Bug 247690 - sysutils/openzfs-kmod performance problem
Summary: sysutils/openzfs-kmod performance problem
Status: New
Alias: None
Product: Ports & Packages
Classification: Unclassified
Component: Individual Port(s) (show other bugs)
Version: Latest
Hardware: amd64 Any
: --- Affects Some People
Assignee: Ryan Moeller
Depends on:
Reported: 2020-07-01 10:03 UTC by rob2g2
Modified: 2020-07-01 21:00 UTC (History)
3 users (show)

See Also:
bugzilla: maintainer-feedback? (freqlabs)

logs + sample test script (33.20 KB, application/x-bzip)
2020-07-01 10:03 UTC, rob2g2
no flags Details
tests with blocksize128k (18.21 KB, application/x-bzip)
2020-07-01 21:00 UTC, rob2g2
no flags Details

Note You need to log in before you can comment on or make changes to this bug.
Description rob2g2 2020-07-01 10:03:31 UTC
Created attachment 216107 [details]
logs + sample test script

I started lots of benchmarks concerning ZFS. Among the test candidates was a Notebook with a single SSD disk where I compared the performance of FreeBSDs ZFS shipped with the base system versus openzfs-kmod. I realized that openzfs-kmod is much slower compared to FreeBSDs base zfs. I used the tool benchmarks/fio (3.20).

If I pick one test, of course on the same hardware, the random-read-write test fired up with the "sync"-ioengine and ashift=12, by the command "fio --name=randrw --rw=randrw --direct=1 --ioengine=sync --bs=8k --numjobs=2 --rwmixread=80 --size=1G --runtime=600 --group_reporting" native-zfs delivers 81.7MiB/s read and 20.5MiB/s write bandwidth whereas openzfs-kmod only reaches 9665KiB/s read and 2422KiB/s write bandwidth.

The difference is quite big on all my tests, with combinations of io-engines posixaio, psync, mmap, sync, pvsync, vsync; zfs with ashift=9 and ashift=12.
This was not really a scientific test, but still showed the obvious difference in performance. fio seems to be a linux tool, however, works on FreeBSD - I don't know how the different io-engines are implemented, but it shows the difference in all tests.

Tests on normal hdds are to come too, but they take of course significantly longer. I am also testing geli/zfs-encryption on FreeBSD and Linux (luks encryption). This is out of scope here, just for info that I can provide logs of different zfs combinations (2-disk raidz1/zmirror, 3-disk raidz1) at a later date if desired.

Attached are 24 log files which include the results of some fio tests (and a dd command output/timing). Also, see the included unenc-openzfs.sh for how/what tests were done.
Comment 1 rob2g2 2020-07-01 10:45:28 UTC
openzfs-kmod-2020060300 was used
Comment 2 Allan Jude freebsd_committer 2020-07-01 17:45:40 UTC
(In reply to rob2g2 from comment #1)
To get a better test, the first thing you will want to do is make sure the 'recordsize' of the dataset you are pointing fio at, is the same as the --bs argument you give fio.

I'll try to look into the logs you attached later.
Comment 3 rob2g2 2020-07-01 21:00:33 UTC
Created attachment 216117 [details]
tests with blocksize128k

did the tests again this time bs=128k as you suggested. The difference is not as dramatic as with the tests before (though the blocksize was the same for openzfs and native zfs)