[c-nsp] QoS concerns

[Pshem Kowalczyk]

Hi,


On 11 June 2010 04:33, Mihai Todor <mihai.todor at gmail.com> wrote:
> Hi Arie,
>
> Thanks for your feedback!
>
> I'll try to clarify my question 1 - it would be useful to know if
> (experience showed) there is such a thing as a maximum for the PQ bandwith
> allocation. Let's say you have just one priority class (in a child policy)
> and you police it at 80% of the available bandwidth (conditioned through
> shaping in the parent policy). Will CBWFQ/MDRR within the child policy be
> able to accurately distribute & assure the provisioned bandwitdh between the
> other classes when the priority class is congested? I'm actualy concerned
> about allocating 70% to the priority class.

We try not to exceed about 35-40% of strict priority traffic on
'internal' (inside the network) interfaces. Firstly to make sure that
the other classes do not get starved but more even more important is
maintaining the actual quality of EF.  With huge amounts of EF
(especially when its an aggregate of many streams) the scheduling
within EF becomes indeterministic, which might impact your jitter.
If 70% of the traffic on an interface is 'priority' then you have to
expect other classes being impacted under congestion. That's the
nature of priority traffic. Actual quality of that policing and then
scheduling might be hardware dependant. We've tested that on ASR9k
(MDRR) and results were as expected - 'lower' classes' where impacted
more then 'higher' ones whilst all of priority was getting through,
and we deemed that acceptable.

kind regards
Pshem