1848090 - Opening a 300MB Chrome profile in Firefox profiler spends a lot of time around GCMinor

Status: RESOLVED FIXED

(Core :: JavaScript Engine, task, P3)

Description

[Mayank Bansal]

Open a 300MB Google Chrome profile in Firefox profiler

Profile of opening the Chrome profile: https://share.firefox.dev/3Oy7p7Y

This may be the expected behaviour or good enough behaviour. Feel free to resolve as invalid.

Comment 1

[Mayank Bansal]

Attachment: Trace-20230810T115935.part01.rar

chrome trace part1

Comment 2

[Mayank Bansal]

Attachment: Trace-20230810T115935.part02.rar

Chrome trace part2

Comment 3

[Mayank Bansal]

Attachment: about:support

Comment 4

[Jon Coppeard (:jonco)]

The JSON parser parser is creating a lot of nursery objects which all end up tenured. We could probably do the same thing we just did for structured clone deserialisation and stop allocating in the nursery after the first nursery collection.

Comment 5

[Jon Coppeard (:jonco)]

Attachment: Bug 1848090 - When parsing JSON, Stop allocating in the nursery after two nursery collections r?sfink

The patch adds a nursery collection callback for the duration of parsing and
counts the collections. After the second collection we switch to allocating
everything in the tenured heap.

JSON parsing allocates from the leaves of the tree upwards, unlike structured
clone deserialization which works from the root downwards. Because of this it
doesn't always make sense to stop nursery allocation after the first
collection, as this doesn't doom the whole data structure to being tenured. We
don't know ahead of time how big the resulting data structure will be but after
two nursery collections then at least half of it will end up tenured.

This change made parsing the JSON data in the bug significantly faster.

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/fc9820aab42f
When parsing JSON, Stop allocating in the nursery after two nursery collections r=sfink

Comment 7

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/fc9820aab42f

Comment 8

[Mayank Bansal]

Profile: https://share.firefox.dev/45fgTvY
So 12.6seconds -> 5.2 seconds
Time spent in GCMinor is all but gone (except in a few places). This is a nice improvement, thanks!