[MPICH] MPI IO doubts

[Robert Latham]

On Thu, Jan 18, 2007 at 09:54:10AM -0400, Luiz Mendes wrote:
> 
> Inside item 3.9.1, the text says that MPI does not specifies how an
> implementation should physically create files. Then it could be a
> linear sequence of bytes. And after says that a file could be create
> as a single file or stripped trough local disks.
> 
> The first doubt is, in NFS divide a MPI File between disks wouldnt
> be possible?

With NFS, you will have multiple (parallel) clients talking to a
single server.  The NFS server could be serving files that are, for
example, hosted on a RAID volume.  It sounds like you are asking about
something along the lines of PVFS.
 
> The second doubt, Files created in MPI way could be operate by systems file
> commands ok? but to see informations inside this file we have to use mpi
> program, isnt it?

While it's possible to create a file only readable by an MPI program,
there are lots of examples of MPI programs creating files that are
read by other tools.  Parallel-NetCDF and HDF5, for example, both use
MPI-IO to create files, but also have a wealth of serial (non-mpi)
tools to look at the data.  This type of file is called
"self-describing", and can be very appealing if you have to exchange
data with collaborators.

==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B