Re: [Caml-list] Distributed computing libraries

[jean-marc alliot <jean-marc.alliot@irit.fr>]

Le 06/06/2012 18:53, Thomas Braibant a écrit :
> Hi list,
>
> There exists several different "distributed computing libraries" in OCaml:
>
> - Functory http://functory.lri.fr/
> - JoCaml http://jocaml.inria.fr/ (not really a library, though)
> - Nproc https://github.com/MyLifeLabs/nproc
> - OCaml for multicore http://www.algo-prog.info/ocmc/web/ (not a library either)
> - Parmap https://gitorious.org/parmap/parmap
> - Plasma Map/Reduce http://plasma.camlcity.org/plasma/
>
> And I am pretty sure that this list is not exhaustive, looking at
> http://caml.inria.fr/cgi-bin/hump.en.cgi?sort=0&browse=77
>
> Maybe it would be interesting for the community if someone could sum
> up the pros and cons of each of this "libraries", and maybe give some
> information about their status (still in development, mature, etc)?
>
> With best regards,
> Thomas Braibant
>
I was one of the two authors of the ocamlpvm interface. It works well 
and has been used for large "industrial" projects without any particular 
problems. These bindings are especially interesting when you need simple 
communications between different processes, and you also need to be able 
to launch processes by hand. The PVM bindings are stable and robust, PVM 
works for multi-core and  network parallelism but PVM is also slowly dying.

We have also been using lately the MPI bindings written by Xavier Leroy. 
The Ocaml bindings+MPI are an extremely efficient way to write parallel 
applications in the MPI paradigm. We compared it to a "home-made" shared 
memory system for exchanging objects between ocaml processes, and MPI is 
much faster (we were using MPICH2 MPI implementation), with, as a bonus, 
all the MPI goodies (such as transparent network parallelism). Well, 
that's a lesson to learn: leave parallelism implementation to the 
professionals.
We had to add one or two functions to Xavier's bindings to do the job 
(the MPI bindings are not complete, and if I remember correctly the 
iprobe function was not implemented). The modification is trivial.
If there are people interested, I could probably dig up the code.

Jean-Marc