You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I just ran the tests on my machine and got the following printout:
Ran 247 tests in 73.629s
I checked and if I run only this test in test_cartesian.py I get:
$ nosetests quantecon/tests/test_cartesian.py:test_performance_C
.
----------------------------------------------------------------------
Ran 1 test in 23.669s
Similar timings for the test_performance_F function:
$ nosetests quantecon/tests/test_cartesian.py:test_performance_F
.
----------------------------------------------------------------------
Ran 1 test in 24.902s
These two tests account for about 2/3 of the total testing time. Is there a way we can not run these tests by default, or at least have their run time not take so long?
We could probably still update the travis script to make sure they are checked by travis.
These long-running tests are actually meant to measure performance by comparing with numpy.
They can be ignored by running nosetests --exclude='.*performance.*' quantecon.
Actually, this seems like the easiest way to differentiate long-running tests from short running ones: decide a naming conventions for the former. (I remember some earlier talks about adding special functions attributes).
The reason why this solution does not completely satisfy me, is that the performance results directly go to /dev/null. So until we start to focus on systematic performance measurements and compare the results, we can as well comment the corresponding code...
I ended up just having nose run the test file(s) relevant to what I was changing so this wasn't a big issue.
I think if we were to implement a suite of performance tests that are monitored via vbench (as suggested in #44) that these tests would be perfect for there.
I think we have talked about why they are slow -- feel free to close this unless you are planning on changing something and want to leave it open as a reminder.
See update to #44. Is vbench worth investing in? It doesn't seem to be active. Pandas results end in 2012. Statsmodels seems to be more up to date with results for April 2014.
Activity
albop commentedon Jan 24, 2015
These long-running tests are actually meant to measure performance by comparing with numpy.
They can be ignored by running
nosetests --exclude='.*performance.*' quantecon.Actually, this seems like the easiest way to differentiate long-running tests from short running ones: decide a naming conventions for the former. (I remember some earlier talks about adding special functions attributes).
The reason why this solution does not completely satisfy me, is that the performance results directly go to
/dev/null. So until we start to focus on systematic performance measurements and compare the results, we can as well comment the corresponding code...sglyon commentedon Jan 28, 2015
@albop thanks of the comments.
I ended up just having nose run the test file(s) relevant to what I was changing so this wasn't a big issue.
I think if we were to implement a suite of performance tests that are monitored via vbench (as suggested in #44) that these tests would be perfect for there.
I think we have talked about why they are slow -- feel free to close this unless you are planning on changing something and want to leave it open as a reminder.
albop commentedon Jan 28, 2015
Sure. Let's wait for the vbench to emerge and close the issue afterwards so that we don't forget about these few lines of code.
mmcky commentedon Aug 5, 2015
See update to #44. Is
vbenchworth investing in? It doesn't seem to be active. Pandas results end in 2012. Statsmodels seems to be more up to date with results for April 2014.oyamad commentedon Aug 18, 2017
Can we add
@attr('slow')totest_performance_*for testing on a local machine(and add?-a speed=slowto.travis.yml)[-]TEST: Tests for cartesian take a long time[/-][+]TEST: Tests for grid_tools take a long time[/+]