This is python profiling primer.

For the purpose of this document, qubes-dev is name of the domain used for postprocessing profiling stats.


yum install gprof2dot graphviz
git clone

If you profile something on dom0, move from repository to dom0:

mkdir -p ~/profiling
qvm-run -p qubes-dev 'cat ~/profiling/' > ~/profiling/
  • WARNING: this will obviously be running third party code which is not signed by ITL nor Fedora. You have been warned.


Identify function responsible for some slow action

You have to select area in which you suspect less than optimal performance. If you do not narrow the area, graphs may be unreadable.

Replace suspect function with probe


def foo(self, bar):
    # function content


def foo(self, *args, **kwargs):
    profile.runctx('self.real_foo(*args, **kwargs)', globals(), locals(),

def real_foo(self, bar):
    # function content

Run application

Beware that some functions may be called often. For example qubesmanager/ gets run once per second. This will produce one pstat file per second.

Remember to revert your changes to application afterwards.

Upload statistics

If you are in dom0:

cd ~/profiling



For every ${basename}.pstats this will produce ${basename}.txt and ${basename}.svg. SVG contains call graph. Text file contains list of all functions sorted by cumulative execution time. You may also try make all-png.

make index.html

This creates index.html with all SVG graphics linked to TXT files. Ready for upload.

make upload


This example is from qubes-manager (qubesmanager/


It is apparent than problem is around get_disk_usage which calls something via It does it 15 times, probably once per VM.