This is a proposal for MPI 2.1, Ballot 4. This is a follow up to: Block lengths of zero in MPI Datatypes in http://www.cs.uiuc.edu/homes/wgropp/projects/parallel/MPI/mpi-errata/index.h... with mail discussion in http://www.cs.uiuc.edu/homes/wgropp/projects/parallel/MPI/mpi-errata/discuss... ___________________________________ Proposal: Add the following paragraph in MPI 1.1, Sect. 3.12, page 62, after line 2 (i.e., after ... "of the types defined by Typesig."): Most datatype constructors have replication count or block length arguments. Allowed values are nonnegative integers. If the value is zero, no elements are generated in the type map nor in the type signature for this entry in the argument list. MPI 1.1, Sect 3.12.1, MPI_TYPE_HINDEXED, page 67, line 22-24 read: IN count number of blocks – also number of entries in array_of_displacements and array_of_blocklengths (integer) but should read: IN count number of blocks – also number of entries in array_of_displacements and array_of_blocklengths (nonnegative integer) MPI 1.1, Sect 3.12.1, MPI_TYPE_STRUCT, page 68, line 19-22 read: IN count number of blocks (integer) – also number of entries in arrays array_of_types, array_of_displacements and array_of_blocklengths IN array of blocklength number of elements in each block (array of integer) but should read: IN count number of blocks (nonnegative integer) – also number of entries in arrays array_of_types, array_of_displacements and array_of_blocklengths IN array of blocklength number of elements in each block (array of nonnegative integer) MPI 2.0, Sect 4.14.1, MPI_TYPE_CREATE_HINDEXED, page 66, line 36-38 read: IN count number of blocks – also number of entries in array_of_displacements and array_of_blocklengths (integer) but should read: IN count number of blocks – also number of entries in array_of_displacements and array_of_blocklengths (nonnegative integer) MPI 2.0, Sect 4.14.1, MPI_TYPE_CREATE_STRUCT, page 67, line 14-18 read: IN count number of blocks (integer) – also number of entries in arrays array_of_types, array_of_displacements and array_of_blocklengths IN array of blocklength number of elements in each block (array of integer) but should read: IN count number of blocks (nonnegative integer) – also number of entries in arrays array_of_types, array_of_displacements and array_of_blocklengths IN array of blocklength number of elements in each block (array of nonnegative integer) ___________________________________ Rationale for this clarification and modification: The outcome of zero-count entries in the type map was not defined. For this, a clarification was needed. The interfaces of HINDEXED and STRUCT was inconsistent to the rest derived datatype routines. This was probably due to editing errors. A meaning of negative values was never defined not intended. Therefore, portable applications could not use negative values. These editing errors are fixed by this proposal. ___________________________________ I hope, I could catch all the problems Jesper mentioned in his mail. Best regards Rolf On Tue, 29 Jan 2008 15:02:10 +0100 Jesper Larsson Traeff <traff_at_it.neclab.eu> wrote:

Dear Rolf,

(thanks for all the work you are putting into this!)

On Tue, Jan 29, 2008 at 12:47:14PM +0100, Rolf Rabenseifner wrote:

...

Reply is off-line from the mailing list.

Dear Jesper and Bill,

I believe that this is not a topic for the one and final shot of MPI 2.1 Ballot 4 in the next (March 2008) meeting. Therefore, I would propose to handle it in MPI 2.2.

In the MPI-1 document (p.66, p.68) index and struct types are treated differently. For index, blocklength is an array of non-negative integers (p.66, line 12), for struct, just an array of integers (p. 68, line 22). Same inconsistency for the new MPI-2 types, p.66 and p.67. I believe that this should at least be made consistent.

I would also be in favor of a small remark that 0 blocklengths are allowed, with the remarks that these are (of course) not included in the type map.

Btw, I will be coming to the meeting in March

best regards

Jesper

...

