[mpich2-dev] MPI_COMM_WORLD failed when using pio-bench on PVFS

[Rob Latham]

On Wed, Apr 01, 2009 at 02:23:59PM +0800, 汪迪 wrote:
> Hi, Thakur
> 
> I install MPICH2 by default configuration, and I install PVFS as
> root on that MPI installation, and again I install MPICH2, having
> "--with-pvfs=pvfs-route" option. I verify my PVFS installation and
> it works fine, and once I activate PVFS, the mpd would start
> automatically backend.  I verify my installation these two days, it
> works fine on any mpi programs including the cpi in the examples
> directory, except pio-bench and noncontig workbench.  pio-bench
> seems work fine if not running on mpi, and access a common file in
> linux file system. Even if it runs on mpich2, if the number of
> process is 1 it also works fine, so does noncontig by the way. The
> only failure occurs when the number of process is more than 1 (and
> more than 2 for noncontig).

Hi 

It sounds like you are doing everything right.  I suspect that you
have found a bug in pio-bench.  That test is quite old and has not
seen a lot of developer attention over the last few years.  

- If you would like to benchmark PVFS I/O performance, I'd like to
  suggest IOR (http://www.ior.sf.net).  

- If you are interested in metadata performance, any number of UNIX
  benchmarks will suffice, but we have lately used 'metarates'
  (http://www.cisl.ucar.edu/css/software/metarates/).   

==rob

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