Bug 138958

Summary: [meta] Run http tests parallel
Product: WebKit Reporter: Csaba Osztrogonác <ossy>
Component: New BugsAssignee: Alexey Proskuryakov <ap>
Status: RESOLVED FIXED    
Severity: Normal CC: ap, bfulgham, commit-queue, dbates, ddkilzer, glenn, ossy
Priority: P2    
Version: 528+ (Nightly build)   
Hardware: Unspecified   
OS: Unspecified   
See Also: https://bugs.webkit.org/show_bug.cgi?id=139225
https://bugs.webkit.org/show_bug.cgi?id=139226
https://bugs.webkit.org/show_bug.cgi?id=139279
Bug Depends on: 136722, 138583, 138894, 138959, 139135, 139149, 139243, 139347    
Bug Blocks:    
Attachments:
Description Flags
proposed patch
dbates: review+
patch for landing none

Csaba Osztrogonác
Reported 2014-11-21 03:30:00 PST
It is a meta bug to collect issues related to running http tests parallel. The final goal is to be able run them parallel as the other tests everywhere and be able remove the complex logic needed to serialize only http tests.
Attachments
proposed patch (37.17 KB, patch)
2014-12-04 16:06 PST, Alexey Proskuryakov 		dbates: review+
DetailsFormatted DiffDiff
patch for landing (37.17 KB, patch)
2014-12-04 16:58 PST, Alexey Proskuryakov 		no flags
DetailsFormatted DiffDiff
Show Obsolete (1) View All Add attachment proposed patch, testcase, etc.
Alexey Proskuryakov
Comment 1 2014-12-03 11:34:07 PST
I think that we are ready to flip the switch (at least on Mac), but I can't figure out which switch to flip. There are still some failures when running with --fully-parallel, but they should not affect normal mode, as far as I can tell.
Csaba Osztrogonác
Comment 2 2014-12-03 12:01:14 PST
(In reply to comment #1) > I think that we are ready to flip the switch (at least on Mac), but I can't > figure out which switch to flip. > > There are still some failures when running with --fully-parallel, but they > should not affect normal mode, as far as I can tell. "-f, --fully-parallel - run all tests in parallel" Ah, this help message is a little bit confusing. Tests are sharded by subdirectories, leaf subdirectories are atomic and can't be distributed to different threads by default. Threads pick up shards from the queue. This fully parallel option disables sharding and each threads can pick up tests from the queue instead of shards. I think we aren't ready for it yet, but it can reveal many flakiness. It would be a good expetiment. Back to parallel http test running. Now http and perf tests are considered as locked tests and locked tests run on 1 thread by default. You can easily override it with: --max-locked-shards command line option or WEBKIT_TEST_MAX_ LOCKED_SHARDS. (cmdline is the stronger one) The default option is defined to 1 in base.py in default_max_locked_shards. First we can override it for ports one by one. And then in base.py once it is stable everywhere. The ideal number would be the same as default_child_processes, but it depends on how much concurrent request can be served by the http server.
Alexey Proskuryakov
Comment 3 2014-12-03 12:36:35 PST
The question is how do I make HTTP tests parallel while keeping perf tests non-parallel.
Csaba Osztrogonác
Comment 4 2014-12-03 13:42:05 PST
(In reply to comment #3) > The question is how do I make HTTP tests parallel while keeping perf tests > non-parallel. I think the easiest way is what I suggested in bug136722 - run them in a different buildstep on the bots.
Csaba Osztrogonác
Comment 5 2014-12-03 13:45:06 PST
One more thing. perf tests can run parallel now with non perf tests. If we increasw max_locked_shards, they will run parallel with more http test thread instead of random non-http test. But 2 perf test won't run parallel, because they are in the same shard.
Alexey Proskuryakov
Comment 6 2014-12-04 16:06:15 PST
Created attachment 242595 [details] proposed patch This makes debug tests run in 8 minutes and 12 seconds on my Mac Pro. Without the patch, they used to take over 15 minutes.
Daniel Bates
Comment 7 2014-12-04 16:48:09 PST
Comment on attachment 242595 [details] proposed patch View in context: https://bugs.webkit.org/attachment.cgi?id=242595&action=review Looks sane to me. > Tools/ChangeLog:17 > + think that these parameters do what we wanted them to do; Apache handles the load Nit: do => did Or wanted => want
Alexey Proskuryakov
Comment 8 2014-12-04 16:58:32 PST
Created attachment 242599 [details] patch for landing
WebKit Commit Bot
Comment 9 2014-12-04 17:36:33 PST
Comment on attachment 242599 [details] patch for landing Clearing flags on attachment: 242599 Committed r176830: <http://trac.webkit.org/changeset/176830>
WebKit Commit Bot
Comment 10 2014-12-04 17:36:38 PST
All reviewed patches have been landed. Closing bug.
Note You need to log in before you can comment on or make changes to this bug.