[c-nsp] configure 876 with qos

[Ivan Gasparik]

On Tuesday 17 July 2007 12:44, Dennis Breithaupt wrote:
> Hello,
>
> I've some questions on that topic, too :)
>
> On Tue, 17 Jul 2007 10:55:41 +0200, Ivan Gasparik <ivan at ig.sk> 
wrote:
> > Hello,
> >
> > The right place to activate policy-map in case of DSL WAN
> > connection is PVC inside interface ATM0:
> >
> > interface ATM0
> >  pvc x/y
> >   cbr <speed>
> >   service-policy output voice
>
> What exactly does the "cbr"-adjustment do under the pvc? Does it
> have an real effect on data flow or is it just an
> administrative/management configuration?

Yes, cbr is a shaping method at ATM layer. There are other methods too 
(vbr, ubr, ... - I'm not an ATM expert :-)), but we use for our 
customers only cbr.

>
> We're having the problem, that we have a load auf 100% on our
> ADSL-lines on the upstream, because we're uploading massive
> background traffic. That leads to horrible high latency times. So
> we want to achieve, that we're only using i.e. 80% of the upload
> bandwidth.

In my opinion the high latency is caused by default setting of 
tx-ring-limit buffer. This buffer is applied in hardware after 
software queueing configured in policy-map. So in case of fully 
exhausted upload your policy-map will do it's work, but the hardware 
queue will add several hundred miliseconds latency to all outgoing 
packets (depends on average packet lenght).
You can decrease the value of tx-ring-limit buffer and decrease the 
time that takes a packet to pass through the interface. VoIP traffic 
configured in policy-map to take strict priority will pass more 
quickly and all other traffic will be queued in software 
(class-default in policy-map) instead of hardware in case of 
congested interface.

>
> Sadly we can't use "bandwidth average" in the service policy,
> because the 836 (in our case), but I think the 876, too, can't
> shape on ATM. (afaik)
>
> Is that the case? Our has cbr an effect on bandwidth limiting, too?

Yes, cbr does ATM traffic shaping.

>
> Can I also combine WRED with this adjustment? My tries with WRED
> were not successfull, because the OutQ was always "0".

WRED is supported in class-default, just configure random-detect 
inside it. Fair queueing is also a good choice.

Ivan


>
> Thanks, regards,
>
> Dennis
> [...]
>
> _______________________________________________
> 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/