> True, but many existing DC switches that do IP-based load balancing have
> no problems load-balancing on UDP 5-tuple. I am positive there's equipment > 
> out there that can do load balancing on GRE keys (or some other part of 
> GRE header), I just haven't encountered it yet (or realized it does that),
> so please fix my ignorance.
..
> Ivan

Hey Ivan,

Looking at the specs of one of the most popular chips there are generic CAM 
rules that allow you to override some of the 'default' hash behaviors. So I 
suspect you could program a number of CAM rules that match the NVGRE Ethertype 
and the inner src/dst IP low bits to give you the hash index .. but of course 
you'd burn CAM quite quicky but that's probably better than trying to use more 
src/dst Ips to get the entropy.

Realistically we can have LAG groups with many 10/40G links that need to be 
spread over. Especially between a monster core router and a monster core 
switch. So using a couple of extra IP addresses per hypervisor is not going to 
be adequate. I.e. does not provide enough entropy for such a wide LAG. So in 
that case and the vanilla ECMP case you'd probably be able to trade CAM space 
v.s. entropy but that's not a cheap solution.

Peter








_______________________________________________
nvo3 mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/nvo3

Reply via email to