Jesper, you have not been at Jan meeting. Therefore a short background. We decided, that MPI 2.1 is mainly the combining of the MPI 1.1 and 2.0 documents to one MPI 2.1 standard. Additionally the existing MPI 1.1 errata, and the MPI-2 electonically voted Ballots 1 & 2, and the Jan. meeting Ballot 3 and the March meeting Ballot 4 are included. Topics that do not get final official reading in these meetings are forwarded to MPI 2.2.

Therefore I try to have mainly items in Ballot 3 and 4 that do not need longer discussions and that are mainly obvious, i.e., that can be shown on one or two slides together with needed background infos.

Implication for me: I do not put your item on my MPI 2.1 agenda.

Okay?

Best regards Rolf

On Mon, 23 Apr 2007 17:10:35 +0200 Jesper Larsson Traeff <traff_at_[hidden]> wrote:

...

(ps: also sent to mpi-core_at_[hidden] - did anybody get it from there?)

Dear MPI forum,

an old discussion (July-September 2001) concluded that blocklengths of value zero are allowed in MPI indexed and struct types. Reading the standard, both text and specification (p.66ff), such "blocks" generate no entries in the corresponding type maps, at least this seems to be the idea - and implementations like mpich2 follow this interpretation, while the original MPICH for instance did not, and maybe with some right?

The question is what a derived datatype is intended to specify. If it specifies just a type map, then clearly blocks of length zero should simply be treated as not being there, and the resulting type map alone defines the extent, lower and upper bounds of the datatype. But if a derived datatype is intended as specifying a layout in memory, zero length blocks could possibly have another meaning (as markers of locations in the layout), and in addition to the type map affect the extent, lower and upper bound of the datatype.

A concrete example

displacements(1)=1 displacements(2)=0 lengths(1)=1 lengths(2)=0

call mpi_type_indexed(2,lengths,displacements,MPI_INTEGER,newtype,code) call mpi_type_commit(newtype,code)

In the first interpretation (datatypes specify type map) the above has extent 4 (assuming MPI_INTEGER is 4 bytes) and lower bound 4

In the second interpretation (datatypes specify layout) the above would have lower bound 0 and extent 8

The point is that in the first interpretation (which is the one at least implicitly prescribed by the standard) the layout that was possibly intended by the user ("something starting at displacement soandso" - irrespective of whether there is actual data at displacement soandso) can be quite arbitrarily collapsed. In an application where datatypes are generated automatically, it might be a high burden on the user to take care of the special cases arising by some blocklenghts just happening to be zero (either he would have to resize, or to put in explcit MPI_LB/MPI_UB markers).

A point could therefore be made for treating blocks of length zero still as "markers" in the specification of a layout - as did for instance the original MPICH implementation. Some changes in the text would be necessary to put this interpretation on a firm foundation. That is not the intention here, but only to see if there are opinions on this issue? A suggestion would be to add a short paragraph stating that zero blocklengths are allowed, and what their effects and side-effects are.

Jesper

Minor clarifications:

p.63, l.39, p.66, l.12, and p.67, l.25 have "nonnegative integer" blocklengths, but p. 68, l.22 only has "integer". To be consistent P. 68, l.22 should also be "nonnegative integer" (negative blocklenghts don't make sense)

(as already pointed out by Steven Huss-Lederman) The specification p.66, l.46ff and similarly for MPI_Type_hindexed and MPI_Type_struct makes little sense (well, is plain wrong) when some B[i] is actually zero. First, l.46 forces some elements into the type map [ (type_0,disp_0+D[0]) ,,,] that shouldn't be there (blocks of length zero shouln't give rise to any elements in the type map), and also l.48 has no real meaning when B[0]. Better would be to add two lines after l.48 giving the general case:

(type_0,disp_0+D[i]*ex),...,(type_{n-1},disp_{n-1}+D[i]*ex),..., (type_0,disp_0+(D[i]+B[i]-1)*ex,...,(type_{n-1},disp_{n-1}+(D[i]+B[i]-1)*ex,

with the proviso that this is empty if B[i]==0 (or even more formally precise/correct)

Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner_at_[hidden] 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: Allmandring 30)

Dr. Rolf Rabenseifner . . . . . . . . . .. email rabenseifner_at_[hidden] 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: Allmandring 30)