On 12/18/2019 3:02 AM, Jens Axboe wrote: > On 12/17/19 3:28 PM, Pavel Begunkov wrote: >> percpu_ref_tryget() has its own overhead. Instead getting a reference >> for each request, grab a bunch once per io_submit_sqes(). >> >> basic benchmark with submit and wait 128 non-linked nops showed ~5% >> performance gain. (7044 KIOPS vs 7423) > > Confirmed about 5% here as well, doing polled IO to a fast device. > That's a huge gain! > Great that you got it for a real drive! As mentioned, after we are done with this one, I have an idea how to optimise it even further, which would work for small QD as well. -- Pavel Begunkov
