1815069 - Make it possible to tweak mozjemalloc's max dirty page sizes dynamically and increase them on foreground content processes

Status: RESOLVED FIXED

(Core :: Performance, defect, P2)

Description

[Olli Pettay [:smaug][[e-mail]]]

Some background for this is in https://bugzilla.mozilla.org/show_bug.cgi?id=1805644#c16

AWSY doesn't show too much difference, possibly because of the 3rd patch on the patch queue I'm about to upload
https://treeherder.mozilla.org/perfherder/compare?originalProject=mozilla-central&originalRevision=6424e727b6adc253c38b7d7d3c861c9360de74da&newProject=try&newRevision=5f6ae029066462d2c4e4c1ceb8d10aa77c41ebae&page=1&framework=4

BrowserTime really likes this
https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=b3dfdc2535422cf85769eef6bb762befaeb9ffb6&newProject=try&newRevision=336b8407412e68151f1afa447bea363288f4b861&page=1&framework=13&showOnlyImportant=1
especially Speedometer* part of those tests
https://treeherder.mozilla.org/perfherder/comparesubtest?originalProject=try&newProject=try&newRevision=336b8407412e68151f1afa447bea363288f4b861&originalSignature=4586009&newSignature=4586009&framework=13&originalRevision=b3dfdc2535422cf85769eef6bb762befaeb9ffb6&page=1&showOnlyImportant=1

The tweaks to the cache sizes in the 2nd patch are based on experimenting with different numbers. The non-default arenas we have really need to be larger than they currently are when running busy code, like speedometer.
It is probably worth to experiment with the limits some more, but it is a bit time consuming to get all the performance and awsy numbers.

Not sure what would be the right component for this, but since this is all about performance, Core: Performance it is :)
I could have also split this to several bugs, but the patches are small and go all together.

I need to still run this on tryserver to ensure I didn't break any checks on debug builds. I wouldn't be surprised if I missed something.

Comment 1

[Olli Pettay [:smaug][[e-mail]]]

Attachment: Bug 1815069 - Make it possible to tweak mozjemalloc's max dirty page sizes dynamically, r=glandium

Comment 2

[Olli Pettay [:smaug][[e-mail]]]

Attachment: Bug 1815069 - Increase page caches on foreground content processes, r=pbone

Depends on D168900

Comment 3

[Olli Pettay [:smaug][[e-mail]]]

Attachment: Bug 1815069 - purge page caches once the CC/GC cycle ends, r=mccr8

jemalloc_free_dirty_pages is surprisingly fast. One may see it in profiles after running speedometer, but
even then it is basically just one sample or so.

Depends on D168901

Comment 4

[Olli Pettay [:smaug][[e-mail]]]

Attachment: Bug 1815069, add dom.memory.foreground_content_processes_have_larger_page_cache pref to control page cache behavior in content processes, r=mccr8

Depends on D168902

Comment 5

[Olli Pettay [:smaug][[e-mail]]]

I'm thinking yet another way to purge. It might not need to use idle tasks, but low priority tasks and it would perhaps purge only to the default levels, not purge all. That might make this less risky from memory usage point of view. Investigating... but feel free to review the patches anyhow. That extra purging would be a followup patch or bug and would be needed only if memory usage turned out to be too high.

Comment 6

[Olli Pettay [:smaug][[e-mail]]]

I rebased the patches and run some more tests (triggered many times, but the confidence is still low) and memory usage looks quite reasonable
https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=968eb82a48a71de0a415a75d1a70f080af18074b&newProject=try&newRevision=a13b303e3131db00545d312caa00c025fc3d0a07&page=1&framework=4

Comment 7

[Olli Pettay [:smaug][[e-mail]]]

Since the API changed and I tweaked some modifiers
https://treeherder.mozilla.org/perfherder/compare?originalProject=try&originalRevision=a6bcb63372e5a145b7bc09d3a05ec569aad16b1e&newProject=try&newRevision=94ccd7acb12e00c1d0ca0f975b82475b18565abe
AWSY looks reasonble and so does Speedometer3

Comment 8

[Paul Bone [:pbone]]

I've triggered speedometer 2 and some other browsertime benchmarks on those pushes also. Some of my own work had some page load regressions and I'd like to know if yours is affected too.

Comment 9

[Olli Pettay [:smaug][[e-mail]]]

See the first comment. At least on Windows this gives massive boost on page load.

