Introduction to MetaPRL Profiling

There are two ways to get the profiling information for MetaPRL - native code profiling and bytecode profiling. Native code profiling would give the amount of time spent in and inside a function, number of calls for each function and for each pair of functions - f1 and f2 - how many times f1 called f2 and how much time was spent in f2 when it was called from f1. The bytecode profiling would give the number of calls for each function and for each branching operator (such as if, match, try ... with) it would give information on how many times each branch was taken. Currently bytecode profiling information would be available only for files that are not pre-processed (although we hope to fix it someday) while in native code everything (including the standard OCaml libraries and even the garbage collection) would be profiled. The bytecode profiling should work on any system capable of compiling and running MetaPRL. The native code profiling should work on any Linux system that uses glibc 2 library. It may also work on other Linux systems and Digital Unix systems.

Native Code Profiling

In order to profile some code, you have to be able to call that code from the MetaPRL top loop. In order to do that you have to either declare you functions as topval in one of the prlc-compiled modules or to add it to the commands variable in the editor/ml/

After you've made all the necessary modifications to the code, set the NATIVE_PROFILING_ENABLED variable in mk/config and run omake. After the code was compiled:

  1. Start native-code version of MetaPRL: "editor/ml/mpopt"
  2. Do all necessary preparation steps
  3. Reset the profiling counters, so that the MetaPRL startup code and your preparation code does not get included into the profiling information: "restart_gmon ();;"
  4. Run the code you want to profile
  5. Dump the profiling information: "stop_gmon ();;". This should create a gmon.out file in the editor/ml directory
  6. Do the cleanup, if necessary
  7. Exit MetaPRL

Now you can produce the actual profiling information by running "gprof mp.opt". Warning: this can produce lots of output (several megabytes) and you may want to redirect the output.

The output of gprof would contain two parts - "flat profile" and "call graph". It would also include some comments that would help you understand its contents.

Bytecode Profiling

This is currently broken since ocamlprof does not know how to deal with preprocessed files. However, this may give you some idea of what is going on in case you want to do bytecode profiling by-hands.

  1. Set the appropriate flags in mk/config and run omake
  2. Start a bytecode version of MetaPRL: editor/ml/mptop
  3. Do all necessary preparation steps
  4. At this point you may want to try to reset all profiling counters to zero to avoid including startup and preparation into the profiling information. I am not sure how to do that, you probably need to go through the
    Profiling.counters: (string * (string * int array)) list ref
    and reset all the arrays to 0.
  5. Run the code you want to profile
  6. Exit MetaPRL

At this point you should have ocamlprof.dump file in your editor/ml directory. Now you can run "ocamlprof <file you want to profile>.ml". That would give you the contents on the original .ml file with all the branches marked with the number. Note - if you want to profile the preprocessed file (most of the MetaPRL files are preprocessed), you have to call ocamlprof on the file created by the preprocessor, not on the original file.