[c-nsp] QoS: wrr-queue queue-limit & bandwidth

[Peter Rathlev]

On Thu, 2008-06-26 at 21:18 -0300, J C <juniper84 at live.com> wrote:

Shouldn't that be "cisco84 at yahoo.com"? ;-)

> I'm having some issues trying to de-cypher and determine real-word
> usage for the following two commands:
> 
> wrr-queue bandwidth
> wrr-queue queue-limit
> 
> I believe I understand the usage of 'wrr-queue bandwidth' more than the
> other command.  My understanding of this command would be to allocate
> weights for queues that would influence bandwidth for queues on the
> egress of an interface.
> 
> I'm sorta lost on the other command.
> 
> Any kind of real-word example would be appreciated.  I learn better
> from backwards engineering than reading Cisco terminology.

The bandwith command set the weights (for a WRR or DWRR scheduling) for
each of the queues regarding how much "raw" bandwith they can use on the
link. If you have a 100mbps link, a "wrr-queue bandwidth percent 10 90"
would reserve 10 mbps for queue 1 and 90 mbps for queue 2.

The queue-limit command splits the internal transmit (for wrr-queue)
buffer up between the queues. So "wrr-queue queue-limit 20 80" reserves
20% of the buffer space for queue 1 and 80% of the buffer space for
queue 2.

Combining these, you could reserve e.g. 10 mbps for queue 1, but maybe
10 MB buffers, and the 90 mbps for queue 2, but maybe only 10 MB
buffers. If both queues use their allocated amount of bandwidth, the
queues never fill up. 

If traffic in queue 1 would start to burst there would be 10 MB of
buffer space to hold on to this burst, letting it out at 10 mbps.
Theoretically this means that queue 1 could burst at 20 mbps for 4
seconds before the buffer was full. This would result in a delay of the
traffic, but no loss.

If traffic in queue 2 would start to burst there would only be 1 MB of
buffer space. So if it bursts at 98 mbps, the queue could hold a seconds
worth of the excess traffic (8 mbps). After this, you would start to see
drops.

Real life use would depend a lot on what you would try to accomplish.
There's a big difference between edge (facing clients), enterprise core,
service provider core and so on. And the priority, threshold and random
detect elements are other important things to consider.

Regards,
Peter