[mpiwg-rma] Changes to the RMA chapter

Rob Latham robl at mcs.anl.gov
Thu Mar 5 09:01:08 CST 2015



On 03/04/2015 08:35 PM, Jim Dinan wrote:
> I agree, no_locks seems to have a clear local semantic.  accumulate_ops
> ad accumulate_ordering need some thought from the WG; I think these were
> intended to be provided at all processes.

if you are going to call it an MPI_Info, then please follow the rules of 
MPI_Info objects already well established in the I/O chapter.  all 
processes must specify the same value for a given MPI_Info key.

RMA has "things that are like infos but don't follow info rules" they 
are called "asserts", right?

==rob

>
>   ~Jim.
>
> On Wed, Mar 4, 2015 at 6:26 PM, William Gropp <wgropp at illinois.edu
> <mailto:wgropp at illinois.edu>> wrote:
>
>     Actually, I think the wording is clear - no process will target that
>     particular window (not window object) with passive target
>     synchronization.  Thus, as noted in the description, no async
>     handing for passive RMA is needed at this process.  It says nothing
>     about whether this process can issue locks against some other
>     process’s window.
>
>     Bill
>
>     On Mar 4, 2015, at 5:17 PM, Balaji, Pavan <balaji at anl.gov
>     <mailto:balaji at anl.gov>> wrote:
>
>      >
>      > If the intention is that it can be local, then the info needs to
>     specify whether it means the process cannot issue locks or whether
>     other processes cannot issue locks on it.  Either way, some change
>     is required.  I agree this is not ticket-0, and will tell the Forum
>     that the RMA WG will put it on the list of things to discuss.  We
>     should do a WG telecon.  A number of issues are queued up for
>     discussion at this point.
>      >
>      >  -- Pavan
>      >
>      >> On Mar 4, 2015, at 5:09 PM, William Gropp <wgropp at illinois.edu
>     <mailto:wgropp at illinois.edu>> wrote:
>      >>
>      >> For the bigger one, yes, this is much harder and absolutely not
>     ticket-0.  no_locks is particularly tricky, because it applies to
>     the window, not the window object, and hence is defined as local.
>      >>
>      >> Bill
>      >>
>      >> On Mar 4, 2015, at 4:59 PM, Balaji, Pavan <balaji at anl.gov
>     <mailto:balaji at anl.gov>> wrote:
>      >>
>      >>>
>      >>> There are two questions below.  The formatting part (whether to
>     use [SAME]) is a smaller issue.  The bigger issue is whether we want
>     to say that all the remaining info keys no_locks, etc., should be
>     the same on all processes.
>      >>>
>      >>> I want to say that, but: (1) need the chapter committee
>     consensus, and (2) I'm not sure it's a ticket-0 change, though we
>     did make that change for same_size as ticket-0.
>      >>>
>      >>> -- Pavan
>      >>>
>      >>>> On Mar 4, 2015, at 4:56 PM, William Gropp <wgropp at illinois.edu
>     <mailto:wgropp at illinois.edu>> wrote:
>      >>>>
>      >>>> That’s an excellent idea - we should use the same language
>     throughout the standard, and the I/O chapter is probably the best
>     source.
>      >>>>
>      >>>> Bill
>      >>>>
>      >>>> On Mar 4, 2015, at 4:39 PM, Balaji, Pavan <balaji at anl.gov
>     <mailto:balaji at anl.gov>> wrote:
>      >>>>
>      >>>>>
>      >>>>> Nathan suggested to use the I/O chapter format where info
>     keys are annotated with "[SAME]" to indicate that all processes must
>     provide the same info key.  That is probably a ticket-0 change.  We
>     need to first agree on whether we want to do this.
>      >>>>>
>      >>>>> -- Pavan
>      >>>>>
>      >>>>>> On Mar 4, 2015, at 4:37 PM, Balaji, Pavan <balaji at anl.gov
>     <mailto:balaji at anl.gov>> wrote:
>      >>>>>>
>      >>>>>>
>      >>>>>> Folks at Forum complained about the addition to the
