Electrolysis/Release Criteria/Slow Script

e10s must not cause the slow-script dialog to appear more often. A/B testing on beta should indicate that the slow script dialog appears no more often than before.

RASCI

• Responsible: azhang
• Accountable: bsmedberg
• Supporting:
• Consulted:
• Informed: cpeterson, elan, release management

Summary

• It seems that there is indeed a higher number of slow-script notices per session on e10s.
• We've verified that we're measuring the same thing across e10s and non-e10s.
• Does this imply that e10s has more slow-script notifications? Possible confounding factors:
  • The dialog on e10s can potentially stop and re-show multiple times (theorized by :jimm). For example, if facebook.com freezes the content process for 12 seconds every minute, non-e10s will block while showing the dialog, while e10s will show the dialog, hide it after 12 seconds, and then reshow it, every minute.
  • Slow script notices can occur in multiple tabs at once in e10s. This seems like a rare occurrence, however, so it shouldn't be significant.
• Code like setInterval(function(){var x=Date.now();while(Date.now()-x<12000){}},30000) can show the slow script dialog on e10s every 30 seconds, while on non-e10s it will only show it once until you select an option in the dialog. It's not really possible to account for what the script might have done on non-e10s while the slow script dialog is shown and the user hasn't selected an option yet.
• :azhang thinks the measure isn't comparable this way - in the non-e10s case, the event triggers under slightly different conditions.

Next steps:

• See if it's possible to make the measure comparable between e10s/non-e10s, or mark it as a non-blocker if not possible.

Technical details:

• SLOW_SCRIPT_NOTICE_COUNT is the telemetry histogram
• There should be no differences in the slow script reporting mechanism between e10s and non-e10s; 1 call to ShowSlowScriptDialog means 1 addition to the histogram regardless of e10s setting. Manual testing supports this.
• The timeout has changed from 2 to 10 in e10s builds (see https://hg.mozilla.org/mozilla-central/rev/63680efe6d55#l35.12 for before and https://hg.mozilla.org/mozilla-central/rev/5002e9c06365#l2.13 for after). However, the value is correct on all beta versions, from 45.0b1 to 45.0b8 - the timeouts are identical between e10s and non-e10s.
• :azhang has verified that beta 45 releases with/without e10s both seem to increment SLOW_SCRIPT_NOTICE_COUNT if and only if there is a slow script notification shown, and that they seem to have the exact same timeout in all cases.

Links:

• bug 1249978 - [e10s] SLOW_SCRIPT_NOTICE_COUNT regressed in beta experiment
• https://gist.github.com/Uberi/f8591cf68aab36bbb758 - notebook prepared by azhang
• bug 1182638 - [e10s][telemetry] slow-script dialog appears twice as often in e10s vs non-e10s

Analysis results: SLOW_SCRIPT_NOTICE_COUNT shows up significantly more often in e10s. Analysis was reviewed by: rvitillo

How slow script notice count works on non-e10s:

• nsGlobalWindow::ShowSlowScriptDialog accumulates the SLOW_SCRIPT_NOTICE_COUNT histogram (https://dxr.mozilla.org/mozilla-central/source/dom/base/nsGlobalWindow.cpp#10677)

How slow script notice count works on e10s:

• [CHILD] nsGlobalWindow::ShowSlowScriptDialog sends a slow script notification via ProcessHangMonitor::NotifySlowScript (https://dxr.mozilla.org/mozilla-central/source/dom/base/nsGlobalWindow.cpp#10652)
• [CHILD] HangMonitorChild::NotifySlowScript invokes HangMonitorChild::NotifySlowScriptAsync to send out the slow script notice (https://dxr.mozilla.org/mozilla-central/source/dom/ipc/ProcessHangMonitor.cpp#352)
• [CHILD] PProcessHangMonitorChild::SendHangEvidence invokes PProcessHangMonitorParent::OnMessageReceived via IPC (https://dxr.mozilla.org/mozilla-central/source/obj-x86_64-unknown-linux-gnu/ipc/ipdl/PProcessHangMonitorChild.cpp#108)
• [PARENT] HangMonitorParent::RecvHangEvidence is called when PProcessHangMonitorParent receives a message from the child process (https://dxr.mozilla.org/mozilla-central/source/obj-x86_64-unknown-linux-gnu/ipc/ipdl/PProcessHangMonitorParent.cpp#288)
• [PARENT] HangMonitorParent::RecvHangEvidence sends out a "process-hang-report" event by creating a HangObserverNotifier task (https://dxr.mozilla.org/mozilla-central/source/dom/ipc/ProcessHangMonitor.cpp#610)
• [PARENT] ProcessHangMonitor.reportHang is called when the ProcessHangMonitor receives a "process-hang-report" event (https://dxr.mozilla.org/mozilla-central/source/browser/modules/ProcessHangMonitor.jsm#66)
• [PARENT] JS function ProcessHangMonitor.reportHang accumulates the SLOW_SCRIPT_NOTICE_COUNT histogram (https://dxr.mozilla.org/mozilla-central/source/browser/modules/ProcessHangMonitor.jsm#403)
• :azhang checked that this is currently the only code path that results in SLOW_SCRIPT_NOTICE_COUNT being incremented.

Bugs

Full Query

ID	Summary	Whiteboard	Keywords	Assigned to
1260769	[e10s] SLOW_SCRIPT_PAGE_COUNT regressed under e10s		regression

1 Total; 1 Open (100%); 0 Resolved (0%); 0 Verified (0%);