Boost logo

Boost-MPI :

Subject: [Boost-mpi] Develop VS master
From: Alain Miniussi (Alain.Miniussi_at_[hidden])
Date: 2017-07-14 14:35:10


Hi,

As you probably know, 1.64 is not building. That reflect the fact that
serialization/master broke backward compatibility (in the detail area, a
part Boost.MPI should not rely on in the first place, but that's another
story).

MPI/develop is ok in that respect, as a mater of fact develop seems to
pass its tests on all the platforms I have access to (linux with g++
icpc and intel MPIs).

I proposed https://github.com/boostorg/mpi/pull/46 in February which is
a merge from develop on master, hoping it would be in for 1.64. It is my
opinion that develop should be regularly merge on master after proper
testing, but that opinion is not shared by all the people with merge
authorization.

You can read the discussion in https://github.com/boostorg/mpi/pull/46
for details. Basically the argument against merging is that it would
break some antique platform, but I was unable to get the informations:
which platform, what breaks, compile error messages etc...

It's been suggested to only move changes to master through cherry
picking, which I'm against since I think it's a maintenance nghtmare
that waste contributor's time and efforts.

So right now, what I would like to see is a merge of develop to master
or a clear description of what to be fixed in develop preventing that.

The pro and cons I can think of:

PROS:

- Master is broken anyway
- All the test I was able to perform passed ok
- develop has some interesting features (cartesian communicator,
variable size versions of global operations to name some)

CONS:

- If I do the merge, I'll be both the guy who submitted the pull request
and the one validating the pool request, which is not good practice IMO.
- some test are failing in
http://www.boost.org/development/tests/develop/developer/mpi.html, but
the only messages I could get indicates a configuration error (some flag
is used that is not supported by the compiler).
- linked to the previous problem: I only have access to some platform
(linus, intel and gnu compilers, Intel and some OpenMPI). As you know,
the cross product of plateform/MPI/compiler version is...big.

So, I would like to have your input on that issue (maybe through
https://github.com/boostorg/mpi/pull/46, maybe not ?) and, very
important, if you could test develop on your platform and share the results.

Do you think we should try to get develop on master for the 1.65 ?

Regards,

Alain

On 14/07/2017 10:17, Alain Miniussi via Boost-mpi wrote:

> Please see the discussion in:
>
> https://github.com/boostorg/mpi/pull/46
>
> I'm going to send an email to the list to see what we can do.
>
> Regards
>
> Alain
>
> On 13/07/2017 20:20, Richard via Boost-mpi wrote:
>> Dear Boost MPI developers and maintainers,
>>
>> since none of the developers or maintainers commented on the ticket
>> within the last half a year, I would like to make sure you are aware of
>> this bug:
>> https://svn.boost.org/trac10/ticket/12723
>>
>> Could someone please see to it that this does not make it to the next
>> release?
>> It's really not difficult to fix and I think it's pretty bad this made
>> it to boost 1.64.
>>
>> But now it's even in 1.65.0 beta 1!!
>>
>> Why doesn't the simple fix in the develop branch of boost mpi get
>> merged?
>> (https://github.com/boostorg/mpi/commit/f5bdcc1ebfe954bb64835f2a0efd94471da42207)
>>
>>
>> I would be really grateful if boost 1.65.0 had a working boost mpi
>> again.
>>
>> Best,
>> Richard
>> _______________________________________________
>> Boost-mpi mailing list
>> Boost-mpi_at_[hidden]
>> https://lists.boost.org/mailman/listinfo.cgi/boost-mpi
>
> _______________________________________________
> Boost-mpi mailing list
> Boost-mpi_at_[hidden]
> https://lists.boost.org/mailman/listinfo.cgi/boost-mpi


Boost-Commit list run by troyer at boostpro.com