[webkit-dev] run-webkit-tests is moving to parallell testing by default (this weekend)

[Dirk Pranke]

We never implemented the "general way of marking subdirectories as
needing to run serially", but it would be easy to do if we needed to
[the 'http' dirs are still special-cased in the code].

There is code now (landed a few months ago) to control how many http
tests run in parallel separately from the main parallelism flag (so
you can run 16 workers but only 2 http tests at a time). I implemented
it but never flipped the switch because it was in the middle of Eric
flipping the bots over to NRWT in the first place. We should try
experimenting with this to see at some point once everything
stabilizes otherwise (this is the max_locked_shards() call in
manager.py; there's no command line flag).

-- Dirk

On Mon, Dec 5, 2011 at 11:17 AM, Ojan Vafai <ojan at chromium.org> wrote:
> Why is that? I don't know about other ports, but AFAIK, chromium writes to a
> mock clipboard and the Apple mac port writes to a local OS clipboard
> instance instead of the global one, specifically to avoid copy/paste tests
> interacting. Even without running tests in parallel, it's probably a good
> idea in order to avoid copy-pastes the developer is doing from affecting the
> test run.
>
> That said, I believe we have a general way of marking subdirectories as
> needing to run serially (which is what we do for http) if there are other
> reasons we need to.
>
> On Mon, Dec 5, 2011 at 11:12 AM, David Levin <levin at google.com> wrote:
>>
>> I believe there are some tests (copy/paste) that it would be very hard to
>> fully shard due to how they work.
>>
>> dave
>>
>>
>>
>> On Mon, Dec 5, 2011 at 11:08 AM, Ojan Vafai <ojan at chromium.org> wrote:
>>>
>>> I looked at one example that didn't exit
>>> early: http://build.webkit.org/builders/SnowLeopard%20Intel%20Release%20%28Tests%29/builds/35153/steps/layout-test/logs/stdio
>>>
>>> In that case, the http tests were the long tail and took 6 minutes longer
>>> than all the other tests. We don't split the http tests up because every
>>> time we've tried it's caused too much flakiness. It's unclear if the
>>> flakiness points to a bug in the test harness (e.g. in how we setup apache)
>>> or to bugs in the tests themselves or both. If someone has time to look into
>>> this, this is probably the biggest benefit to be found in NRWT runtime when
>>> running tests in parallel.
>>>
>>> FYI, NRWT outputs a log of the runtime after each run:
>>>
>>> 2011-12-03 03:09:30,018 58036 printing.py:462 INFO       worker/9:  4696
>>> tests, 1746.63 secs
>>> 2011-12-03 03:09:30,018 58036 printing.py:462 INFO       worker/8:  1177
>>> tests, 1693.47 secs
>>> 2011-12-03 03:09:30,018 58036 printing.py:462 INFO       worker/3:  1408
>>> tests, 2033.51 secs
>>> 2011-12-03 03:09:30,018 58036 printing.py:462 INFO       worker/2:   941
>>> tests, 2119.65 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/1:  1121
>>> tests, 2041.97 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/0:  1453
>>> tests, 2515.75 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/7:  1189
>>> tests, 1731.12 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/6:  3556
>>> tests, 2114.37 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/5:   948
>>> tests, 2097.13 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO       worker/4:  1411
>>> tests, 1716.66 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO      worker/15:   795
>>> tests, 2027.16 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO      worker/14:  1123
>>> tests, 1732.72 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO      worker/13:   425
>>> tests, 2021.25 secs
>>> 2011-12-03 03:09:30,019 58036 printing.py:462 INFO      worker/12:  1175
>>> tests, 1710.09 secs
>>> 2011-12-03 03:09:30,020 58036 printing.py:462 INFO      worker/11:  3462
>>> tests, 2096.30 secs
>>> 2011-12-03 03:09:30,020 58036 printing.py:462 INFO      worker/10:  1449
>>> tests, 1722.68 secs
>>> 2011-12-03 03:09:30,020 58036 printing.py:462 INFO    31120.45
>>> cumulative, 1945.03 optimal
>>>
>>> That shows you that, if we fully sharded all the tests, they would in
>>> theory take 1945 seconds to run, but worker/0 (the worker that runs the http
>>> tests) took 2515 seconds to run.
>>>
>>> Ojan
>>>
>>> On Mon, Dec 5, 2011 at 6:55 AM, Adam Roben <aroben at apple.com> wrote:
>>>>
>>>> On Dec 2, 2011, at 6:55 PM, Eric Seidel wrote:
>>>>
>>>> > The SnowLeopard bot went from a 1 hr 4 min (!?!) cycle time, to 38 min
>>>> > (still !?!).
>>>>
>>>> I suspect our Mac test bots could use a dose of RAM. Many of them only
>>>> have 3GB, since when you're running tests one by one you don't really need
>>>> much more.
>>>>
>>>> -Adam
>>>>
>>>> _______________________________________________
>>>> webkit-dev mailing list
>>>> webkit-dev at lists.webkit.org
>>>> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>>>
>>>
>>>
>>> _______________________________________________
>>> webkit-dev mailing list
>>> webkit-dev at lists.webkit.org
>>> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>>>
>>
>
>
> _______________________________________________
> webkit-dev mailing list
> webkit-dev at lists.webkit.org
> http://lists.webkit.org/mailman/listinfo.cgi/webkit-dev
>