On Sun, Jun 11, 2023 at 03:27:57PM +0300, Sagi Grimberg wrote: > Maybe I'm doing something wrong? Didn't expect to find a regression > against mainline on the default setup.
Hi Sagi, I ran the latest branch against ed8ad9728a where it forked off master. master achieves fewer IOPS. It looks like the regression you saw was solved by the changes I made last week. Both "master" and "modified" are running with 1 IOThread: IOPS ------ randread 4k 64 master-1 213504 randread 4k 64 master-2 212650 randread 4k 64 master-3 211699 randread 4k 64 master-4 211940 randread 4k 64 master-5 214110 randread 4k 64 modified-1 234708 randread 4k 64 modified-2 236014 randread 4k 64 modified-3 235328 randread 4k 64 modified-4 235742 The improvement is around 10%. You can find the benchmark configuration and raw data here: https://gitlab.com/stefanha/virt-playbooks/-/tree/1a464c0676fe9133fb244d8a2dd1439001c7bc42 The configuration is in go.yml, plays/benchmark.yml, files/test.xml.j2, and files/fio.sh. The raw data is in notebook/fio-output/ and you can explore the Jupyter notebook by running notebook/go.sh. Stefan
signature.asc
Description: PGP signature