On Fri, 2008-02-01 at 01:45 -0800, Shirley Ma wrote: > I filed a patch back two years ago to split CQ. The feedback was it > didn't benefit mthca since it had only one interrupt shared between send > and recv. More context switches were generated when splitting CQ. Then I > decided to wait the multiple interrupt vector to be implemented in > device driver layer to push this patch later with multiple interrupt > vectors (This is the next item on my list). The why the performance got > improved without multiple interrupt support here? >
In my approach, I use two CQs but the send does not generate interrupts. Instead I use polling right after I post to the send queue. Splitting the CQ is also a preparation for using unsignaled send queue, this significantly reducing the overhead of polling the CQ. As I mentioned, this approach significantly improves small UDP messages send rate. _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
