[MPI3 Fortran] (j3.2006) (SC22WG5.3909) [ukfortran] [MPI3 Fortran] MPI non-blocking transfers

[N.M. Maclaren]

On Jan 23 2009, Bill Long wrote:
>
>Since a WAIT statement could occur in a subprogram (at least ones the 
>compiler does not know about, such as ones not intrinsic or inlined) 
>then it is part of the required implementation that operations on a 
>variable with the asynchronous attribute not be moved across such a 
>call.  Included in the class of calls that block code motion would be 
>those to the MPI routines.  Having the asynchronous attribute in the 
>caller scope effectively solves the code motion problem.

Agreed.

> Having it on 
>the corresponding dummy probably solves it as well.

What corresponding dummy?  There may not BE one!

>Jim pointed out that asynchronous may not be that useful in the end.  I 
>think this is an important point.  As noted above, asynchronous can 
>solve the code motion problem.  It is not clear what you mean by "solve" 
>in regard to the other main problem - copy-in/copy-out of actual 
>arguments.   What the asynchronous attribute will do is cause a program 
>that would otherwise have to include copy-in/copy-out code at the call 
>to the MPI_Ixxx routine to, instead, get a hard error at compile time 
>based on a constraint violation.  It does not make the program "work", 
>which is probably somewhere on the user's agenda.

That is correct.  However, to make arbitrary broken programs work is
beyond any mortal's ability!  The best that we can hope for is to minimise
the amount of rewriting needed.

> To get a working 
>program, there are two possible solutions that I see:
>
>1) Require that the user always supply a contiguous actual argument 
>corresponding to the buffer dummy  (in which case it seems to make more 
>sense for THAT to be the requirement on the user in the first place), or
>
>2) Provide a Fortran interface than can handle non-contiguous sections 
>by having assumed-shape corresponding dummies.  This solution off-loads 
>the bulk of the work to the MPI folks. But is does seem more robust.

Neither of those solutions will work.  Sorry.

Consider the situation where procedure Fred allocates a buffer, and then
passes it down through two call chains:  Fred->Joe->Bert->Alf->MPI_Isend and
Fred->Pete->Nick->Greg->MPI_Wait.  The buffer is then used in Greg.  The
requirement is that neither of the call chains performs a copy-in/copy-out
at any stage.  It is perfectly legal to pass a non-contiguous section to
an assumed size dummy (which is then contiguous); it doesn't matter WHAT the
interface to MPI_Isend is - that assumed size array is a copy of the buffer
that was allocated in Fred.

The point is that the above problem has ALREADY been addressed for Fortran
asynchronous I/O.  If the politics are such that some vendors will refuse
to accept that solution for genuinely asynchronous transfers, then we have
a big problem.  But, make no mistake, it's political and not technical.


Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QH, England.
Email:  nmm1 at cam.ac.uk
Tel.:  +44 1223 334761    Fax:  +44 1223 334679