Re: [RFC PATCH v1 00/30] fs: inode->i_version rework and optimization

[bfields@fieldses.org (J. Bruce Fields)]

On Thu, Dec 22, 2016 at 09:42:04AM -0500, Jeff Layton wrote:
> On Thu, 2016-12-22 at 00:45 -0800, Christoph Hellwig wrote:
> > On Wed, Dec 21, 2016 at 12:03:17PM -0500, Jeff Layton wrote:
> > > 
> > > Only btrfs, ext4, and xfs implement it for data changes. Because of
> > > this, these filesystems must log the inode to disk whenever the
> > > i_version counter changes. That has a non-zero performance impact,
> > > especially on write-heavy workloads, because we end up dirtying the
> > > inode metadata on every write, not just when the times change. [1]
> > 
> > Do you have numbers to justify these changes?
> 
> I have numbers. As to whether they justify the changes, I'm not sure.
> This helps a lot on a (admittedly nonsensical) 1-byte write workload. On
> XFS, with this fio jobfile:

To me, the interesting question is whether this allows us to turn on
i_version updates by default on xfs and ext4.

When Josef looked at doing that previously he withdrew the patch due to
performance regressions.  I think the most useful thread started here:

	http://lkml.kernel.org/r/1337092396-3272-1-git-send-email-josef@redhat.com

Skimming quickly....  I think the regression was also in the small-write
case.  So apparently that was thought to reveal a real problem?

So if you've mostly eliminated that regression, then that's good
motivation for your patches.  (Though I think in addition to comparing
the patched and unpatched i_version case, we need to compare to the
unpatched not-i_version case.  I'm not clear whether you did that.)

--b.

> 
> --------------------8<------------------
> [global]
> direct=0
> size=2g
> filesize=512m
> bsrange=1-1
> timeout=60
> numjobs=1
> directory=/mnt/scratch
> 
> [f1]
> filename=randwrite
> rw=randwrite
> --------------------8<------------------
> 
> Unpatched kernel:
>   WRITE: io=7707KB, aggrb=128KB/s, minb=128KB/s, maxb=128KB/s, mint=60000msec, maxt=60000msec
> 
> Patched kernel:
>   WRITE: io=12701KB, aggrb=211KB/s, minb=211KB/s, maxb=211KB/s, mint=60000msec, maxt=60000msec
> 
> So quite a difference there and it's pretty consistent across runs. If I
> change the jobfile to have "direct=1" and "bsrange=4k-4k", then any
> variation between the two doesn't seem to be significant (numbers vary
> as much between runs on the same kernels and are roughly the same).
> 
> Playing with buffered I/O sizes between 1 byte and 4k shows that as the
> I/O sizes get larger, this makes less difference (which is what I'd
> expect).
> 
> Previous testing with ext4 shows roughly the same results. btrfs shows
> some benefit here but significantly less than with ext4 or xfs. Not sure
> why that is yet -- maybe CoW effects?
> 
> That said, I don't have a great test rig for this. I'm using VMs with a
> dedicated LVM volume that's on a random SSD I had laying around. It
> could use testing on a wider set of configurations and workloads.
> 
> I was also hoping that others may have workloads that they think might
> be (postively or negatively) affected by these changes. If you can think
> of any in particular, then I'm interested to hear about them.
> 
> -- 
> Jeff Layton <jlayton@redhat.com>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html