Linux-Fsdevel Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Amir Goldstein <amir73il@gmail.com>
To: Mel Gorman <mgorman@techsingularity.net>
Cc: Jan Kara <jack@suse.cz>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	linux-kernel <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] fsnotify: Rearrange fast path to minimise overhead when there is no watcher
Date: Mon, 8 Jun 2020 21:12:44 +0300	[thread overview]
Message-ID: <CAOQ4uxgcUHuqiXFPO5mX=rvDwP-DOoTZrXvpVNphwEMFYHtyCw@mail.gmail.com> (raw)
In-Reply-To: <20200608180130.GJ3127@techsingularity.net>

> > > didn't look too closely at your series as I'm not familiar with fsnotify
> > > in general. However, at a glance it looks like fsnotify_parent() executes
> > > a substantial amount of code even if there are no watchers but I could
> > > be wrong.
> > >
> >
> > I don't about substantial, I would say it is on par with the amount of
> > code that you tries to optimize out of fsnotify().
> >
> > Before bailing out with DCACHE_FSNOTIFY_PARENT_WATCHED
> > test, it also references d_inode->i_sb,  real_mount(path->mnt)
> > and fetches all their ->x_fsnotify_mask fields.
> >
> > I changed the call pattern from open/modify/... hooks from:
> > fsnotify_parent(...);
> > fsnotify(...);
> >
> > to:
> > fsnotify_parent(...); /* which calls fsnotify() */
> >
> > So the NULL marks optimization could be done in beginning of
> > fsnotify_parent() and it will be just as effective as it is in fsnotify().
> >
>
> Something like that may be required because
>
>                               5.7.0                  5.7.0                  5.7.0                  5.7.0
>                             vanilla      fastfsnotify-v1r1      fastfsnotify-v2r1          amir-20200608
> Amean     1       0.4837 (   0.00%)      0.4630 *   4.27%*      0.4597 *   4.96%*      0.4967 *  -2.69%*
> Amean     3       1.5447 (   0.00%)      1.4557 (   5.76%)      1.5310 (   0.88%)      1.6587 *  -7.38%*
> Amean     5       2.6037 (   0.00%)      2.4363 (   6.43%)      2.4237 (   6.91%)      2.6400 (  -1.40%)
> Amean     7       3.5987 (   0.00%)      3.4757 (   3.42%)      3.6543 (  -1.55%)      3.9040 *  -8.48%*
> Amean     12      5.8267 (   0.00%)      5.6983 (   2.20%)      5.5903 (   4.06%)      6.2593 (  -7.43%)
> Amean     18      8.4400 (   0.00%)      8.1327 (   3.64%)      7.7150 *   8.59%*      8.9940 (  -6.56%)
> Amean     24     11.0187 (   0.00%)     10.0290 *   8.98%*      9.8977 *  10.17%*     11.7247 *  -6.41%*
> Amean     30     13.1013 (   0.00%)     12.8510 (   1.91%)     12.2087 *   6.81%*     14.0290 *  -7.08%*
> Amean     32     13.9190 (   0.00%)     13.2410 (   4.87%)     13.2900 (   4.52%)     14.7140 *  -5.71%*
>
> vanilla and fastnotify-v1r1 are the same. fastfsnotify-v2r1 is just the
> fsnotify_parent() change which is mostly worse and may indicate that the
> first patch was reasonable. amir-20200608 is your branch as of today and
> it appears to introduce a substantial regression albeit in an extreme case
> where fsnotify overhead is visible. The regressions are mostly larger
> than noise with the caveat it may be machine specific given that the
> machine is overloaded. I accept that adding extra functional to fsnotify
> may be desirable but ideally it would not hurt the case where there are
> no watchers at all.
>

Of course.
And thanks for catching this regression even before I posted the patches :-)

> So what's the right way forward? The patch as-is even though the fsnotify()
> change itself may be marginal, a patch that just inlines the fast path
> of fsnotify_parent or wait for the additional functionality and try and
> address the overhead on top?
>
>

Let me add your optimizations on top of my branch with the needed
adaptations and send you a branch for testing.

Thanks,
Amir.

  reply	other threads:[~2020-06-08 18:12 UTC|newest]

Thread overview: 14+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-06-08 14:05 Mel Gorman
2020-06-08 15:03 ` Amir Goldstein
2020-06-08 16:06   ` Mel Gorman
2020-06-08 16:26     ` Amir Goldstein
2020-06-08 18:01       ` Mel Gorman
2020-06-08 18:12         ` Amir Goldstein [this message]
2020-06-08 20:39           ` Amir Goldstein
2020-06-10 12:59             ` Mel Gorman
2020-06-10 13:24               ` Amir Goldstein
2020-06-12  9:38                 ` Amir Goldstein
2020-06-12 12:42                   ` Mel Gorman
2020-06-08 15:19 ` Jan Kara
2020-06-08 16:50   ` Mel Gorman
2020-06-08 17:45     ` Amir Goldstein

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CAOQ4uxgcUHuqiXFPO5mX=rvDwP-DOoTZrXvpVNphwEMFYHtyCw@mail.gmail.com' \
    --to=amir73il@gmail.com \
    --cc=jack@suse.cz \
    --cc=linux-fsdevel@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mgorman@techsingularity.net \
    --subject='Re: [PATCH] fsnotify: Rearrange fast path to minimise overhead when there is no watcher' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).