Sunday, August 13, 2006

Python and kcachegrind

I've recently needed to profile some very subtle issues that cropped up in a customer's python application. However, when I tried to use hotshot, I consistently got tracebacks. After some digging around on the net, I saw folks saying that profiling is basically busted in python2.4 (and then I remembered Itamar saying basically the same thing at PyCon 2006 when we were looking at web2 slowness).

To get around this, I built python2.5 from svn and copied its cProfile, _lsprof and pstats files to my python2.4 libs. This was a complete desperation move and I totally didn't expect it to work -- but it did (with only a warning about a version mismatch).

Earlier this year, JP and Itamar updated an lsprof patch to work as a standalone. However, I've never done any profiling in python, so it took a few minutes to get up to speed. Looking at the patch source and the python2.5 cProfile docs and then doing the usual dir() and help() on cProfile.Profile in the python interpreter is what helped the most.

To give others new to profiling a jumpstart, I'm including a quick little toy howto below.

Import the junk:
>>> import os
>>> import cProfile
>>> import lsprofcalltree

Define a silly test function:
>>> def myFunc():
... myPath = os.path.expanduser('~/kcrw_s')
... print "Hello, world! This is my home:"
... print myPath

Define a profile object and run it:
>>> p = cProfile.Profile()
Hello, world! This is my home:
<cProfile.Profile object at 0xb7c87304>

Get the stats in a form kcachegrind can use and save it:
>>> k = lsprofcalltree.KCacheGrind(p)
>>> data = open('prof.kgrind', 'w+')
>>> k.output(data)
>>> data.close()

You can now open up the prof.kgrind file in kcachegrind and view the (in this case, very uninteresting) results to your heart's content.

Technorati Tags: ,