Multilink traffic utilization

[Tshi M]

Hi,

Other than the logical bandwidth difference between let say 4T1 vs. 5T1s, is there any other difference? I am asking because I just saw one T1 that is part of a bundle of 3T1s reporting high utilization.

Regards,

p.s. any good link will be good.

[Rick Morris]

With multi-link or any other type of NxT1 application you get what is called load sharing. It is not a 50% balance as some people think. Traffic may prefer one link over another, or traffic coming in through the provider may use one link over the other. To avoid over utilization on a link you can set up BGP and shape your traffic or set up SLA's to balance out over used lines.

[Tshi M]

Thanks Rick for the quick reply. According to this link: http://www.cisco.com/en/US/products/hw/modules/ps2033/products_white_paper09186a0080091d4b.shtml

Multilink provides load balancing and not load sharing. I have read elsewhwere that there is difference between load sharing and load balancing.

Thanks,

[Rick Morris]

Despite what Cisco says, you proved that this is not the case with your set-up.

I do not argue the fact that Cisco says this, and believe that by the design of Multi-link that it "should" but as we see in the real world it is not always that way. Cisco does one thing...book answers, and then there is the rest of us with real world experience.

You should get close to 50/50 but never a balance. I think the closes I have ever come was 60/40 and that was on a good day.

If you spike one line I would check errors, and routing with the carrier.

[Tshi M]

:-) I second you on books answer and really world experience. You mentioned SLA in your first posting, do you have any sample config that i can use as a template?

Regards,

[Rick Morris]

Here is a good link to start at.

http://www.cisco.com/en/US/docs/ios/12_4/ip_sla/configuration/guide/hsla_c.html

[Joseph W. Doherty]

As Rick notes, it's not unusual that individual bundle links don't show equal utilization. Your Cisco reference, in your 2nd post, also notes different effective load balancing efficiencies for different technologies in Table 1.

Perhaps to clarify the difference between load sharing vs. load balancing: "Load sharing is the ability to distribute outgoing traffic (or influence the flow of incoming traffic) over multiple paths. Load balancing is the ability to split the load toward the same destination (host or IP prefix) over multiple paths."

[Tshi M]

Thanks Joseph and Rick,

Yes, I do understand the difference between the two definitions. I guess my original question was if there were any difference other than the virtual bandwidth between using odds or even numbers of circuits. I was told that if you use an even number of circuits, the traffic will evently distributed but if you use an odd number of circuits, one of the circuit will have more traffic. Logically this theory makes because you obviously can't evenly shared with odd number. But I was looking for official docs that support it and a method of getting closer to that.

[pauloroque]

I disagree...

A multilink (PPP) is an aggregation of many different physical links that acts as a single logical link (a bundle). In this case the load sharing/balancing is precisely equal over all members of the bundle. This is different of having many links toward a single destination. In this case the load sharing/balance is not precise.

PPP multilink permits things such as packet interleave and fragmentation over the members.

Some very small variation may occur if interleave is not enabled.

I have a setup where there is 3 links in a bundle. In there, the traffic is distributed precisely over the 3 members.

Is 'ppp multilink interleave' enabled in your config? Some type of queue mechanisms are not compatible with interleave and fragmentation. What is the amount variation do you notice in your implementation?

PRoque

[Tshi M]

Hi Paulo,

Unfortunately I don't have access to the router since it belongs to the service provider. I will try to gather more info from them.

Regards

[Joseph W. Doherty]

"I was told that if you use an even number of circuits, the traffic will evently distributed but if you use an odd number of circuits, one of the circuit will have more traffic. "

Could be true based on how technolgy distribute traffic across links. For instance, could that have been in reference to Etherchannel? For MLP, if it's not doing some type of link hash, but if doing round-robin (as mentioned in you reference document) or if sending traffic to next interface without traffic, shouldn't be true.