Bug 113089 – 5.6% memory regression on page_cycler_intl1

• Status: NEW
• Component: WebCore Misc.

[Rafael Weinstein]

[Remember to CC the gardeners (sheriffbot:sheriffs on #webkit). If there's a clear culprit, assign the bug to that person, otherwise, assign it to the current gardener]

Most likely regression range: http://trac.webkit.org/log/?verbose=on&rev=146412&stop_rev=146398

REGRESSIONS
http://chromium-perf.appspot.com/?tab=chromium-rel-win7-webkit&graph=ws_final_r&rev=189577&history=150&master=ChromiumWebkit&testSuite=page_cycler_intl1&details=true

May be: https://bugs.webkit.org/show_bug.cgi?id=109021, or https://bugs.webkit.org/show_bug.cgi?id=112088

[Rafael Weinstein]

Note that there was also a page load improvement in this same range:

http://chromium-perf.appspot.com/?tab=chromium-rel-win7-webkit&graph=times&trace=t&rev=189739&history=150&master=ChromiumWebkit&testSuite=page_cycler_morejs&details=true

[Rafael Weinstein]

Also, there was an improvement in DHTML:

http://chromium-perf.appspot.com/?tab=chromium-rel-win7-webkit&graph=ws_peak_r&trace=ws_pk_r&rev=189708&history=150&master=ChromiumWebkit&testSuite=page_cycler_dhtml&details=true

[Peter Kasting]

I am skeptical that this is a WebKit-related change.

The two changes mentioned in comment 0 are the only changes in the range that should affect Chromium.  But looking at the patches, bug 109021 is mostly disabled since Chromium doesn't enable CSS3_TEXT, and bug 112088 clearly doesn't touch memory usage at all.

The Chromium changes for this test run are at http://build.chromium.org/f/chromium/perf/dashboard/ui/changelog.html?url=%2Ftrunk%2Fsrc&range=189475%3A189440&mode=html .  Is it possible any of those is the culprit?

[Tien-Ren Chen]

Is there many RTL pages in this test suite? And is accelerated overflow scrolling enabled on the tests? If both answers are yes, then https://bugs.webkit.org/show_bug.cgi?id=112088 might result in more memory usage since it fixes a bug with accelerated scrolling that the negative part of the overflow can't be scrolled into view. (so some of the tiles were never allocated)

[Peter Kasting]

(In reply to comment #4)
> Is there many RTL pages in this test suite? And is accelerated overflow scrolling enabled on the tests? If both answers are yes, then https://bugs.webkit.org/show_bug.cgi?id=112088 might result in more memory usage since it fixes a bug with accelerated scrolling that the negative part of the overflow can't be scrolled into view. (so some of the tiles were never allocated)

Hmm, that's a possibility.  Since this is the "intl1" suite, it could have a lot of RTL pages.

Tien-Ren, what do you think about temporarily reverting your change, seeing if the working set drops, and if so, closing this bug and considering this an "expected memory-size increase due to bug fix"?  Rafael, do you know if in that case we'd need to update some size expectation lines anywhere?

[Tien-Ren Chen]

(In reply to comment #5)
> (In reply to comment #4)
> > Is there many RTL pages in this test suite? And is accelerated overflow scrolling enabled on the tests? If both answers are yes, then https://bugs.webkit.org/show_bug.cgi?id=112088 might result in more memory usage since it fixes a bug with accelerated scrolling that the negative part of the overflow can't be scrolled into view. (so some of the tiles were never allocated)
> 
> Hmm, that's a possibility.  Since this is the "intl1" suite, it could have a lot of RTL pages.
> 
> Tien-Ren, what do you think about temporarily reverting your change, seeing if the working set drops, and if so, closing this bug and considering this an "expected memory-size increase due to bug fix"?  Rafael, do you know if in that case we'd need to update some size expectation lines anywhere?

I'm not very familiar with the page cycler test suite. What is the easiest steps to get it running on my workstation? We can make local builds to compare the memory usage.

[Peter Kasting]

(In reply to comment #6)
> I'm not very familiar with the page cycler test suite. What is the easiest steps to get it running on my workstation? We can make local builds to compare the memory usage.

Not sure offhand.  I'm hoping the perf sheriff or some Chromium troopers would know this.  According to build.chromium.org, the current perf sheriff is bulach, who's probably in Europe and thus unlikely to be around.  Try asking on #chromium for sheriffs or troopers who might know this, or ask the infrastructure people?

[James Robinson]

Does the regression show up on the perf canaries?  That will tell us whether the regression was in Chromium or in WebKit and hopefully provide a tighter range.  Please check that before reverting anything.

[Peter Kasting]

(In reply to comment #8)
> Does the regression show up on the perf canaries?  That will tell us whether the regression was in Chromium or in WebKit and hopefully provide a tighter range.  Please check that before reverting anything.

This bot was a perf canary.

[Peter Kasting]

For now I'm going to assign this to trchen to follow up on while I continue sheriffing, but I'm on the CC list in case something comes up.