Profiling can mean a few different things, so here are the current ways!
Profiling with Valgrind¶
This option is Linux only. See Profiling with Valgrind for details.
Profiling with cProfile¶
Mantid supports profiling of its underlying code using the cProfile module from python.
Launch Mantid Workbench from the command line using the script in the installation folder, adding the
--profile modifier with a path to the output file.
path/to/install/bin/workbench --profile path/to/outputFile.prof
This profile can then either be sent to the developers upon request or be analysed using the python module snakeviz.
Profiling with Yappi¶
cProfile can only profile the current thread, this makes it useful for diagnosing PyQt performance issues (such as random hangs when a user does something), but poor for diagnosing workflow algorithm issues.
Yappi (Yet Another Python Profiler) is multithreading aware. cProfile which will show everything up to the exec, or run call. Yappi will profile the thread and show any slow points within the separate thread.
Note: This will not with QThreads and may crash or be unable to profile inside. It is recommended you migrate workers by inheriting from IQtAsync, see async_qt_adaptor.py for details. Using native threading enables tooling, makes testing simple and produces readable stack traces.
To use Yappi instead of cProfile simple append –yappi to the list of arguments after profile, e.g.:
# If you do not have yappi installed python3 -m pip install yappi --user # Run with Yappi profiler ./MantidWorkbench --profile name_of_profile.out --yappi
KCachegrind can be used to view profiling data, see Profiling with Valgrind for more details on usage.
Profiling an algorithm¶
On Linux, the build can be configured to generate algorithm profiling information. See AlgorithmProfiler for details.
Other Profiling Tools¶
KCachegrind visualizes callgrind output.
See Profiling With Valgrind for help on running callgrind
Takes snapshot of run and prints percentage of calls in functions
See here for a list of other tools: http://www.pixelbeat.org/programming/profiling/
Start/stop recording of program using a button
Not as detailed or flexible as callgrind