[petsc-dev] Deprecated MPI functions

Barry Smith bsmith at mcs.anl.gov
Thu Feb 5 13:22:00 CST 2015


  I have no problem with someone making a branch that goes through the code and replaces all deprecated versions with their replacements and then submits it to the code review process (which, of course, means putting the code in next and then putting it into master :-)). I don't have any particular interest in doing this non-productive work myself.

   Barry

> On Feb 5, 2015, at 1:05 PM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Folks, these warnings are really noisy.  It has been 18 years since
> MPI-2 was published.  Who does not have an MPI-2.0 implementation
> available?
> 
> If we have to pay the cost of maintaining compatibility with MPI-1 when
> most users have MPI-3 implementations, I'd like to know who specifically
> is holding us up so that I can publicly shame them.
> 
> Jed Brown <jed at jedbrown.org> writes:
> 
>> Lisandro Dalcin <dalcinl at gmail.com> writes:
>>> Why not just use a configure test for MPI_VERSION to #define
>>> PETSC_MPI_VERSION, 
>> 
>> Could just use MPI_VERSION.
>> 
>>> and then if PETSC_MPI_VERSION<2 #definine the new functions as the old
>>> ones? Do you expect a MPI implementation advertising itself as MPI-2
>>> but not providing the new functions for attribute access?
>> 
>> Some implementations provided some or all of the new functions before
>> advertising themselves as MPI_VERSION=2.  There are probably others that
>> advertise as MPI_VERSION=2 but are incomplete, though they're not
>> supposed to do this.
>> 
>>>> [1] Something is wrong with the process when 17 years after a
>>>> standardization that involved the relevant vendors, we still cannot rely
>>>> on systems having upgraded.
>>> 
>>> I totally agree. One way of make pressure for the upgrades is to just
>>> stop supporting them :-), of making the support more difficult to get.
>> 
>> Yes, but for the same reason that PETSc works with a 10-year old Python
>> and a C compiler that has been abandoned for 15 years (MS), we have
>> historically tried to work with very old/obscure MPI implementations.




More information about the petsc-dev mailing list