test item scheduling does not schedule N items to N workers

[nchammas]

I may just be misunderstanding how xdist works, but I can't seem to get the number of workers I requested to all accept tasks in parallel:

$ py.test -v -n 8
================================ test session starts =================================
platform darwin -- Python 3.5.1, pytest-2.8.5, py-1.4.31, pluggy-0.3.1 -- .../pytest-xdist-testcase/venv/bin/python3
cachedir: .cache
rootdir: .../pytest-xdist-testcase, inifile: 
plugins: xdist-1.13.1
[gw0] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw1] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw2] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw3] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw4] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw5] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw6] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw7] darwin Python 3.5.1 cwd: .../pytest-xdist-testcase
[gw0] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw1] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw2] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw3] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw4] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw5] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw6] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
[gw7] Python 3.5.1 (default, Dec  7 2015, 21:59:10)  -- [GCC 4.2.1 Compatible Apple LLVM 7.0.0 (clang-700.1.76)]
gw0 [8] / gw1 [8] / gw2 [8] / gw3 [8] / gw4 [8] / gw5 [8] / gw6 [8] / gw7 [8]
scheduling tests via LoadScheduling

test_tests.py::test_1[0] 
test_tests.py::test_1[2] 
test_tests.py::test_2[0] 
test_tests.py::test_2[2] 
[gw0] PASSED test_tests.py::test_1[2] 
[gw7] PASSED test_tests.py::test_1[0] 
[gw2] PASSED test_tests.py::test_2[0] 
[gw4] PASSED test_tests.py::test_2[2] 
test_tests.py::test_2[1] 
test_tests.py::test_1[3] 
test_tests.py::test_1[1] 
test_tests.py::test_2[3] 
[gw7] PASSED test_tests.py::test_1[1] 
[gw0] PASSED test_tests.py::test_1[3] 
[gw4] PASSED test_tests.py::test_2[3] 
[gw2] PASSED test_tests.py::test_2[1] 

============================= 8 passed in 64.12 seconds ==============================

Watching the output as it comes, it looks like the second block of 4 tests doesn't start running until the first block of 4 completes.

Am I misunderstanding something? Why don't all 8 workers start running tests at the same time? Why doesn't gw1, for example, get any tests?

[nonatomiclabs]

Do you have 8 physical cores on your computer?

If not, I think it is normal to see only 4 processes active.

[RonnyPfannschmidt]

it clearly starts 8 workers, however due to a design problem thats sheduled to be fixed in the next bigger sprint it distributes tests in batches of 2 items, and thus sheduling is rather subpar

for your own case, the problem will elevae as the number of tests gows tho

AFAIR this problem was noted on the py.test issue tracker already

its also related to #17 #18 and #20

[nchammas]

@Filaton

Do you have 8 physical cores on your computer?

If not, I think it is normal to see only 4 processes active.

I only have 2 virtual cores. I don't think this is normal, as @RonnyPfannschmidt's comment confirms. I expect the 8 workers to all start accepting tasks at the same time. This should have nothing to do with my physical cores.

Of course, if I run 8 tasks on a machine with 2 cores, that means there will be a lot of context switching between the tasks. That's fine with me since my tests are I/O-heavy.

@RonnyPfannschmidt

AFAIR this problem was noted on the py.test issue tracker already

Do you have the py.test issue number?

Going forward, should xdist issues be reported here or on the main py.test issue tracker? I only searched here before reporting this issue. Apologies if it's a dup. I think it would be better if xdist issues were tracked exclusively in the xdist repo.

[nonatomiclabs]

@nchammas Sorry, my fault. In fact, I meant that if you have only 4 cores, it's normal that tests will pass by group of 4 (supposing that they last the same time).
I'll avoid commenting issues in the morning… ;)

[RonnyPfannschmidt]

a current problem with xdist is, that the current runtest protocol needs a next item, the issue arises from a legacy problem as non function scoped fixtures/caches got introduced,
its very likely unfixable before pytest 3.0

[nck974]

Is there any plan on doing something about this now that pytest is in version >7?

[RonnyPfannschmidt]

the need in pytest for next-item has not been elevated or changed,
possible resolutions for enabling a sheduler passtrough have been discussed, no plans have been made due to lack of manpower