Closed Bug 1317261 Opened 8 years ago Closed 8 years ago

Split tasks of reftests on Windows 7 debug build

Categories

(Infrastructure & Operations Graveyard :: CIDuty, task)

task
Not set
normal

Tracking

(Not tracked)

RESOLVED FIXED

People

(Reporter: xidorn, Assigned: aobreja)

Details

Attachments

(2 files, 6 obsolete files)

In bug 1258916, I'm importing a large set of tests from the CSSWG's test repo, which seems to significantly increase the time to run reftests on Windows 7 debug build, which now constantly exceeds the 7200s limit. Is it possible to split the reftest tasks there, or extend the time limit? (Not sure what's the right component for this issue...)
Moving to proper component...
Component: Build Config → Buildduty
Product: Core → Release Engineering
QA Contact: bugspam.Callek
Assignee: nobody → aobreja
Attached patch bug1317261_bb.patch (obsolete) — Splinter Review
The patch to split in 2 chunks the tasks of reftests on Windows 7 debug build.
Attachment #8810828 - Flags: review?(kmoir)
Ru is also going to timeout. Does this patch also cover that?
Attached file diff.txt (obsolete) —
cat old | wc -l 14937 cat new | wc -l 14950 cat old | grep 'Windows 7' | grep ' reftest ' | grep -v 'e10s' | grep 'debug' | wc -l 13
R, Ru, and R-e10s are all going to timeout, so they may need to be splitted as well. See https://treeherder.mozilla.org/#/jobs?repo=try&revision=ee051ed6ab12
Attachment #8810828 - Flags: review?(kmoir) → review+
Win8 reftests don't run by default on Try - I think you should opt-in to those as I expect they'll need the same treatment here.
Flags: needinfo?(xidorn+moz)
Andrei, the Try push from comment 5 also shows OSX debug around 90-100min, which is pretty high. In bug 1317434, I'm splitting both OSX and Windows debug e10s reftests into two chunks to get their runtimes down to something saner, so maybe you can just get both platforms here as well.
Flags: needinfo?(aobreja)
Looks like Win8 debug reftests need split as well: https://treeherder.mozilla.org/#/jobs?repo=try&revision=a3b480a42bd2c4f6262c3476f7df80597c4c03cf R-e10s takes ~109min, and R seems to exceed the time limit.
Flags: needinfo?(xidorn+moz)
Attached patch bug1317261_bb.patch (obsolete) — Splinter Review
This patch split in 2 chunks all tests from Windows7,Windows8 and OSX for: reftest reftest-noaccel reftest_e10s reftest-noaccel_e10s Ryan I know that you also make this split for *e10-s for Windows 8,7 and OSX in bug 1317434. So what should be the order here,first to push your patch and I will recreate this one (with changes only for reftest and reftest_noaccel on Win 7,8 as OSX tests only occur for e10-s) or should I push this patch first and modify yours.
Attachment #8810828 - Attachment is obsolete: true
Flags: needinfo?(aobreja)
Attachment #8811721 - Flags: review?(ryanvm)
Attached file diff (obsolete) —
cat new | wc -l 14813 cat old | wc -l 14721
Attachment #8810829 - Attachment is obsolete: true
Comment on attachment 8811721 [details] [diff] [review] bug1317261_bb.patch Review of attachment 8811721 [details] [diff] [review]: ----------------------------------------------------------------- I'll rebase around yours, thanks for asking! That said, it looks like OSX only split reftest-e10s and not the non-e10s variant as well? ::: mozilla-tests/config.py @@ +3223,5 @@ > + WEB_PLATFORM_REFTESTS_E10S + MOCHITEST_BC_7_E10S + MOCHITEST_CLIPBOARD_E10S + \ > + MOCHITEST_GPU_E10S + REFTEST_NOACCEL_E10S_TWO_CHUNKS + WEB_PLATFORM_TESTS_CHUNKED_MORE_E10S > + BRANCHES['try']['platforms'][platform][slave_platform]['opt_unittest_suites'] += \ > + base_tests + REFTEST_NOACCEL_E10S + WEB_PLATFORM_TESTS_CHUNKED_E10S > + Rather that duplicating this logic, just split Win10 too.
Attachment #8811721 - Flags: review?(ryanvm)
Attached patch bug1317261_bb_v1.patch (obsolete) — Splinter Review
Split in 2 chunks all tests that run on Win 7,8 and OSX for : reftest reftest-noaccel reftest_e10s reftest-noaccel_e10s
Attachment #8811721 - Attachment is obsolete: true
Attachment #8811741 - Flags: review?(kmoir)
Attached file diff_v1 (obsolete) —
cat new | wc -l 14813 cat old | wc -l 14721
Attachment #8811741 - Flags: review?(kmoir) → review+
This may not be strictly necessary now. After experimenting some optimization, it seems the run time can be significantly reduced. In https://treeherder.mozilla.org/#/jobs?repo=try&revision=e20e45bf6c9d44c6c05a5f21600990afc7104b3b, majority of the tests now run under 100min, which gives us a large enough margin. The oranges there are from some unexpected test result (UNEXPECTED PASS?!), not timeout anymore. The time is still much longer than Linux builds, so you may still want to apply this split, but it's no longer a blocker.
No longer blocks: 1258916
Attached file diff
cat old |wc -l 14721 cat new |wc -l 14827 also spliting reftest and reftest_noaccel for OSX
Attachment #8811722 - Attachment is obsolete: true
Attachment #8811742 - Attachment is obsolete: true
In production. Initial results are also showing only a minor increase in overall runtime, so that's encouraging from an overhead and load standpoint.
Status: NEW → RESOLVED
Closed: 8 years ago
Resolution: --- → FIXED
Product: Release Engineering → Infrastructure & Operations
Product: Infrastructure & Operations → Infrastructure & Operations Graveyard
You need to log in before you can comment on or make changes to this bug.

Attachment

General

Created:
Updated:
Size: