Engine Profiling: Difference between revisions
cleanup |
|||
| Line 84: | Line 84: | ||
'''Extract the data''' | '''Extract the data''' | ||
<code><pre> | <code><pre> | ||
BIN_PATH=~/Projects/spring/installs/cmake/master/linux64/debug/spring | BIN_PATH=~/Projects/spring/installs/cmake/master/linux64/debug/spring | ||
| Line 132: | Line 131: | ||
The result will look about like this: | The result will look about like this: | ||
http://wiki.jrfonseca.googlecode.com/hg/xdot-profile2.png | http://wiki.jrfonseca.googlecode.com/hg/xdot-profile2.png | ||
== Windows == | == Windows == | ||
Revision as of 13:29, 20 November 2010
Development < Engine Profiling
Engine Profiling
Linux
Intro
There are several ways to profile an application.
gprof
The compiler injects special code in the binary that self profiles it when it runs. this is what cmake -DCMAKE_BUILD_TYPE=PROFILING is here for. We are not going to use this. It does not interfere with normal program's work-flow. With spring, it gets so slow that it is unusable for profiling anything in late game (which is where it is needed the most). It requires special compiling options.
oprof
OProfile uses statistical sampling. It is a kernel module, which every n clock ticks, registers which function is currently being processed. It is lightweight; it may run with almost 0 performance impact. This method always profiles your profiles your whole system; relevant info has to be filtered out after the profiling run has finished. The downside of statistical profiling, is that it requires at least a certain sample size to achieve good accuracy. Also remember, that oprof still requires the debug symbols, to resolve addresses into function names. It can be turned on/off at any time, so for instance, you may skip spring loading so you would only profile in-game runtime. You can also merge multiple profiling sessions.
Setting up oprof
Required stuff:
- oprof kernel module
- oprof package
Using oprof
| Note: Modern CPUs have special statistical sampling timing functions to make profiling with software like oprof fast. |
You will have to load the kernel module. The profiling daemon is controlled using opcontrol as root. As regular user, you can do post-processing using opreport.
Recommended settings
- use a call-graph depth of size 16
- disable kernel profiling & enable user-space profiling
- separate each application in its own profile
- separate each thread in its own profile; you can always merge them later, using opreport
A typical profiling session
# We need to be root
su -
# In case the kernel module is not already loaded ...
insmod oprofile
# This has to be sent only once; it will be saved
opcontrol --event=CPU_CLK_UNHALTED:1000000:0x0:0:1 \
--callgraph=16 \
--separate=thread,library \
--no-vmlinux
# -- Bring spring into a state where you want to start profiling --
# start the profiler
opcontrol --start
# -- In spring, do whatever you want to profile --
# stop the profiler
opcontrol --stop
# in order to make the profile data available to user-space,
# you have to dump the data
opcontrol --dump
Post processing oprof data
With the settings we used above, oprofile indexes stuff by thread- and process-ID. In order to generate a report, we will have to tell which program, which threads and which sampling sessions to include. If you do not care much, you can just merge thread data and sampling data.
Extract the data
BIN_PATH=~/Projects/spring/installs/cmake/master/linux64/debug/spring
DATA_FILE_PATH=./springProfilingData.oprof
# these two will be used later
GRAPH_FILE_PATH=./springProfilingGraph.dot
IMAGE_FILE_PATH=./springProfilingGraph.png
# Caution: this would take a long time!
# ~ 3h for a spring in-game time of 5min
# --debug-info \
# without it, it shrinks to ~ 10s
opreport \
--long-filenames \
--callgraph \
--demangle=smart \
--merge=tid,tgid \
--symbols ${BIN_PATH} \
--output-file ${DATA_FILE_PATH}
--callgraph tells it to plot a full call-graph table, which can be parsed by other tools.
If instead you want a flat CSV table, omit it.
Converting the data file to a dot-graph file
gprof2dot takes oprof data and outputs a nice dot-graph.
To generate the dot graph out of the --callgraph version, use this command:
gprof2dot --format=oprofile -o=${GRAPH_FILE_PATH} ${DATA_FILE_PATH}
Then you can either use XDot to view it directly, or create an image out of the graph file.
Createing an image out of the graph file
We will use GraphViz's dot utility to render to an image.
dot -Tpng ${GRAPH_FILE_PATH} > ${IMAGE_FILE_PATH}
The -Tpng part can be replaced by -Tsvg, -Tgif and more.
The result will look about like this:
Windows
<Please add>