Re: QoS settings not mapped correctly per pkey ?

[sebastien dugue <sebastien.dugue-6ktuUTfB/bM@public.gmane.org>]

On Thu, 03 Dec 2009 11:04:16 +0200
Yevgeny Kliteynik <kliteyn-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org> wrote:

> sebastien dugue wrote:
> >   Hi Yevgeny,
> > 
> > On Thu, 03 Dec 2009 10:01:28 +0200
> > Yevgeny Kliteynik <kliteyn-LDSdmyG8hGV8YrgS2mwiifqBs+8SCbDb@public.gmane.org> wrote:
> > 
> >> Sebastien,
> >>
> >> I noticed that you found the problem in IPoIB child 
> >> interfaces configuration. Glad that this worked out well. 
> >>
> >> My question is about the note that you left in the issue:
> >>
> >>  " It looks like in 'datagram' mode, the SL weights
> >>    do not seem to be applied, or maybe this is an
> >>    artifact of IPoIB in 'datagram mode' "
> >>
> >> Have you checked that in this mode you do get the right
> >> SL for each child interface by shutting off the relevant
> >> SL (mapping it to VL15)?
> > 
> >   Yes, SL to VL mapping is OK.
> > 
> >> If yes, then what you're saying is that you see that
> >> interfaces use the right SL and VL, but you don't see
> >> any arbitration between VLs?
> > 
> >   Right, whatever the weights I put in the vlarbs tables have absolutely
> > no effect when IPoIB is in datagram mode. I don't know if it's
> > an arbitration problem (don't think so) or an IPoIB problem.
> > 
> >   Could be that due to the 2044 bytes MTU in datagram mode, iperf
> > spends much time not doing transfers and fails to provide enough
> > data to the interfaces. Don't know.
> > 
> >   Once I switched to connected mode, with a 65520 bytes MTU, things
> > started to work OK with a much better overall combined bandwidth.
> 
> OK, then "a much better overall combined bandwidth" is an
> answer here. VL arbitration kicks in only when you saturate
> the link. If you don't, there's no point doing arbitration,
> because HW is able so serve any packet that comes w/o the
> need to prioritize.

  Yep, that's the conclusion I came to.

  Might be interesting to find where the bottleneck is though, that
prevents saturating the link.

  Sébastien.

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html