[Pvfs2-users] The unexpected results of running MPI-IO on PVFS2

[Robert Latham]

Hi Xuechen.  I think both your messages are related so I"m going to
resond to them together below:


On Fri, Nov 16, 2007 at 08:46:08PM -0500, xuechen zhang wrote:
> I just configure my cluster according to the instruction of quick
> start.  And everything goes well until I run my first MPI-IO code.
> I got the unexpected results which are listed below.

Let's see what the debugging tools suggest:
- what does 'pvfs2-ping' say?  

(you might need to fix up your /etc/fstab entry.. pvfs2-ping will tell
you for sure)

- do 'pvfs2-ls' and 'pvfs2-cp' work for you?

> [xuechen at cluster9 exp1]$ mpiexec -np 1 sample
...
>         MPI_File_open(MPI_COMM_WORLD,"/mnt/pvfs2/matrix",MPI_MODE_RDWR|MPI_MODE_CREATE,MPI_INFO_NULL,&fh);

Do you have the file system mounted yet?  You don't have to if you
are using MPI-IO (I run tests that way quite often), but you'll have
to add the 'pvfs2:' prefix to your file name if you don't have the
kernel interface up yet.

On Mon, Nov 19, 2007 at 02:40:08PM -0500, xuechen zhang wrote:
> I exactly did the instructions of PVFS2-install quick start. However,
> when I use MPI-IO to access PVFS2, it still can not find the mount
> point. Below is what I got after running test code "simple".
> 
> [xuechen at cluster9 exp1]$ mpiexec -n 1 simple -fname /mnt/pvfs2/matrix

It's pretty important to check error codes: 'simple' does not, though
it really should (I'll fix that right now).  Your test code doesn't
either.  I suspect in both cases the MPI_File_open call is failing. 

> Can anyone help me find the reason?

Can you try a different test?  How about 'noncontig_coll2', which
*does* check the return value of every MPI call. 

I'm happy to see another user of MPI-IO: I hope we can get you back on
track.

==rob

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