[c-nsp] FW: Overruns

[Brett Frankenberger]

On Thu, Jan 27, 2011 at 08:53:18AM +0000, Nick Hilliard wrote:
> On 27/01/2011 07:57, Mohammad Khalil wrote:
> >its on Cisco 7606-S , the connection is port channel with 5 physical interfaces
> 
> Oh, you Really Don't Want To Do That(tm).   For etherchannels on
> EARL7 architecture, if you want your load balancing to be roughly
> equal, you need to ensure that your port channels are configured
> with either 2, 4 or 8 physical interfaces.  The reason for this is
> due to limitations on the EARL7 chip on the sup720 - specifically,
> there are only 3 bits of bucket space, which means 1) no more than 8
> active links and 2) severe limitations in the load balancing
> algorithm.
> 
> If you have 5 physical interfaces, the load balancing will work out
> (in the optimal case) as 2:2:2:1:1.  This means that you effectively
> have (2+2+2+1+1)/(2+2+2+2+2) = 4/5 of the total etherchannel
> capacity available for traffic.  I.e. you're actually not gaining
> anything by using more than 4 physical links.

Of course he is.  With five links, assuming the traffic hashes evenly
across the 8 buckets, he effectvly has 4GBps of throughput available. 
If one of the five links fails, he still has 4GBps of throughput
available.

With four links, assuming the traffic hashes evenly across the 8
buckets, he effectively has 4Gbps of throughput available.  But if one
of the four links fails, then he'll hash at 3:3:2 and effectively have
2.67Gbps available.  

In other words, the fifth link doesn't add any throughput benefit in
the "everything is working" case -- four or five active links offers
the same throughput -- but it offers a significant redundancy benefit. 
With five links, loss of one link is no impact to cpacity; with four,
loss of one link is a 33% reduction in capacity.

     -- Brett