H5Part development and discussion

["Howison, Mark" <mhowison AT brown.edu>]

Hi Sergey,

I was able to recreate your error on two different systems I tried: my MacBook Pro with openmpi 1.5 and franklin at NERSC with Cray MPT (both with H5Part 1.6.2, HDF5 1.8.5-patch1). I wonder if it could be a bug in the SGI MPT stack, especially since the debug output you have points to ADIOI_Free? Is there anyway to open a bug report with SGI?

Thanks

Mark

On Fri, Oct 15, 2010 at 2:31 PM, Howison, Mark <mhowison AT brown.edu> wrote:

Hi Sergey, thanks for reported this bug, and the previous one with the testf program. I will take a look at fixing these this weekend. Mark

On Wed, Oct 6, 2010 at 4:43 AM, Sergey Litvinov <slitvinov AT gmail.com> wrote:

I am using fortran API of H5Part on HLRB-II supercomputer in Germany.
http://www.lrz.de/services/compute/hlrb/

I found that H5Part fails if I have more then one processor with zero particles.

I can reproduce the error with a test case. Please find in the attachments

testfzero.F90: a test case I compile. (One can put
it in place of test/testf.F90 or modify test/Makefile.am
to have it as a new test case)

I compile  hdf5-1.8.5-patch1 with debug enabled and set environment variable
export HDF5_DEBUG="trace"
config.log.hdf5: logfile of hdf5 configuration

H5Part-1.6.2
config.log.H5part: logfile of H5Part configuration

np.2: is an output of mpirun -prefix "%g: " -np 2 ./testf  (OK)
np.3: is an output of mpirun -prefix "%g: " -np 3 ./testf  (fails)

I use
> mpirun --version
MPI: libxmpi.so 'SGI MPT 1.26  09/04/10 08:45:46'

> mpif90
mpif90 for LRZ wrapper for SGI MPT
ifort (IFORT) 10.1 20090817

> mpicc
mpicc for LRZ wrapper for SGI MPT
icc (ICC) 10.1 20090817