[Mpi-comments] MPI_Dims_create(..., ndims=0, ...) semantic ?

Fri Nov 11 11:53:58 CST 2016

I’ll add something, but I disagree that this is a change - the standard is implicitly quite clear - ndim can be an integer, and 0 is an integer.  

Bill

William Gropp
Acting Director and Chief Scientist, NCSA
Director, Parallel Computing Institute
Thomas M. Siebel Chair in Computer Science
University of Illinois Urbana-Champaign

On Nov 10, 2016, at 2:52 AM, Rolf Rabenseifner <rabenseifner at hlrs.de> wrote:

> Yes I agree: Bill's solution is better, but it requires that the 
> MPI libraries may Need to be fixed, because then, ndimm==0 && nnodes==1
> is no error.
> Currently, it is undefined, i.e., a MPI library may handle it
> as an erronous input and may call the error handler. 
> I.e., this solution would require a change log entry that
> all MPI library developers check there code.
> 
> Proposal: I would add after the Example 7.1 on MPI-3.1 page 293 line 42
> 
>  If ndims is zero, then nnodes must be one, and the argument
>  dims will be ignored, i.e., there will be no output.
> 
> Changelog in "Changes in MPI-4.0":
> 
>  Section 7.5.2 on page 792.
>  MPI_DIMS_CREATE was extended to additionally allow a zero-dimensional 
>  Cartesian topology, i.e., input ndims==0 together with nnodes==1.
> 
> Bill, please can you provide the ticket/issue/... with probably 
> better wording ...
> 
> Rolf
> 
> ----- Original Message -----
>> From: "wgropp" <wgropp at illinois.edu>
>> To: "Comments on the MPI Standard" <mpi-comments at lists.mpi-forum.org>
>> Sent: Wednesday, November 9, 2016 7:04:44 PM
>> Subject: Re: [Mpi-comments] MPI_Dims_create(..., ndims=0, ...) semantic ?
> 
>> I agree with the analysis of the output but disagree with the conclusion - to
>> make the behavior consistent with that of MPI_Cart_create, ndim == 0 should be
>> permitted (with nnodes == 1) but there is, as noted, no output other than a
>> return value / error value set of MPI_SUCCESS. Otherwise the use of
>> MPI_Dims_create with MPI_Cart_create in the general case looks like
>> 
>> if (ndim > 0) MPI_Dims_create(…)
>> MPI_Cart_create(…)
>> 
>> while the intent of the changes for MPI_Cart_create was to allow
>> 
>> MPI_Dims_create(…)
>> MPI_Cart_create(…)
>> 
>> even if ndim == 0.
>> 
>> Bill
>> 
>> William Gropp
>> Acting Director and Chief Scientist, NCSA
>> Director, Parallel Computing Institute
>> Thomas M. Siebel Chair in Computer Science
>> University of Illinois Urbana-Champaign
>> 
>> 
>> 
>> 
>> 
>> On Nov 9, 2016, at 9:17 AM, Rolf Rabenseifner < [ mailto:rabenseifner at hlrs.de |
>> rabenseifner at hlrs.de ] > wrote:
>> 
>> 
>> 
>> We added all that the MPI_CART_... routines work fine with ndims==0.
>> 
>> We oversaw that MPI_DIMS_CREATE is also a candidate for ndims==0.
>> 
>> Reading the standard, I would say that MPI_DIMS_CREATE
>> - with ndims==0 can never have any output, because
>> the ndims Array has no element,
>> - with ndims==0 and nnodes != 1 is erronous
>> because a zero-dim-cart-topology has alwas exactly one node
>> (and erronous implies always undefined behavior,
>> i.e., the MPI library may catch or ignore this error),
>> - with ndims==0 and nnodes == 1 would be okay but would not
>> produce any output (because ndims has no entries),
>> - but compared to the other CART routines where ndims==0
>> behaviour is always explicitely specified, I would conclude
>> that MPI_DIMS_CREATE with ndims==0 has always undefined
>> behavior, i.e., that also ndims==0 and nnodes == 1 may be
>> treated by the MPI library as erronous.
>> 
>> Therefore, I would recommend to clarify the standard by
>> defining ndims of MPI_DIMS_CREATE as "positive integer"
>> instead of only "integer".
>> 
>> Rolf
>> 
>> ----- Original Message -----
>> 
>> 
>> From: "wgropp" < [ mailto:wgropp at illinois.edu | wgropp at illinois.edu ] >
>> To: "Comments on the MPI Standard" < [ mailto:mpi-comments at lists.mpi-forum.org |
>> mpi-comments at lists.mpi-forum.org ] >