Comment 10

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5216330caa3c
Make it possible to tweak mozjemalloc's max dirty page sizes dynamically, r=glandium,pbone
https://hg.mozilla.org/integration/autoland/rev/4b1f970d7dfa
Increase page caches on foreground content processes, r=pbone
https://hg.mozilla.org/integration/autoland/rev/cd6590c2d0c1
purge page caches once the CC/GC cycle ends, r=mccr8,pbone
https://hg.mozilla.org/integration/autoland/rev/2cc01888068f
add dom.memory.foreground_content_processes_have_larger_page_cache pref to control page cache behavior in content processes, r=mccr8

Comment 11

[BugBot [:suhaib / :marco/ :calixte]]

The severity field for this bug is set to S3. However, the Performance Impact field flags this bug as having a high impact on the performance.
:smaug, could you consider increasing the severity of this performance-impacting bug? Alternatively, if you think the performance impact is lower than previously assessed, could you request a re-triage from the performance team by setting the Performance Impact flag to ??

For more information, please visit auto_nag documentation.

Comment 12

[Iulian Moraru]

https://hg.mozilla.org/mozilla-central/rev/5216330caa3c
https://hg.mozilla.org/mozilla-central/rev/4b1f970d7dfa
https://hg.mozilla.org/mozilla-central/rev/cd6590c2d0c1
https://hg.mozilla.org/mozilla-central/rev/2cc01888068f

Comment 13

[Alexandru Ionescu (needinfo me) [:alexandrui]]

(In reply to Pulsebot from comment #10)

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5216330caa3c
Make it possible to tweak mozjemalloc's max dirty page sizes dynamically,
r=glandium,pbone
https://hg.mozilla.org/integration/autoland/rev/4b1f970d7dfa
Increase page caches on foreground content processes, r=pbone
https://hg.mozilla.org/integration/autoland/rev/cd6590c2d0c1
purge page caches once the CC/GC cycle ends, r=mccr8,pbone
https://hg.mozilla.org/integration/autoland/rev/2cc01888068f
add dom.memory.foreground_content_processes_have_larger_page_cache pref to
control page cache behavior in content processes, r=mccr8

== Change summary for alert #37493 (as of Fri, 03 Mar 2023 18:23:14 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)	Performance Profiles
4%	speedometer3	linux1804-64-shippable-qr	fission webrender	101.63 -> 105.52	Before/After

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=37493

Comment 14

[Andrej (:aglavic)]

(In reply to Pulsebot from comment #10)

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5216330caa3c
Make it possible to tweak mozjemalloc's max dirty page sizes dynamically,
r=glandium,pbone
https://hg.mozilla.org/integration/autoland/rev/4b1f970d7dfa
Increase page caches on foreground content processes, r=pbone
https://hg.mozilla.org/integration/autoland/rev/cd6590c2d0c1
purge page caches once the CC/GC cycle ends, r=mccr8,pbone
https://hg.mozilla.org/integration/autoland/rev/2cc01888068f
add dom.memory.foreground_content_processes_have_larger_page_cache pref to
control page cache behavior in content processes, r=mccr8

== Change summary for alert #37572 (as of Wed, 08 Mar 2023 02:12:09 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
37%	perf_reftest_singletons link-style-cache-1.html	windows10-64-shippable-qr	e10s fission stylo webrender	452.43 -> 284.55
35%	perf_reftest_singletons link-style-cache-1.html	linux1804-64-shippable-qr	e10s fission stylo webrender	500.32 -> 324.13
33%	perf_reftest_singletons inline-style-cache-1.html	linux1804-64-shippable-qr	e10s fission stylo webrender	1,425.74 -> 955.07
31%	perf_reftest_singletons inline-style-cache-1.html	windows10-64-shippable-qr	e10s fission stylo webrender	1,425.34 -> 977.16
8%	pdfpaint	windows10-64-shippable-qr	e10s fission stylo webrender-sw	563.19 -> 515.41
...	...	...	...	...
2%	tp5o	windows10-64-shippable-qr	e10s fission stylo webrender-sw	227.03 -> 222.02

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=37572

Comment 15

[Andra Esanu (needinfo me)]

