[j-nsp] CoS buffer size

[Dan Peachey]

On 24 June 2015 at 21:05, Saku Ytti <saku at ytti.fi> wrote:

> On 24 June 2015 at 22:29, Dan Peachey <dan at illusionnetworks.com> wrote:
>
> Hey,
>
> > I thought the weights were determined by the %? The weights are then used
> > to schedule the queues appropriately. Even if the queues are in excess,
> > they should be weighted correctly?
>
> I tested this when Trio came out, and couldn't get QoS working,
> because my AFB class was congesting my BE class fully. Solution was to
> configure guaranteed-rate==shaping-rate, at which point the
> percentages were honoured. This is actually what DPCE generation did
> by default. I think the reason for grate being 0 by default, is
> because trio supports additional level of shaping, in which use-case
> it may make sense.
>
> As I recall you can oversub grate, but I can't recall testing it (For
> my scenarios, majority of traffic is non-qos traffic, so it's not an
> issue either way).
>
>
Hi Saku,

I've run some tests and experienced the same thing as you. I had BE traffic
affecting one of my other queues which was not oversubscribed and caused
packet loss. Setting the guaranteed rate = shaping rate seemed to fix it.
Seems odd to me that this needs to be done. Documentation I've read appears
to suggest that in PIR mode (no guaranteed-rate set) the per-queue
guarantee/transmit rate is calculated from the shaper rate and when a queue
exceeds it's guaranteed rate it is in excess, but this doesn't appear to be
the case and queues are always in excess.

Thanks,

Dan