Re: Octave profile timing

[David Grundberg]

On 04/13/2010 06:27 PM, Anne Ikonen wrote:
> Hi,
> I was looking for MATLAB timing profile from OCTAVE, but didn't find.
> Then I made a plan to measure OCTAVE timings. Several alternatives 
> were found:
> 1) Inserting
>      t(linenum, callcounter) = clock();
>     timelines in the code. This works, but isn't automatic, so big 
> files cause lots of additional work done.
> 2) Inserting
>      t(linenum, callcounter) = clock();
>     timelines in the code AUTOMATICALLY before actually calling the 
> scripts or functions.
>     Then you get:
>        myfunction.m  ----> modified myfunction.m including timelines.
>     This also works, but needs a facility to maintain timelines in 
> m-files. An easy automatic code looks for suitable lines in the m-file 
> and inserts a timeline everywhere.
>     However, this modifies m-files, which isn't very good thing in the 
> long run.
> 3) Could simulate OCTAVE engine and call eval for every octave code 
> fraction with timing measurement:
>      octaveline   ---->       eval ( 'timestart; octaveline ; timestop;')
>     this also works, but simulation can become rather complicated 
> trying to do all the same things as octave engine does.
> 4) Could dispatch octave engine "eval"-functionality to include timing 
> start and stop around executed code.
>     How this can be done? eval is a builtin function in octave. And 
> timing results need to be collected also.
> 5) Can find out average times for every octave action during eval and 
> report results as estimations of algorithm times.
>     This is not exact, but gives a good view of timings in an algorithm.
>     This works, but isn't automatic, so big files cause lots of 
> additional work done. And exceptional timings are not found this way.
> 6) Could try to use external facilities to measure timing points to 
> octave engine proceedings. Works and shows code location timings, but 
> since code location contents are not exactly known and can even vary, 
> isn't the best possible way to measure octave timings.
> Is MATLAB profile on/off available for octave somewhere outside octave 
> pages? It wasn't mentioned in octave manuals.
> Sincerely
> Jari-Pekka Ikonen
> email: address@hidden 
> <mhtml:%7B346A1EDF-6639-4BF9-A047-A83A328AABDF%7Dmid://00000362/%21x-usc:mailto:address@hidden>


Octave doesn't have a profiler yet, but I think no one would protest if 
there were one.

While writing m-code and shell scripts can give some results, I want 
Octave specialized to do profiling. I want statement-per-statement 
measurements. I want per-function measurements. I think the best way is 
to change the core Octave so that it can profile m-code on its own 
merit. A rigorous way to implement a profiler is by extending or 
subclassing the parse tree evaluator (written in C++). Someone needs to 
be interested in this, take on this project and make a prototype 
implementation. That's all it takes. (I realize this is on the help 
list, but anyway.)

I'm sorry I can't give a more helpful answer.

hth,
David