On Wed, Jun 15, 2011 at 12:40:32PM +0200, Peter Hallin wrote: > On 2011-06-15 04:26, Jonathan Gray wrote: > > Claudio has recently synced ix to a newer version of the Intel code, > > if you can try again with -current this should work. > > Great, thanks. I tried it out today, but the performance wasn't exactly > what I had expected. > > This is a dual port card, so I bridged ix0 with ix1 and ran iperf > between two hosts through the card. The hosts were connected with > 1Gbase as I have no other 10Gbase cards. > > It only got up to about 450 Mbit/s (standard TCP iperf test) and there > was an unbalance between ix0 and ix1 when it came to interrupts. ix1 > only got up to 4000 intr/s whereas ix1 got up to 13000 intr/s (when > monitoring live with systat vmstat). The CPU intr usage was around 50-60%. > > Could this possibly be related to the ppb problems we've seen with em(4) > cards?
Yes this seems low indeed. You could try use rdomains and route exec with a cable between the two ports to make packets go over the interfaces if you don't have another host with 10G interfaces. If you've had ppb problems with em I'd imagine they would affect ix as well if it is sitting behind the same ppb.
