Profiling Code

GProf: non intrusive overall timing.

Official doc with information about how to interpret the output: http://google-perftools.googlecode.com/svn/trunk/doc/cpuprofile.html

Short documentation:

Install google performance tools:

Ubuntu

sudo apt-get install google-perftools graphviz

OSX

brew install google-perftools graphviz

###CPU Profiling Then when you run your code, you need to define CPUPROFILE and LD_PRELOAD

Ubuntu

CPUPROFILE=/tmp/my_executable.prof LD_PRELOAD=/usr/lib/libprofiler.so.0 ./my_executable

OSX

CPUPROFILE=/tmp/my_executable.prof DYLD_INSERT_LIBRARIES=/usr/local/Cellar/google-perftools/2.1/lib/libprofiler.dylib ./my_executable

Finally, generate the call-graph pdf:

Ubuntu

google-pprof --pdf my_executable /tmp/my_executable.prof > my_executable_profiling.pdf

OSX

pprof --pdf my_executable /tmp/my_executable.prof > my_executable_profiling.pdf

The call graph then looks like:

---

---

Other useful variables:

CPUPROFILE_FREQUENCY=x

where the default x = 100 (value is in Hz).

sm::timing: detailed but intrusive timing.

// Pull in the sm_timing definitions
#include <sm_timing/Timer.h>

//Enable timing.
typedef sm::timing::Timer Timer;

//Disable timing by switching the typedef.
typedef sm::timing::DummyTimer Timer;

After adding this definition you can now time specific code segments by adding a timer call:

Timer expensive_operation_timer("ExpensiveOperation");
DoExpensiveOperation();
expensive_operation_timer.Stop();

When you want to print out the timing information you call:

sm::timing::Timing::Print(std::cout);
//or similarly
LOG(INFO) << sm::timing::Timing::Print();

You will get a table containing the following information:

SM Timing
---------
//a                 b c               d                 e             f           g           
ExpensiveOperation  5 00:00:09:00.12 (00:00:08:56.67 +- 00:00:00.15) [00:00:08.00 00:00:09.000]

Where the columns denote:

a) Name of the timer.
b) Number of calls.
c) Overall time of all calls.
d) Mean time of a call.
e) Std-dev of the times.
f) Min time of a call.
g) Max time of a call.

Memory profiling

Then when you run your code, you need to define CPUPROFILE and LD_PRELOAD

Ubuntu

HEAPPROFILE=/tmp/mybin.hprof LD_PRELOAD=/usr/lib/libtcmalloc.so.4 ./my_executable

OSX

HEAPPROFILE=/tmp/mybin.hprof DYLD_INSERT_LIBRARIES=/usr/local/Cellar/google-perftools/2.1/lib/libtcmalloc.dylib ./my_executable

Finally, generate the call-graph pdf:

Ubuntu

google-pprof --pdf my_executable /tmp/profile.0001.heap > my_executable_heap_profiling.pdf

OSX

pprof --pdf my_executable /tmp/profile.0001.heap > my_executable_heap_profiling.pdf

Depending on your settings you will get many *.heap files that are created during the lifetime of your application.