Roland Dreier wrote: >> This is only for completeness sake. Performance will be poor compared to two >> (or more) >> Linux systems using this. This is expected to be really a corner and do not >> expect this >> to happen. Hence is there for completeness only. If I interpreted this >> correctly, this >> is was what Roland suggested (and I agree with the suggestion). > > No, this isn't really what I suggested. I suggested one receive to > start with, and if this receive is consumed, then post a full set of > receives to fill the QP receive queue and use it normally. No > separate CQ, no kernel log message. This still might run into RNR > issues if more than one message is sent right away. > > I woudn't be that surprised if AIX or Solaris or some other OS chooses > to use the same QP for sending and receiving, and I'm not sure we want > to be in a situation where that doesn't work well. Although maybe we > can defer worrying about this until someone runs into the issue in > practice. > > I guess you should post your latest rolled up patch sowe can see where > we stand on this. >
Without a separate CQ, how do the receive work completion handlers distinguish between the different receive queues of rx_qp and tx_qp -any suggestions? Pradeep _______________________________________________ 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
