[MPICH] problems with writing in mpi, slaves processes doesnt write

[Luiz Mendes]

Hi Rob, Geoff, all

I have got it work. The problem was located at file permission. Always i
have to change the file owner to nobody and nogroup, because the other PC
seems to allow writes only with this latter owner.

Now, when i change the owner, it works, and didnt log messages about io
permission denied anymore.

Info argument, the fourth argument of the function MPI_File_open, could be
used to pass MPI ROMIO hints?

I have been obtaining low values about bandwith (15MB/s - local and non
local processes) compared to only local processes situations (350 MB/s). Is
it normal? Would be SSH as rpc caller the reason for this? The previous
measures were obtained with execution using 2 processes and using mpirun -np
2 iotest.

The network here is 100mbits ethernet, with local disc at 7200rpm and the
other with 5400rpm both IDE interface.

Thanks for attention, i appreciate a lot, furthermore the emails of this
list are helping me a lot.

Luiz Mendes

2007/1/17, Robert Latham <robl at mcs.anl.gov>:
>
> On Tue, Jan 16, 2007 at 08:54:00AM -0400, Luiz Mendes wrote:
> > I have two PCS here. One of them, the master node and the another one
> that
> > is a slave.
> >
> > I installed mpich 1.2.7p1 and configured the two machines in
> > Machine.LINUXfile.
> >
> > Well, but when i test io operations with 2 processes, the slave process
> > doesnt write. Why?
> >
> > i have installed mpich 1.2.7p1 on the 2 pcs in the same folder,
> > /usr/share/MPI-1, and after i export /usr/share/MPI-1 from the master
> node
> > to the slave node, mouting it in /usr/share/MPI-1 in slave PC.
> >
> > What should i do to make it works?
>
> Nothing sounds wrong from your description.  If your test program is
> small enough, you should post it and maybe we can either reproduce the
> bug on our end, or point out something you may have overlooked in your
> MPI code.
>
> > And, do you know something about use pvfs2+mpich 1.2.7p1?
>
> The PVFS support in MPICH-1.2.7p1 is fair, but not as good as what you
> can find in MPICH2-1.0.5.  The version in MPICH2 contains many bug
> fixes and performance improvements.
>
> ==rob
>
> --
> Rob Latham
> Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
> Argonne National Lab, IL USA                 B29D F333 664A 4280 315B
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20070117/60cf5f33/attachment.htm>