in mind static unequal cost load balancing. Dynamic is also possible, I believe, with at least EIGRP. Further, on the roadmap for PfR, dynamic load balancing is planned for some of the IGPs. These too you might keep a watch on in the future.