On Wed, 2017-12-13 at 19:02 -0500, Jeff Layton wrote:
> On Thu, 2017-12-14 at 10:03 +1100, Dave Chinner wrote:
> > On Wed, Dec 13, 2017 at 03:14:28PM -0500, Jeff Layton wrote:
> > > On Wed, 2017-12-13 at 10:05 -0500, J. Bruce Fields wrote:
> > > > This is great, thanks.
> > > >
> > > > On Wed, Dec 13, 2017 at 09:19:58AM -0500, Jeff Layton wrote:
> > > > > With this, we reduce inode metadata updates across all 3 filesystems
> > > > > down to roughly the frequency of the timestamp granularity, particularly
> > > > > when it's not being queried (the vastly common case).
> > > > >
> > > > > The pessimal workload here is 1 byte writes, and it helps that
> > > > > significantly. Of course, that's not what we'd consider a real-world
> > > > > workload.
> > > > >
> > > > > A tiobench-example.fio workload also shows some modest performance
> > > > > gains, and I've gotten mails from the kernel test robot that show some
> > > > > significant performance gains on some microbenchmarks (case-msync-mt in
> > > > > the vm-scalability testsuite to be specific), with an earlier version of
> > > > > this set.
> > > > >
> > > > > With larger writes, the gains with this patchset mostly vaporize,
> > > > > but it does not seem to cause performance to regress anywhere, AFAICT.
> > > > >
> > > > > I'm happy to run other workloads if anyone can suggest them.
> > > > >
> > > > > At this point, the patchset works and does what it's expected to do in
> > > > > my own testing. It seems like it's at least a modest performance win
> > > > > across all 3 major disk-based filesystems. It may also encourage others
> > > > > to implement i_version as well since it reduces the cost.
> > > >
> > > > Do you have an idea what the remaining cost is?
> > > >
> > > > Especially in the ext4 case, are you still able to measure any
> > > > difference in performance between the cases where i_version is turned on
> > > > and off, after these patches?
> > >
> > > Attached is a fio jobfile + the output from 3 different runs using it
> > > with ext4. This one is using 4k writes. There was no querying of
> > > i_version during the runs. I've done several runs with each and these
> > > are pretty representative of the results:
> > >
> > > old = 4.15-rc3, i_version enabled
> > > ivers = 4.15-rc3 + these patches, i_version enabled
> > > noivers = 4.15-rc3 + these patches, i_version disabled
> > >
> > > To snip out the run status lines:
> > >
> > > old:
> > > WRITE: bw=85.6MiB/s (89.8MB/s), 9994KiB/s-11.1MiB/s (10.2MB/s-11.7MB/s), io=50.2GiB (53.8GB), run=600001-600001msec
> > >
> > > ivers:
> > > WRITE: bw=110MiB/s (115MB/s), 13.5MiB/s-14.2MiB/s (14.1MB/s-14.9MB/s), io=64.3GiB (69.0GB), run=600001-600001msec
> > >
> > > noivers:
> > > WRITE: bw=117MiB/s (123MB/s), 14.2MiB/s-15.2MiB/s (14.9MB/s-15.9MB/s), io=68.7GiB (73.8GB), run=600001-600001msec
> > >
> > > So, I see some performance degradation with -o iversion compared to not
> > > having it enabled (maybe due to the extra atomic fetches?), but this set
> > > erases most of the difference.
> >
> > So what is the performance difference when something is actively
> > querying the i_version counter as fast as it can (e.g. file being
> > constantly stat()d via NFS whilst being modified)? How does the
> > performance compare to the old code in that case?
> >
>
> I haven't benchmarked that with the latest set, but I did with the set
> that I posted around a year ago. Basically I just ran a similar test to
> this, and had another shell open doing statx(..., STATX_VERSION, ...);
> the thing in a tight loop.
>
> I did see some performance hit vs. the case where no one is viewing it,
> but it was still significantly faster than the unpatched version that
> was incrementing the counter every time.
>
> That was on a different test rig, and the patchset has some small
> differences now. I'll see if I can get some hard numbers with such a
> testcase soon.
I reran the test on xfs, with another shell running test-statx to query
the i_version value in a tight loop:
WRITE: bw=110MiB/s (115MB/s), 13.6MiB/s-14.0MiB/s (14.2MB/s-14.7MB/s), io=64.5GiB (69.3GB), run=600001-600001msec
...contrast that with the run where I was not doing any queries:
WRITE: bw=129MiB/s (136MB/s), 15.8MiB/s-16.6MiB/s (16.6MB/s-17.4MB/s), io=75.9GiB (81.5GB), run=600001-600001msec
...vs the unpatched kernel:
WRITE: bw=86.7MiB/s (90.0MB/s), 9689KiB/s-11.7MiB/s (9921kB/s-12.2MB/s), io=50.8GiB (54.6GB), run=600001-600002msec
There is some clear peformance impact when you are running frequent
queries of the i_version.
My gut feeling is that you could probably make the new code perform
worse than the old if you were to _really_ hammer the inode with queries
for the i_version (probably from many threads in parallel) while doing a
lot of small writes to it.
That'd be a pretty unusual workload though.
--
Jeff Layton <jlayton@xxxxxxxxxx>
