Subject: Re: [Boost-mpi] No arg init
From: Jeremiah Willcock (jewillco_at_[hidden])
Date: 2012-11-29 11:23:28
On Thu, 29 Nov 2012, Alain O Miniussi wrote:
> Based on http://www.mpi-forum.org/docs/mpi-20-html/node43.htm#Node43.
> Only implementations conforming to MPI1.1 are allowed to require passing
> the actual values provided to main, MPI-2 does not.
> I don't know what kind of compatibility we want to maintain wrt
> MPI1,(afaic, stuff that is in the range 2.x-3.x would be fine) but that
> change would only concern a subset of MPI 1.1 implementations, and
> passing the arguments would still be allowed.
> But I guess the main question is: do we want to deal with 3 versions of
> the MPI standard ?
> Also, if Boost.MPI is presented as the successor of the C++ bindings,
> and if I am correct in assuming that the MPI::Init function always
> provided both versions, it would be weird to maintain a pre-existent
Another issue with MPI versions is that Boost.MPI currently uses functions
such as MPI_Address that have been removed in MPI 3.0. Is that something
that should be addressed in the future? I think the replacements for some
of them did not exist before MPI 2.0.
-- Jeremiah Willcock
> On Thu, 2012-11-29 at 08:33 -0600, Hal Finkel wrote:
>> I thought the system was arranged as it was because some MPI implementations pass special arguments from the runtime system to the MPI library via special command-line arguments. Passing the args to MPI_Init allows the library to intercept those arguments (and erase them from the arg list seen by the rest of the program). I don't know if any real implementations work this way, but we may want to be careful about discouraging passing the real arg variables to MPI_Init.
>> ----- Original Message -----
>>> From: "Matthias Troyer" <troyer_at_[hidden]>
>>> To: "alain miniussi" <alain.miniussi_at_oca.eu>, "Discussion of Boost.MPI development" <boost-mpi_at_[hidden]>
>>> Sent: Thursday, November 29, 2012 5:17:35 AM
>>> Subject: Re: [Boost-mpi] No arg init
>>> Hi Alain,
>>> You can indeed provide no arg initialization and just fake arguments
>>> for MPI-1.
>>> On Nov 29, 2012, at 12:13 PM, Alain O Miniussi
>>> <alain.miniussi_at_oca.eu> wrote:
>>>> While working on the thread issue, a question arises regarding the
>>>> arg initialization.
>>>> I understand that some MPI1.x might not allow it, but should
>>>> refect that choice ? Why can't we just pass dummy empty (argc,argv)
>>>> MPI_Init(_thread) since, after all, no specific values are required
>>>> This would simplify our API, and I do not see any down side.
>>>> Any though ?
>>>> Alain Miniussi
>>>> Responsable Tech. Centre de Calcul Haute Performance
>>>> Obs. de la CÃ´te d'Azur |Mont Gros: +33 4 92 00 30 09
>>>> BP 4229 |Sophia : +33 4 83 61 85 44
>>>> 06304 Nice Cedex 4 |https://crimson.oca.eu
>>>> Boost-mpi mailing list
>>> Boost-mpi mailing list
> Alain Miniussi
> Responsable Tech. Centre de Calcul Haute Performance
> Obs. de la CÃ´te d'Azur |Mont Gros: +33 4 92 00 30 09
> BP 4229 |Sophia : +33 4 83 61 85 44
> 06304 Nice Cedex 4 |https://crimson.oca.eu
> Boost-mpi mailing list
Boost-Commit list run by troyer at boostpro.com