Dotan Barak wrote:
If one wishes to use maximum inline data, he should open a QP with
maximum number of s/g entries in the QP->SQ
(ibv_qp_init_attr.cap.max_send_sge). The actual number of bytes which
can be send as inline will be returned in the
attribute ibv_qp_init_attr.cap.max_inline_data after ibv_create_qp
will be finished.
Its correct that the maximal inline length can be get this way, but
please note that the inline size as a function of the send s/g entries
has "steps" so for some values of n=|s/g| f(n+1) = f(n) but n+1 would
consume more memory.
Basically, inline sending tends to improve latency (only one DMA
transaction, to fetch the descriptor) but it increases the memory
consumption and the CPU utilization (since the user space hw library
does a copy). My take here is that it should be used with care, going
simple you can create the QP with the s/g count you need per your design
and do inline send only when possible under this count, etc.
Or.
_______________________________________________
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