>> Sent: Wednesday, November 9, 2016 3:18:10 PM
>> Subject: Re: [Mpi-comments] MPI_Dims_create(..., ndims=0, ...) semantic ?
>> 
>> 
>> 
>> The standard doesn’t specify this, though I recall us discussing this and IIRC
>> agreeing that ndims equal to zero was well-defined and not an error. And there
>> is nothing in the standard that suggests that it should be an error, just
>> nothing that says it shouldn’t.
>> 
>> Bill
>> 
>> William Gropp
>> Acting Director and Chief Scientist, NCSA
>> Director, Parallel Computing Institute
>> Thomas M. Siebel Chair in Computer Science
>> University of Illinois Urbana-Champaign
>> 
>> 
>> 
>> 
>> 
>> On Nov 8, 2016, at 11:40 PM, Gilles Gouaillardet < [ [ mailto:gilles at rist.or.jp
>> | mailto:gilles at rist.or.jp ] |
>> [ mailto:gilles at rist.or.jp | gilles at rist.or.jp ] ] > wrote:
>> 
>> 
>> 
>> Folks,
>> 
>> 
>> Does the standard specifiy the behavior of MPI_Dims_create() when the ndims
>> argument is zero ?
>> 
>> 
>> the standard only says negative input values of dims[i] are erroneous, but does
>> not say anything about ndims.
>> 
>> ndims is the number of Cartesian dimensions (integer), and though a negative
>> number likely looks erroneous
>> 
>> to most of us, a zero value is incorrect from a mathematical point of view (and
>> i wish i remembered that ...)
>> 
>> 
>> Could someone tell me how a zero value should be interpreted ?
>> 
>> (e.g. valid/invalid/up to the implementor ?)
>> 
>> 
>> FWIW, currently
>> 
>> in MPICH, this is a no-op.
>> 
>> In Open MPI, this is an error by default, but if parameter check is manually
>> disabled, then this is a no-op too.
>> 
>> 
>> Cheers,
>> 
>> 
>> Gilles
>> 
>> _______________________________________________
>> mpi-comments mailing list
>> [ [ mailto:mpi-comments at lists.mpi-forum.org |
>> mailto:mpi-comments at lists.mpi-forum.org ] | [
>> mailto:mpi-comments at lists.mpi-forum.org | mpi-comments at lists.mpi-forum.org ] ]
>> [ https://lists.mpi-forum.org/mailman/listinfo/mpi-comments |
>> https://lists.mpi-forum.org/mailman/listinfo/mpi-comments ]
>> 
>> 
>> _______________________________________________
>> mpi-comments mailing list
>> mpi-comments at lists.mpi-forum.org
>> https://lists.mpi-forum.org/mailman/listinfo/mpi-comments
>> 
>> --
>> Dr. Rolf Rabenseifner . . . . . . . . . .. email [ mailto:rabenseifner at hlrs.de |
>> rabenseifner at hlrs.de ]
>> High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530
>> University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
>> Head of Dpmt Parallel Computing . . . [ http://www.hlrs.de/people/rabenseifner |
>> www.hlrs.de/people/rabenseifner ]
>> Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)
>> _______________________________________________
>> mpi-comments mailing list
>> [ mailto:mpi-comments at lists.mpi-forum.org | mpi-comments at lists.mpi-forum.org ]
>> https://lists.mpi-forum.org/mailman/listinfo/mpi-comments
>> 
>> 
>> _______________________________________________
>> mpi-comments mailing list
>> mpi-comments at lists.mpi-forum.org
>> https://lists.mpi-forum.org/mailman/listinfo/mpi-comments
> 
> -- 
> Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner at hlrs.de
> High Performance Computing Center (HLRS) . phone ++49(0)711/685-65530
> University of Stuttgart . . . . . . . . .. fax ++49(0)711 / 685-65832
> Head of Dpmt Parallel Computing . . . www.hlrs.de/people/rabenseifner
> Nobelstr. 19, D-70550 Stuttgart, Germany . . . . (Office: Room 1.307)
> _______________________________________________
> mpi-comments mailing list
> mpi-comments at lists.mpi-forum.org
> https://lists.mpi-forum.org/mailman/listinfo/mpi-comments

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpi-forum.org/pipermail/mpi-comments/attachments/20161111/cd310abe/attachment-0001.html>
-------------- next part --------------
_______________________________________________
mpi-comments mailing list
mpi-comments at lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpi-comments