I used dd to test
            dd if=/tmp/mount/home/ubuntu/os_disk_partition/genesisTool/4Fz
of=/dev/null bs=1M count=53687091200 skip=0
iflag=skip_bytes,count_bytes,direct &

            dd if=/tmp/mount/home/ubuntu/os_disk_partition/genesisTool/4Fz
of=/dev/null  bs=1M count=53687091200 skip=53687091200
iflag=skip_bytes,count_bytes,direct &

            dd if=/tmp/mount/home/ubuntu/os_disk_partition/genesisTool/4Fz
of=/dev/null  bs=1M count=53687091200 skip=107374182400
iflag=skip_bytes,count_bytes,direct &

            dd if=/tmp/mount/home/ubuntu/os_disk_partition/genesisTool/4Fz
of=/dev/null  bs=1M count=53687091200 skip=161061273600
iflag=skip_bytes,count_bytes,direct &

Here the file /tmp/mount/home/ubuntu/os_disk_partition/genesisTool/4Fz
is a file in the ext4 file system that is accessed via a loop device.

Also in the above change we have the default version to be always 1
worker. One can change the number of workers to handle the performance
to their needs. In our case we saw that we got good performance with 4
threads even for sequential io.

On Wed, Jan 22, 2020 at 5:40 PM Ming Lei [off-list ref] wrote:

On Wed, Jan 22, 2020 at 4:11 AM muraliraja.muniraju
[off-list ref] wrote:

quoted

Below is the dd results that I ran with the worker and without the worker changes.
Enhanced Loop has the changes and ran with 1,2,3,4 workers with 4 dds running on the same loop device.
Normal Loop is 1 worker(the existing code) with 4 dd's running on the same loop device.
Enhanced loop
1 - READ: io=21981MB, aggrb=187558KB/s, minb=187558KB/s, maxb=187558KB/s, mint=120008msec, maxt=120008msec
2 - READ: io=41109MB, aggrb=350785KB/s, minb=350785KB/s, maxb=350785KB/s, mint=120004msec, maxt=120004msec
3 - READ: io=45927MB, aggrb=391802KB/s, minb=391802KB/s, maxb=391802KB/s, mint=120033msec, maxt=120033msec
4 - READ: io=45771MB, aggrb=390543KB/s, minb=390543KB/s, maxb=390543KB/s, mint=120011msec, maxt=120011msec
Normal loop
1 - READ: io=18432MB, aggrb=157201KB/s, minb=157201KB/s, maxb=157201KB/s, mint=120065msec, maxt=120065msec
2 - READ: io=18762MB, aggrb=160035KB/s, minb=160035KB/s, maxb=160035KB/s, mint=120050msec, maxt=120050msec
3 - READ: io=18174MB, aggrb=155058KB/s, minb=155058KB/s, maxb=155058KB/s, mint=120020msec, maxt=120020msec
4 - READ: io=20559MB, aggrb=175407KB/s, minb=175407KB/s, maxb=175407KB/s, mint=120020msec, maxt=120020msec

Could you share your exact test command?

Multiple jobs may hurt performance in case of sequential IOs on HDD backend.
Also the 1st version of the loop dio patch uses normal wq, I remembered that
random IOperformance isn't improved much, meantime sequential IO perf drops
with normal wq, whentesting SSD backend.

So I took kthread worker.

Thanks,
Ming Lei