Caml performance

Caml performance

[Michael Hicks]

[ Malheureusement, je n'ecris pas bien le francais; alors, j'ecrivait (?) en
Anglais seulement ... ]

I've been trying to compile OCaml to be instrumented with gprof.  When I do
so, it's possible to get profile information for the ocamlrun executable
(the interpreter), but not for any of the additional libraries (like Unix
and Threads), presumably because they are "linked" in with the bytecode file
that's being executed, but do not reside in the ocamlrun executable itself.
Is there any way around this?  In particular, I'm trying to find out how
much time is spent in the thread scheduler, and gprof seemed like a good way
to find this out.

Secondly, I noticed that the interpreter was not compiled, by default, with
-O on.  Is there any reason for this?  By contrast, the threads library
scheduler does have -O turned on, so it makes me think that gcc might do
some optimizations that are not gc-safe.

Thanks!
Mike

-- 
Michael Hicks
Ph.D. Candidate, the University of Pennsylvania
http://www.cis.upenn.edu/~mwh            mailto://mwh@dsl.cis.upenn.edu
"Since my last report, this employee has reached rock bottom and shows
signs of starting to dig."  -- taken from actual performance report

Re: Caml performance

[Xavier Leroy]

> I've been trying to compile OCaml to be instrumented with gprof.  When I do
> so, it's possible to get profile information for the ocamlrun executable
> (the interpreter), but not for any of the additional libraries (like Unix
> and Threads), presumably because they are "linked" in with the bytecode file
> that's being executed, but do not reside in the ocamlrun executable itself.

Yes, in some sense.  If you use external libraries, a special-purpose
runtime system is created and linked with your program, and the
ocamlrun executable itself is not used at all.

> Is there any way around this?  In particular, I'm trying to find out how
> much time is spent in the thread scheduler, and gprof seemed like a good way
> to find this out.

A simple way to do this is to recompile the whole OCaml distribution
with the "-p" flag.  In config/Makefile, add "-p" to the variables
BYTECCCOMPOPTS and BYTECCLINKOPTS.  Then do make clean; make all; make
install.

For good measure, also use "ocamlc -custom -ccopt -p ..." when linking
your application, to make sure "-p" is passed to the C compiler when
linking the custom runtime system.

> Secondly, I noticed that the interpreter was not compiled, by default, with
> -O on.

That is not true.  The Makefile in byterun/ adds the "-O" option itself.
Maybe a better place for that option would be in config/Makefile
itself.  But at any rate the bytecode interpreter is compiled with -O.

> By contrast, the threads library
> scheduler does have -O turned on, so it makes me think that gcc might do
> some optimizations that are not gc-safe.

Oh no, we don't let gcc mess with the GC... The GC interface is very
explicitely written down in the C code of the interpreter.  Any
optimization that would mess with the GC interface would be
semantically incorrect in a very big way.

Best regards,

- Xavier Leroy