[c-nsp] Etherchannel load-balacing change on live network

[James Ventre]

On Thu, Jul 7, 2016 at 4:57 PM, Gert Doering <gert at greenie.muc.de> wrote:

>
> Since these things are not flow based, there is no "current" traffic -
> every new packet gets an individual decision, so when you turn the
> switch, the next packet gets switched to the new rule, and all previous
> packets are already gone.
>


​It's not flow based, but if changing the hashing algorithm also withdraws
and (re)assigns ​the hash result buckets on the channel member ports, as it
reprograms the ASIC, that can cause packet loss.  Years ago I tested the
hash bucket reassignment on a 3750E and I saw a handful of packets dropped
at 1gbps speeds.  I was using a Smartbits, but I don't recall the
tested/fixed packet sizes that I used.