> We have software which uses the mellanox shared receive queue and we > are seeing a substantial performance penalty (on the order of 4-5 > usec on 0 byte messages).
Are you comparing code that sends short messages either by rdma write where the reciever does polling on memory (as done in some IB MPIs "short message protocols") to code that uses send/recv with SRQ --or-- code that does send/recv where in one run the recv is to "non shared" queue (that is regular QP) and in the other run to SRQ (QP attached to SRQ)? Or. _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
