Profiling¶
This is considered an advanced topic mainly of interest to server developers.
Introduction¶
Sometimes it can be useful to try to determine just how efficient a particular piece of code is, or to figure out if one could speed up things more than they are. There are many ways to test the performance of Python and the running server.
Before digging into this section, remember Donald Knuth’s words of wisdom:
[…]about 97% of the time: Premature optimization is the root of all evil.
That is, don’t start to try to optimize your code until you have actually identified a need to do so. This means your code must actually be working before you start to consider optimization. Optimization will also often make your code more complex and harder to read. Consider readability and maintainability and you may find that a small gain in speed is just not worth it.
Simple timer tests¶
Python’s timeit
module is very good for testing small things. For example, in order to test if it
is faster to use a for
loop or a list comprehension you could use the following code:
import timeit
# Time to do 1000000 for loops
timeit.timeit("for i in range(100):\n a.append(i)", setup="a = []")
<<< 10.70982813835144
# Time to do 1000000 list comprehensions
timeit.timeit("a = [i for i in range(100)]")
<<< 5.358283996582031
The setup
keyword is used to set up things that should not be included in the time measurement,
like a = []
in the first call.
By default the timeit
function will re-run the given test 1000000 times and returns the total
time to do so (so not the average per test). A hint is to not use this default for testing
something that includes database writes - for that you may want to use a lower number of repeats
(say 100 or 1000) using the number=100
keyword.
Using cProfile¶
Python comes with its own profiler, named cProfile (this is for cPython, no tests have been done
with pypy
at this point). Due to the way Evennia’s processes are handled, there is no point in
using the normal way to start the profiler (python -m cProfile evennia.py
). Instead you start the
profiler through the launcher:
evennia --profiler start
This will start Evennia with the Server component running (in daemon mode) under cProfile. You could
instead try --profile
with the portal
argument to profile the Portal (you would then need to
start the Server separately).
Please note that while the profiler is running, your process will use a lot more memory than usual.
Memory usage is even likely to climb over time. So don’t leave it running perpetually but monitor it
carefully (for example using the top
command on Linux or the Task Manager’s memory display on
Windows).
Once you have run the server for a while, you need to stop it so the profiler can give its report.
Do not kill the program from your task manager or by sending it a kill signal - this will most
likely also mess with the profiler. Instead either use evennia.py stop
or (which may be even
better), use @shutdown
from inside the game.
Once the server has fully shut down (this may be a lot slower than usual) you will find that
profiler has created a new file mygame/server/logs/server.prof
.
Analyzing the profile¶
The server.prof
file is a binary file. There are many ways to analyze and display its contents,
all of which has only been tested in Linux (If you are a Windows/Mac user, let us know what works).
We recommend the
Runsnake visualizer to see the processor usage
of different processes in a graphical form. For more detailed listing of usage time, you can use
KCachegrind. To make KCachegrind work with
Python profiles you also need the wrapper script
pyprof2calltree. You can get pyprof2calltree via
pip
whereas KCacheGrind is something you need to get via your package manager or their homepage.
How to analyze and interpret profiling data is not a trivial issue and depends on what you are
profiling for. Evennia being an asynchronous server can also confuse profiling. Ask on the mailing
list if you need help and be ready to be able to supply your server.prof
file for comparison,
along with the exact conditions under which it was obtained.
The Dummyrunner¶
It is difficult to test “actual” game performance without having players in your game. For this
reason Evennia comes with the Dummyrunner system. The Dummyrunner is a stress-testing system: a
separate program that logs into your game with simulated players (aka “bots” or “dummies”). Once
connected these dummies will semi-randomly perform various tasks from a list of possible actions.
Use Ctrl-C
to stop the Dummyrunner.
Warning: You should not run the Dummyrunner on a production database. It will spawn many objects and also needs to run with general permissions.
To launch the Dummyrunner, first start your server normally (with or without profiling, as above). Then start a new terminal/console window and active your virtualenv there too. In the new terminal, try to connect 10 dummy players:
evennia --dummyrunner 10
The first time you do this you will most likely get a warning from Dummyrunner. It will tell you to
copy an import string to the end of your settings file. Quit the Dummyrunner (Ctrl-C
) and follow
the instructions. Restart Evennia and try evennia --dummyrunner 10
again. Make sure to remove that
extra settings line when running a public server.
The actions perform by the dummies is controlled by a settings file. The default Dummyrunner
settings file is evennia/server/server/profiling/dummyrunner_settings.py
but you shouldn’t modify
this directly. Rather create/copy the default file to mygame/server/conf/
and modify it there. To
make sure to use your file over the default, add the following line to your settings file:
DUMMYRUNNER_SETTINGS_MODULE = "server/conf/dummyrunner_settings.py"
Hint: Don’t start with too many dummies. The Dummyrunner defaults to taxing the server much more intensely than an equal number of human players. A good dummy number to start with is 10-100.
Once you have the dummyrunner running, stop it with Ctrl-C
.
Generally, the dummyrunner system makes for a decent test of general performance; but it is of course hard to actually mimic human user behavior. For this, actual real-game testing is required.