1561888 - 6.03% Heap Unclassified (linux64-shippable-qr) regression on push 431c5142ff3ad102328d90cb29470b20ccd1271e (Thu June 13 2019)

Status: RESOLVED WONTFIX

(Core :: Graphics: WebRender, defect, P1)

Description

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

We have detected an awsy regression from push:

https://hg.mozilla.org/integration/autoland/pushloghtml?changeset=431c5142ff3ad102328d90cb29470b20ccd1271e

As author of one of the patches included in that push, we need your help to address this regression.

Regressions:

6% Heap Unclassified linux64-shippable-qr opt stylo tp6 169,863,343.96 -> 180,102,028.55

You can find links to graphs and comparison views for each of the above tests at: https://treeherder.mozilla.org/perf.html#/alerts?id=21466

On the page above you can see an alert for each affected platform as well as a link to a graph showing the history of scores for this test. There is also a link to a treeherder page showing the jobs in a pushlog format.

To learn more about the regressing test(s), please see: https://wiki.mozilla.org/AWSY/Tests

Comment 1

[Glenn Watson [:gw]]

This is reasonably expected, at least for now. With this change we're more aggressively caching content in surfaces, in order to save power by not re-rendering content that is the same each frame, which has a memory cost.

There are, however, some changes coming in the next few weeks that will reduce this cost (depending on the test case), by using more granular tile sizes and detecting regions that are a single solid color only.

However, it's unlikely these will land in the next two weeks, due to the beta soft freeze next week, and me being on PTO the following week.

You can assign this to me, if you like, and I'll follow up when I work on and land the changes referenced above?

Comment 2

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

Sure gw, thanks!

Comment 3

[Glenn Watson [:gw]]

I investigated this regression today.

Previously, the tests in tp6 were disabling picture caching because of the presence of multiple scroll roots, which weren't supported.

The patch that regressed heap usage in this test allows picture caching to work on these test cases. Because of this, the WR code allocates some extra tile and dependency buffers to store the cached output of rendering frames.

So I think it's reasonable to say this is expected, and we can close as WONTFIX or similar. Does that sound reasonable?

Comment 4

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

Yes, it's reasonable, thanks!