Linux-mm Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: SJ Park <sj@kernel.org>
To: SJ Park <sj@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Jonathan Corbet <corbet@lwn.net>,
	Randy Dunlap <rdunlap@infradead.org>,
	Shuah Khan <skhan@linuxfoundation.org>,
	damon@lists.linux.dev, linux-doc@vger.kernel.org,
	linux-mm@kvack.org, arijit.desaii@crusoe.ai
Subject: Re: [RFC PATCH 0/4] mm/damon/reclaim: enable auto-tuning features by default
Date: Fri, 20 Nov 2026 17:53:23 -0800	[thread overview]
Message-ID: <20261121015324.83753-1-sj@kernel.org> (raw)
In-Reply-To: <20261120060741.92121-1-sj@kernel.org>

On Thu, 19 Nov 2026 22:07:34 -0800 SJ Park <sj@kernel.org> wrote:

> The initial version of DAMON_RECLAIM was supposed to be used after
> manual tuning of its parameters.  Among the parameters, the monitoring
> intervals (sample_interval and aggr_interval) and the coldness threshold
> (min_age) are particularly important to be optimally tuned.  The optimum
> values for the parameters tend to be highly dependent on the workload
> and the system.  Hence no small amount of tuning efforts are required to
> each DAMON_RECLAIM use case.
> 
> To reduce the tuning efforts, we recently introduced two auto-tuning
> features.  Intervals auto-tuning (autotune_monitoring_intervals)
> automatically tunes monitoring intervals to provide practical and
> efficient monitoring results.  PSI-driven quota auto-tuning
> (quota_autotune_feedback) automatically tunes the coldness threshold.
> 
> Because the auto-tuning features were introduced after the initial
> version, the features are disabled by default.  That is to keep the
> default behaviors unchanged.  However, enabling the auto-tuning features
> is only expected to provide more benefit to the users.  Disabling those
> by default only makes the adoption of the features slow.  Also, it is
> difficult to imagine real users who use DAMON_RECLAIM with its default
> parameter values.
> 
> Enable the auto-tuning features by default.

My colleague, Arijit, found this is not a good idea.  He is now working on
proving the theory and making a patch.

DAMON_RECLAIM receives the coldness throughput via 'min_age' parameter, which
has 2 minutes as the default value.  DAMON_RECLAIM converts that into the
number of aggregation intervals for the time, and set it as the min_age_region
of the damos_access_pattern for its reclamation DAMOS scheme.  DAMON_RECLAIM
does the conversion when it starts working, or a parameter is updated online,
using commit_inputs parameter.  It doesn't update the min_age_region value when
the aggregation interval is auto-tuned.

This patchset enables the intervals auto-tuning features so that the
aggregation intervals span from 100 milliseconds to 200 seconds (up to 2,000
x).  As a result, the effective min_age can span from 2 minutes to 4,000
minutes.

In usual real workload, the auto-tuning converges the aggregation interval to a
few seconds (20-200x).  The effective min_age will be about 40 minutes to 400
minutes.  Even on long term real world workload, that could be too much
threshold.  The user will show DAMON_RECLAIM is reclaiming only unexpectedly
small amount of memory.

PSI-driven coldness threshold auto-tuning can make the effective min_age only
longer, so it doesn't help.

One strightforward workaround would be setting the min_age a small value, like
2-4 aggregation intervals (200-400 milliseconds).  Then, in usual case, the
effective min_age becomes 4 seconds to 40 seconds.  It might feel like too
small threshold.  PSI-driven coldness threshold auto-tuning will make it
better.

Yet another workaround would be monitoring the auto-tuned interval and update
the min_age online, using the commit_inputs parameter.

We introduced age_ms DAMOS filter type [1] to fix this kind of issue.  Long
term solution for DAMON_RECLAIM would be modifying it to use the age_ms DAMOS
filter instead of the aggregation interval based min_age_region.

My colleague, Arijit, is working on the long term solution.  Until it is
arrived, I will keep the RFC tag of this patch series.

[1] https://lore.kernel.org/20261106020559.98352-1-sj@kernel.org


Thanks,
SJ

[...]


      parent reply	other threads:[~2026-11-21  1:53 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-20  6:07 [RFC PATCH 0/4] mm/damon/reclaim: enable auto-tuning features by default SJ Park
2026-11-20  6:07 ` [RFC PATCH 1/4] mm/damon/reclaim: enable monitoring intervals auto-tune " SJ Park
2026-11-20  6:07 ` [RFC PATCH 2/4] mm/damon/reclaim: set memory PSI 0.05% as default quota goal SJ Park
2026-11-20  6:07 ` [RFC PATCH 3/4] mm/damon/reclaim: disable watermarks by default SJ Park
2026-11-20  6:07 ` [RFC PATCH 4/4] Docs/admin-guide/mm/damon/reclaim: update for changed default parameters SJ Park
2026-11-21  1:53 ` SJ Park [this message]

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=20261121015324.83753-1-sj@kernel.org \
    --to=sj@kernel.org \
    --cc=akpm@linux-foundation.org \
    --cc=arijit.desaii@crusoe.ai \
    --cc=corbet@lwn.net \
    --cc=damon@lists.linux.dev \
    --cc=linux-doc@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=rdunlap@infradead.org \
    --cc=skhan@linuxfoundation.org \
    /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
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox