FS#28581 - [hdf5] OpenMPI support
Attached to Project:
Arch Linux
Opened by Simone (junki.gnu) - Wednesday, 22 February 2012, 09:53 GMT
Last edited by Ronald van Haren (pressh) - Monday, 07 January 2013, 20:56 GMT
Opened by Simone (junki.gnu) - Wednesday, 22 February 2012, 09:53 GMT
Last edited by Ronald van Haren (pressh) - Monday, 07 January 2013, 20:56 GMT
|
Details
Description:
HDF5 is compiled without MPI support, and this introduce some difficulties when parallel scientific computing frameworks (such as Trilinos) are compiled. The main problem, on the other hand, is that hdf5 with mpi enabled is not thread-safe. What do you think about that? |
This task depends upon
Closed by Ronald van Haren (pressh)
Monday, 07 January 2013, 20:56 GMT
Reason for closing: Implemented
Additional comments about closing: hdf5-openmpi-1.8.10-1
Monday, 07 January 2013, 20:56 GMT
Reason for closing: Implemented
Additional comments about closing: hdf5-openmpi-1.8.10-1
I don't know if serial and parallel version can live in peace together ... probably not.
I'm glad to help if needed!
FS#25098) as these are also not compatible with thread-safe.enabling 'unsupported' flag. I don't know if it's safe enough.
No problem with Fortran.
There's a big issue that prevents compilation: mpicc and fakeroot
don't like each other: https://bbs.archlinux.org/viewtopic.php?id=117493.
The suggested solution doesn't apply here, because make install calls
mpicc as a linker (through libtool).
Please find enclosed a first not-working draft of PKGBUILD. I'll keep trying.
applied to openmpi source.
Indeed, fakeroot mpicc -show still fails (segfault) with repo
version.
After recompiling openmpi with this patch, hdf5 compilation runs
smoothly *if* fortran is not enabled. Otherwise I obtain:
FCLD libhdf5_fortran.la
/usr/bin/ld: cannot find -l-L/usr/lib/gcc/x86_64-unknown-linux-gnu/4.6.2
collect2: ld returned 1 exit status
Probably there's an easy way to prevent this, but as far as I see
from fortran/src/Makefile there's an extra -rpath at line 646.
Resuming:
- openmpi patch: have I to file a bug?
- hdf5-openmpi compiles fine *without* fortran interface, fails otherwise;
- paraview keeps working if hdf5 is replaced by hdf5-openmpi.
Last version of PKGBUILD, *without* Fortran (see above).
Is it possible also to change hdf5 package in order to match this new
structure? I mean name from hdf5 to hdf5-serial, and a "provides=('hdf5')" and
"replaces=('hdf5')" into PKGBUILD.
How do we procede?
Thanks!