[openib-general] uverbs performance; ibv_pingpong poll vs sleep

[Roland Dreier]

    Steven> I have compaired the data rates using ibv_pingpong with
    Steven> and without the -e option. Therefore, using polling and
    Steven> waiting the CQ events (sleeping).

    Steven> Is there any way to trade off the data rate with the CPU
    Steven> usage (I was thinking of some timeout from polling).

I suppose you could have some sort of adaptive polling scheme that
spins polling for a while and then sleeps waiting for an event.
However, as Michael said, it's probably better to use pipelining and
post multiple send work requests.  This hides the latency of getting a
completion event by keeping the HCA busy.

 - R.