> I'm not sure if the SRP stack (initiator side) polls or interrupts, > but I seem to be hitting a wall at ~120K transactions/second.
Interesting. The SRP initiator is using interrupts to call srp_completion(), which then services all pending completions and rearms the CQ. So it is essentially interrupt driven -- there is no polling mechanism a la NAPI for network devices. > Is there any way to coalesce interrupts w/ SRP? Not at the moment, although one could do a hack with ib_modify_cq() a la IPoIB (assuming you're using ConnectX HCA). A bigger piece of work would be some sort of NAPI framework -- I think people have alreay looked at "NAPI for blockdevices" as part of work to optimize for high-IOPS SSDs, but I don't have pointers at the moment. - R. _______________________________________________ general mailing list general@lists.openfabrics.org http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
