This is a Python script to convert the output from many profilers into a dot graph.
- read output from:
- prune nodes and edges below a certain threshold;
- use an heuristic to propagate time inside mutually recursive functions;
- use color efficiently to draw attention to hot-spots;
- work on any platform where Python and Graphviz is available, i.e, virtually anywhere.
If you want an interactive viewer for the graphs generated by gprof2dot, check xdot.py.
gprof2dot currently fulfills my needs, and I have little or no time for its maintenance. So I'm afraid that any requested features are unlikely to be implemented, and I might be slow processing issue reports or pull requests.
- Python: known to work with version 2.7 and 3.3; it will most likely not work with earlier releases.
- Graphviz: tested with version 2.26.3, but should work fine with other versions.
apt-get install python graphviz
Usage: gprof2dot.py [options] [file] ... Options: -h, --help show this help message and exit -o FILE, --output=FILE output filename [stdout] -n PERCENTAGE, --node-thres=PERCENTAGE eliminate nodes below this threshold [default: 0.5] -e PERCENTAGE, --edge-thres=PERCENTAGE eliminate edges below this threshold [default: 0.1] -f FORMAT, --format=FORMAT profile format: axe, callgrind, hprof, json, oprofile, perf, prof, pstats, sleepy, sysprof or xperf [default: prof] --total=TOTALMETHOD preferred method of calculating total time: callratios or callstacks (currently affects only perf format) [default: callratios] -c THEME, --colormap=THEME color map: color, pink, gray, bw, or print [default: color] -s, --strip strip function parameters, template parameters, and const modifiers from demangled C++ function names -w, --wrap wrap function names --show-samples show function samples -z ROOT, --root=ROOT prune call graph to show only descendants of specified root function -l LEAF, --leaf=LEAF prune call graph to show only ancestors of specified leaf function --skew=THEME_SKEW skew the colorization curve. Values < 1.0 give more variety to lower percentages. Values > 1.0 give less variety to lower percentages
perf record -g -- /path/to/your/executable perf script | c++filt | gprof2dot.py -f perf | dot -Tpng -o output.png
opcontrol --callgraph=16 opcontrol --start /path/to/your/executable arg1 arg2 opcontrol --stop opcontrol --dump opreport -cgf | gprof2dot.py -f oprofile | dot -Tpng -o output.png
If you're not familiar with xperf then read this excellent article first. Then do:
Start xperf as
xperf -on Latency -stackwalk profile
Run your application.
Save the data. ` xperf -d output.etl
Start the visualizer:
In Trace menu, select Load Symbols. Configure Symbol Paths if necessary.
Select an area of interest on the CPU sampling graph, right-click, and select Summary Table.
In the Columns menu, make sure the Stack column is enabled and visible.
Right click on a row, choose Export Full Table, and save to output.csv.
Then invoke gprof2dot as
gprof2dot.py -f xperf output.csv | dot -Tpng -o output.png
VTune Amplifier XE
Collect profile data as (also can be done from GUI):
amplxe-cl -collect hotspots -result-dir output -- your-app
Visualize profile data as:
amplxe-cl -report gprof-cc -result-dir output -format text -report-output output.txt gprof2dot.py -f axe output.txt | dot -Tpng -o output.png
See also Kirill Rogozhin's blog post.
/path/to/your/executable arg1 arg2 gprof path/to/your/executable | gprof2dot.py | dot -Tpng -o output.png
python -m profile -o output.pstats path/to/your/script arg1 arg2 gprof2dot.py -f pstats output.pstats | dot -Tpng -o output.png
python cProfile (formerly known as lsprof)
python -m cProfile -o output.pstats path/to/your/script arg1 arg2 gprof2dot.py -f pstats output.pstats | dot -Tpng -o output.png
python hotshot profiler
The hotshot profiler does not include a main function. Use the hotshotmain.py script instead.
hotshotmain.py -o output.pstats path/to/your/script arg1 arg2 gprof2dot.py -f pstats output.pstats | dot -Tpng -o output.png
java -agentlib:hprof=cpu=samples ... gprof2dot.py -f hprof java.hprof.txt | dot -Tpng -o output.png
See Russell Power's blog post for details.
