[c-nsp] GE Servers in data center with teaming interfaces

[Nick Hilliard]

On 19/01/2011 12:00, Arne Svennevik wrote:
>> On 18/01/2011 22:27, chris stand wrote:
>>> Do any of you support / host servers from GE ?
>>> Anyone have any of them in LACP port channels ?
>>
>> Yes. Be very careful with your switch uplinks if you do this, because
>> applications can be badly behaved and can cause serious problems due
>> to microbursting.
>
> Would you mind to elaborate this? Along with some pointers on what to look for to detect it?

This was a very badly worded reply to the OP - I had something specific in 
mind when writing it, and then clicked "send" before reading the nonsense I 
had written.

There are generally two reasons for implementing LAGs: resilience and 
capacity.  If you're implementing them for pure resilience and your traffic 
levels are pretty log, then that's fine.  However, if it's for capacity 
reasons and you're dealing with bursty traffic, you need to be careful with 
your uplink capacity, particularly if your uplink bearer channels are the 
same capacity as your uplink bearer channels.

As always, it comes back to badly behaved applications creating 
microbursts, bad hashing algorithms used for LAG load distribution and bad 
monitoring for not catching these issues before they become a problem.

Gert Doering posted a good example of this a couple of months ago:

http://www.mail-archive.com/cisco-nsp@puck.nether.net/msg33606.html

Nick