[mpich-discuss] Funky MPI_File_write_at Parameters

[William Gropp]

The offset is likely 8 bytes, not 4 (it depends on the MPI  
implementation and file system environment), and the write buffer must  
be non-null unless a datatype with an absolute address is used.  My  
guess is that the offset is 8 bytes, and that is throwing off the  
interpretation of the traceback.

Bill

On Apr 11, 2011, at 7:45 PM, Ellis H. Wilson III wrote:

> Hi all,
>
> Back with another odd-ball tracing question.  When I trace IOR using  
> the
> following parameters:
>
> IOR -a MPIIO -b 4m -t 2m -i 1 -o pvfs2:/tmp/pvfs2client/test.dat
>
> I get what I feel are very odd tracing parameters shown below for the
> MPI_File_write_at function:
>
> MPI_File_write_at(0x9ae9ab0, 4194304, NULL, -1211088896, 0x200000,
> 0x4c00010d
>
> Now there are 6 parameters, which I expected from the specification,  
> but
> it seems to me like they might be out of order.  P1 seems to be the  
> file
> handle, P2 seems to be the offset, P3 indicates my lack of a buffer,
> which I believe to be fine, but P4 is the real oddity as the count
> argument should be nonnegative and here we have -1211088896.  0x200000
> corresponds to 2097152 decimal so I would think that makes more  
> sense as
> a count (-t means transfers of 2megs, which is dead on for that) and I
> know the last parameter shown should have been my datatype and sitting
> in position 5, but it is in 6.  Status object is nowhere to be found  
> as
> it ends there (unless status object somehow got switched to position  
> 4,
> which is my guess currently).  These parameters are similarly weird  
> for
> MPI_File_read_at.
>
> I am tracing MPI calls using LANL-Trace and I've looked through the  
> code
> for LANL-Trace and IOR and everything seems to check out except the
> difference between how IOR internally handles offset and count (both
> long long int) and how mpich handles offset and count (int).  I don't
> think this is the issue however as I've added some quick casts to IOR
> and the problem persists.  I am truly lost on why this is happening,  
> but
> then again perhaps I am misunderstanding something fundamental about  
> this.
>
> If anyone can confirm that at the very least, I am right in thinking
> this looks weird it would be very helpful.
>
> Thanks,
>
> ellis
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss

William Gropp
Director, Parallel Computing Institute
Deputy Director for Research
Institute for Advanced Computing Applications and Technologies
Paul and Cynthia Saylor Professor of Computer Science
University of Illinois Urbana-Champaign