[c-nsp] mpls and etherchannel

[Lukas Tribus]

> On 20/Apr/16 20:12, Lukas Tribus wrote:
>
>> If we would know the platform, we could tell you more, but generically
>> speaking, ECMP doesn't behave any better or worse than port-channeling.
>
> Beg to differ on this one.
>
> We have converted some LACP links to native IP/MPLS to get ECMP because
> of issues with some of the payload not being equally hashed across the
> LACP member links.

What I meant (with "generically speaking"), is that there is no reason for
a box to differ in the load-balancing behavior just because of ECMP vs LACP,
because the NP lookup to create the load balancing hash has the same
exact cost (its not like LACP is a encapsulation so the NP has to look
deeper into the packet).

That doesn't mean all boxes use the same load balancing hash algorithm
for ECMP and LACP, it just means that there is no reason this data
plane load balancing algorithm has to differ between LACP and
ECMP.

I'm talking about the load-balancing hash here. Its obvious that
ECMP can load-balance in all kinds of l3 topologies where as
LACP only balances towards the same box.


ASR9k is a platform that behaves the same from a load-balancing
hash perspective with both ECMP and LACP afaik.



>> Stop it. Don't do per packet load-balancing. Repeat after me:
>> I won't per packet-load balance.
>
> We have a practical use-case where per-packet load balancing works
> better than per-flow load balancing.
>
> This has worked without issue as the fibre connections are inside the
> data centre (less than 10m between the devices). This is on a Juniper MX
> router.

Sure. You know what you are doing, you probably tested this extensively
before deploying and knew about the OOO danger. But its not someting
I would recommend anyone without thorough understanding of the
problem.


cheers,
lukas