[c-nsp] Max performance 6148(A--GE-TX boards

[Patrick McEvilly]

We have been experiencing high input packet drops on interfaces on
6148a's when we are less than ~200mb/s on the 8 ports combined in one
case and <100Mb/s in another, not anywhere close to the 8-1 over
subscription as advertised.  It also got much worse when we went from
catos to IOS last week, not sure why.  Don't expect too much from a
6148a module.




Matt Buford wrote:
>> Can anyone comment on this ? Does this mean we can get a max of 6 Gig
>> throughput on a 6148A card and max 2 Gbit on a 6148 ? Or do these
>> numbers only apply to etherchannels ? I don't seem to find the right
>> performance figures for these cards.
> 
> Issue 1:
> 
> Each group of 8 ports only supports a gigabit.  If you get 2 ports in the 
> same range of 8 both trying to download a gigabit, each one is only going to 
> receive 500 mbits.  If you have a mix of high bandwidth servers and low 
> bandwidth servers, you should probably spread the high bandwidth servers out 
> so there is only one big bandwidth user per group of 8 ports.
> 
> Issue 2:
> 
> When using etherchannel, a copy of EVERY PACKET for the entire etherchannel 
> bundle gets sent to the ASIC.  Since the ASIC only has a 1gbit capacity, 
> this means that the entire etherchannel bundle maxes out at 1gb.  But wait! 
> It isn't just the etherchannel bundle that maxes out at 1gb.  It is the 
> etherchannel bundle plus all other ports on the same ASIC.
> 
> The result:
> 
> So, lets imagine you make an etherchannel with port 3/1 and 4/1.  Both cards 
> are 6148-GE-TX.  Your etherchannel downloads at 1gbit.  Now, someone else 
> (not in the Etherchannel) on port 3/2 starts downloading.  The downloader 
> AND the etherchannel now have to share the 1gbit restriction, since the ASIC 
> handling 3/1-8 has to receive a copy of every etherchannel packet plus every 
> packet for port 3/2 - all through a 1gbit pipe.
> 
> The moral of the story is don't etherchannel on WS-6148-GE-TX cards (unless 
> doing so only for redundancy and not for bandwidth). 
> 
> _______________________________________________
> cisco-nsp mailing list  cisco-nsp at puck.nether.net
> https://puck.nether.net/mailman/listinfo/cisco-nsp
> archive at http://puck.nether.net/pipermail/cisco-nsp/