On Tue, 2021-11-30 at 06:26 -0800, Bart Van Assche wrote: > On 11/30/21 12:43 AM, Bean Huo wrote: > > On Fri, 2021-11-19 at 11:57 -0800, Bart Van Assche wrote: > > > On my test setup this patch increases IOPS from 2736 to 22000 > > > (8x) > > > for the > > > following test: > > > > > > for hipri in 0 1; do > > > fio --ioengine=io_uring --iodepth=1 --rw=randread \ > > > --runtime=60 --time_based=1 --direct=1 --name=qd1 \ > > > --filename=/dev/block/sda --ioscheduler=none --gtod_reduce=1 > > > \ > > > --norandommap --hipri=$hipri > > > done > > > > For 4KB random read, direct IO, and iodepth=1, we did not see an > > improvement in IOPS due to this patch. Maybe the test case above is > > not > > sufficient. > > Hi Bean, > > Which test has been run? Polling is only enabled if --hipri=1 is > specified > to fio and --ioengine=io_uring is the only I/O engine on Android that > supports > that flag. > > Thanks, > > Bart. Hi Bart, It is the test case in your commit message. If iodepth=1, there is no performance improvement. Increase the io-depth to multiple, for example, let iodepth= CPU core counter. I see that the interrupt overhead is significantly increased when the request is completed, so IO_polling will win compared to the interrupt mode. Kind regards, Bean