A node in the output graph represents a function and has the following layout:
+------------------------------+ | function name | | total time % ( self time % ) | | total calls | +------------------------------+
- total time % is the percentage of the running time spent in this function and all its children;
- self time % is the percentage of the running time spent in this function alone;
- total calls is the total number of times this function was called (including recursive calls).
An edge represents the calls between two functions and has the following layout:
total time % calls parent --------------------> children
- total time % is the percentage of the running time transfered from the children to this parent (if available);
- calls is the number of calls the parent function called the children.
Note that in recursive cycles, the total time % in the node is the same for the whole functions in the cycle, and there is no total time % figure in the edges inside the cycle, since such figure would make no sense.
The color of the nodes and edges varies according to the total time % value. In the default temperature-like color-map, functions where most time is spent (hot-spots) are marked as saturated red, and functions where little time is spent are marked as dark blue. Note that functions where negligible or no time is spent do not appear in the graph by default.
Frequently Asked Questions
How can I generate a complete call graph?
gprof2dot.py generates a partial call graph, excluding nodes and edges with little or no impact in the total computation time. If you want the full call graph then set a zero threshold for nodes and edges via the
--edge-thres options, as:
gprof2dot.py -n0 -e0
The node labels are too wide. How can I narrow them?
The node labels can get very wide when profiling C++ code, due to inclusion of scope, function arguments, and template arguments in demangled C++ function names.
If you do not need function and template arguments information, then pass the
--strip option to strip them.
If you want to keep all that information, or if the labels are still too wide, then you can pass the
--wrap, to wrap the labels. Note that because
dot does not wrap labels automatically the label margins will not be perfectly aligned.
Why there is no output, or it is all in the same color?
Likely, the total execution time is too short, so there is not enough precision in the profile to determine where time is being spent.
You can still force displaying the whole graph by setting a zero threshold for nodes and edges via the
--edge-thres options, as:
gprof2dot.py -n0 -e0
But to get meaningful results you will need to find a way to run the program for a longer time period (aggregate results from multiple runs).
Why don't the percentages add up?
You likely have an execution time too short, causing the round-off errors to be large.
See question above for ways to increase execution time.
Which options should I pass to gcc when compiling for profiling?
Options which are essential to produce suitable results are:
-g: produce debugging information
-fno-omit-frame-pointer: use the frame pointer (frame pointer usage is disabled by default in some architectures like x86_64 and for some optimization levels; it is impossible to walk the call stack without it)
If you're using gprof you will also need
-pg option, but nowadays you can get much better results with other profiling tools, most of which require no special code instrumentation when compiling.
You want the code you are profiling to be as close as possible as the code that you will be releasing. So you should include all options that you use in your release code, typically:
-O2: optimizations that do not involve a space-speed tradeoff
-DNDEBUG: disable debugging code in the standard library (such as the assert macro)
However many of the optimizations performed by gcc interfere with the accuracy/granularity of the profiling results. You should pass these options to disable those particular optimizations:
-fno-inline-functions: do not inline functions into their parents (otherwise the time spent on these functions will be attributed to the caller)
-fno-inline-functions-called-once: similar to above
-fno-optimize-sibling-calls: do not optimize sibling and tail recursive calls (otherwise tail calls may be attributed to the parent function)
If the granularity is still too low, you may pass these options to achieve finer granularity:
-fno-default-inline: do not make member functions inline by default merely because they are defined inside the class scope
-fno-inline: do not pay attention to the inline keyword Note however that with these last options the timings of functions called many times will be distorted due to the function call overhead. This is particularly true for typical C++ code which expects that these optimizations to be done for decent performance.
See the full list of gcc optimization options for more information.
See the wiki for external resources, including complementary/alternative tools.