One more thing regarding latency - you can use ethtool to modify these parameters if your application is latency sensitive.
On Sun, 2008-01-06 at 10:59 +0200, Eli Cohen wrote: > On Sun, 2008-01-06 at 10:47 +0200, Or Gerlitz wrote: > > Hi Eli, > > > > The patch below in OFED 1.3 does not check the device capabilities and hence > > always fail on non connectx systems. Can you fix it such that we will not > > get all those "why I we mthca0: failed to modify CQ params prints in the > > logs". > I chose to use KERN_INFO to indicate that it is not a sever situation. > Perhaps I should just remove the message. > > > > Other then that and maybe even more important... I understand that it hard > > codes > > ipoib to ask for delivery of interrupt only after MAX (16 packets received, > > 10 us > > elapsed since first packet received), correct? so every simple ping-pong > > test that > > measures IPoIB latency under small packet rate will have now 10us added to > > its latency? > I did not notice that this deteriorates ping pong latency. Could you > check if it does? > > > > > I wonder if we can come up with simple way-in-the-middle here. > > > > thanks, > > > > Or. > > ------------------------------------ > > Set default CQ moderation parameters > > > > Set the default params to make sure they are applied. Otherwise > > we get poor performance. > > > > Signed-off-by: Eli Cohen <[EMAIL PROTECTED]> > > > > Index: ofa_kernel-1.3/drivers/infiniband/ulp/ipoib/ipoib_verbs.c > > =================================================================== > > --- ofa_kernel-1.3.orig/drivers/infiniband/ulp/ipoib/ipoib_verbs.c > > 2007-11-19 18:05:29.000000000 +0200 > > +++ ofa_kernel-1.3/drivers/infiniband/ulp/ipoib/ipoib_verbs.c > > 2007-11-19 18:06:52.000000000 +0200 > > @@ -181,6 +181,9 @@ > > goto out_free_mr; > > } > > > > + if (ib_modify_cq(priv->cq, 16, 10)) > > + printk(KERN_INFO "%s: failed to modify CQ params\n", ca->name); > > + > > if (ib_req_notify_cq(priv->cq, IB_CQ_NEXT_COMP)) > > goto out_free_cq; > > _______________________________________________ ewg mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
