[afnog] IP transit and Load Balancing

[Nishal Goburdhan]

On 9 Jun 2015, at 18:25, Joey ESQUIBAL wrote:

> Dear AfNOG,
>
> I trust this e-mail finds everyone well.
> May I ask your opinion and or best practice about having multiple IP 
> transit to 4 different IPLC's.
> For argument’s sake, I have allocated sample capacities on each 
> circuit below:
>
> IPLC 1 (1 x STM4) => IP Transit 1
> IPLC 2 (1 x STM4) => IP Transit 2
> IPLC 3 (2 x STM1) => IP Transit 1
> IPLC 4 (2 x STM1) => IP Transit 2
>
> All circuit ends to different geographical locations.
>
> I’m thinking of having two IP transit (which I think should be good 
> already) and advertise specific prefix(es) on each peer.

hi joey,

in general having two transit providers is a good thing.  three is still 
ok, but more than that (ie. four and greater) becomes a nightmare to 
deal with, so, you’re already off to a good start, if you have just 
two;  no reason to change that!

the next part of multihoming deals with knowing when, and where, to 
advertise your subnets.  since these are in four separate locations that 
will make it a little more difficult.  one (medium term) suggestion i 
would make, is to figure out how to get them to, say, two locations.  
being able to make consistent advertisements from centralised locations 
helps to make your network easier to manage as your grow.

immediately, you’ll want to announce your aggregate prefixes across 
all of these links.  then, a good strategy would be to chop up the 
aggregate into half (initially) and advertise these more specifics (as 
well as the aggregates) across the links you need to load - in this 
case, your STM-4 circuits.  you might then also need to chop up those 
again, to again, artificially create load.  that’s ok to do really.

the trick here is to remember that you _should_ advertise your aggregate 
out of all of the circuits (should there be some sort of failure).  and 
i would strongly advocate that you work towards centralising the two 
locations that you pull bandwidth into your network;  eg. a pop in LDN 
and/or AMS, or similar… because then you can manage this 
de-aggregation inside your network, and make more effective use of your 
bandwidth.

i’ve not gone into a lot of detail, since there’s a lot that can be 
written here.  feel free to ping me if you need more specifics.

—n.