>     "same_size" info key which added the following phrase:
>      >>>>>>
>      >>>>>> "and that all processes have provided this info key with the
>     same value"
>      >>>>>>
>      >>>>>> (this change was not a part of any ticket, and was made by
>     the chapter committee)
>      >>>>>>
>      >>>>>> There were two concerns:
>      >>>>>>
>      >>>>>> 1. This is a backward incompatible change (as in,
>     applications were not required to give the same value earlier),
>     though folks understood why this is needed.  They were OK as long as
>     we add a Errata/Changelog item saying so.
>      >>>>>>
>      >>>>>> 2. The property that all processes should give the same info
>     key was true for other keys in that section as well (no_locks,
>     accumulate_ordering, accumulate_ops).  So the same sentence should
>     be added to them as well.
>      >>>>>>
>      >>>>>> The Forum wants to see a draft of these changes.  I'd like
>     to propose adding the same phrase ("and that all processes have
>     provided this info key with the same value") to the other info keys
>     as well.  What do folks think?  Is this something that we agree on
>     (technically, it is possible to have different info keys in the job,
>     but it probably doesn't make much sense).
>      >>>>>>
>      >>>>>> Thoughts?
>      >>>>>>
>      >>>>>> -- Pavan
>      >>>>>>
>      >>>>>> --
>      >>>>>> Pavan Balaji  ✉️
>      >>>>>> http://www.mcs.anl.gov/~balaji
>      >>>>>>
>      >>>>>> _______________________________________________
>      >>>>>> mpiwg-rma mailing list
>      >>>>>> mpiwg-rma at lists.mpi-forum.org
>     <mailto:mpiwg-rma at lists.mpi-forum.org>
>      >>>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>      >>>>>
>      >>>>> --
>      >>>>> Pavan Balaji  ✉️
>      >>>>> http://www.mcs.anl.gov/~balaji
>      >>>>>
>      >>>>> _______________________________________________
>      >>>>> mpiwg-rma mailing list
>      >>>>> mpiwg-rma at lists.mpi-forum.org
>     <mailto:mpiwg-rma at lists.mpi-forum.org>
>      >>>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>      >>>>
>      >>>> _______________________________________________
>      >>>> mpiwg-rma mailing list
>      >>>> mpiwg-rma at lists.mpi-forum.org
>     <mailto:mpiwg-rma at lists.mpi-forum.org>
>      >>>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>      >>>
>      >>> --
>      >>> Pavan Balaji  ✉️
>      >>> http://www.mcs.anl.gov/~balaji
>      >>>
>      >>> _______________________________________________
>      >>> mpiwg-rma mailing list
>      >>> mpiwg-rma at lists.mpi-forum.org
>     <mailto:mpiwg-rma at lists.mpi-forum.org>
>      >>> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>      >>
>      >> _______________________________________________
>      >> mpiwg-rma mailing list
>      >> mpiwg-rma at lists.mpi-forum.org <mailto:mpiwg-rma at lists.mpi-forum.org>
>      >> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>      >
>      > --
>      > Pavan Balaji  ✉️
>      > http://www.mcs.anl.gov/~balaji
>      >
>      > _______________________________________________
>      > mpiwg-rma mailing list
>      > mpiwg-rma at lists.mpi-forum.org <mailto:mpiwg-rma at lists.mpi-forum.org>
>      > http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>
>     _______________________________________________
>     mpiwg-rma mailing list
>     mpiwg-rma at lists.mpi-forum.org <mailto:mpiwg-rma at lists.mpi-forum.org>
>     http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>
>
>
>
> _______________________________________________
> mpiwg-rma mailing list
> mpiwg-rma at lists.mpi-forum.org
> http://lists.mpi-forum.org/mailman/listinfo.cgi/mpiwg-rma
>

-- 
Rob Latham
Mathematics and Computer Science Division
Argonne National Lab, IL USA



More information about the mpiwg-rma mailing list