(In reply to Iulian Moraru from comment #12)

https://hg.mozilla.org/mozilla-central/rev/5216330caa3c
https://hg.mozilla.org/mozilla-central/rev/4b1f970d7dfa
https://hg.mozilla.org/mozilla-central/rev/cd6590c2d0c1
https://hg.mozilla.org/mozilla-central/rev/2cc01888068f

== Change summary for alert #37518 (as of Sat, 04 Mar 2023 14:04:49 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)	Performance Profiles
6%	outlook fcp	windows10-64-shippable-qr	fission warm webrender	98.85 -> 93.33	Before/After
5%	outlook SpeedIndex	windows10-64-shippable-qr	cold fission webrender	1,322.12 -> 1,262.00	Before/After
4%	outlook fcp (geomean)	windows10-64-shippable-qr	fission warm webrender	98.53 -> 94.37	Before/After
4%	outlook fcp (mean)	windows10-64-shippable-qr	fission warm webrender	98.71 -> 94.55	Before/After
4%	wikia ContentfulSpeedIndex	windows10-64-shippable-qr	fission warm webrender	868.95 -> 834.42
...	...	...	...	...	...
3%	nytimes LastVisualChange	windows10-64-shippable-qr	cold fission webrender	2,048.42 -> 1,989.00	Before/After

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=37518

Comment 16

[Alexandre Poirot [:ochameau]]

Good news coming from DevTools performance tests as well.

Some raw JS computations have been made faster:

• 20% improvement on server.protocoljs, means that overall DevTools Client <=> Server communication has been made significantly faster! I imagine that the test allocates tons of objects when doing this particular stress test.
• 6 to 12% improvement on webconsole.open. The console opens faster against our various test pages.
• 2% improvement on source-map. This is pure JS computation, but generating source-map could be overall faster.

== Change summary for alert #37501 (as of Sat, 04 Mar 2023 01:59:15 GMT) ==

Regressions:

Ratio	Test	Platform	Options	Absolute values (old vs new)
15%	damp console.log-in-loop-content-process-node	linux1804-64-shippable-qr	e10s fission stylo webrender-sw	47.81 -> 54.85
4%	damp console.log-in-loop-content-process-date	windows10-64-shippable-qr	e10s fission stylo webrender-sw	66.34 -> 68.98

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)
14%	damp server.protocoljs.DAMP	windows10-64-shippable-qr	e10s fission stylo webrender-sw	1,579.23 -> 1,358.87
14%	damp server.protocoljs.DAMP	windows10-64-shippable-qr	e10s fission stylo webrender	1,581.24 -> 1,364.75
12%	damp server.protocoljs.DAMP	linux1804-64-shippable-qr	e10s fission stylo webrender	1,180.21 -> 1,034.21
12%	damp custom.webconsole.open.DAMP	windows10-64-shippable-qr	e10s fission stylo webrender-sw	654.64 -> 577.45
12%	damp custom.webconsole.open.DAMP	windows10-64-shippable-qr	e10s fission stylo webrender	653.73 -> 577.32
...	...	...	...	...
2%	damp console.log-in-loop-content-process-window	windows10-64-shippable-qr	e10s fission stylo webrender-sw	213.56 -> 208.48

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=37501

Comment 17

[Alex Finder]

== Change summary for alert #37456 (as of Thu, 02 Mar 2023 18:41:12 GMT) ==

Improvements:

Ratio	Test	Platform	Options	Absolute values (old vs new)	Performance Profiles
5%	reddit fcp	linux1804-64-shippable-qr	fission warm webrender	171.62 -> 162.88	Before/After
5%	cnn SpeedIndex	android-hw-a51-11-0-aarch64-shippable-qr	cold webrender	3,397.38 -> 3,228.50
5%	cnn FirstVisualChange	android-hw-a51-11-0-aarch64-shippable-qr	warm webrender	2,390.08 -> 2,278.08
5%	cnn SpeedIndex	android-hw-a51-11-0-aarch64-shippable-qr	warm webrender	2,493.58 -> 2,377.17
5%	cnn PerceptualSpeedIndex	android-hw-a51-11-0-aarch64-shippable-qr	warm webrender	2,466.00 -> 2,353.33
...	...	...	...	...	...
4%	cnn FirstVisualChange	android-hw-a51-11-0-aarch64-shippable-qr	cold webrender	2,972.18 -> 2,847.67

For up to date results, see: https://treeherder.mozilla.org/perfherder/alerts?id=37456