spack: can't build hdf5 using intel-mpi
Expected Result
hdf5 installed with ^intel-mpi
Actual Result
[sajid@xrmlite ~]$ spack install hdf5+mpi ^intel-mpi %intel@18.0.3
==> intel-mpi is already installed in /home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/intel-mpi-2018.2.199-2k65vtkd3pfhdoq3ds5p23utfcr7qv7o
==> zlib is already installed in /home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/zlib-1.2.11-vonxam4cye6oarnfuv4tviycduvx53ma
==> Installing hdf5
==> Using cached archive: /home/sajid/packages/spack/var/spack/cache/hdf5/hdf5-1.10.2.tar.gz
==> Staging archive: /home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2.tar.gz
==> Created stage in /home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt
==> No patches needed for hdf5
==> Building hdf5 [AutotoolsPackage]
==> Executing phase: 'autoreconf'
==> Executing phase: 'configure'
==> Error: ProcessError: Command exited with status 77:
'/home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2/configure' '--prefix=/home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt' '--enable-unsupported' '--disable-threadsafe' '--disable-cxx' '--disable-hl' '--disable-fortran' '--without-szlib' '--enable-build-mode=production' '--enable-shared' 'CFLAGS=-fPIC' 'CXXFLAGS=-fPIC' 'FCFLAGS=-fPIC' '--enable-parallel' 'CC=/home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/intel-mpi-2018.2.199-2k65vtkd3pfhdoq3ds5p23utfcr7qv7o/compilers_and_libraries/linux/mpi/intel64/bin/mpiicc' '--with-zlib=/home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/zlib-1.2.11-vonxam4cye6oarnfuv4tviycduvx53ma'
2 errors found in build log:
30 checking for config /home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2/config/site-specific/host-phys.northwestern.edu... no
31 checking for config /home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2/config/site-specific/host-northwestern.edu... no
32 checking for config /home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2/config/site-specific/host-edu... no
33 checking build mode... production
34 checking for gcc... /home/sajid/packages/spack/opt/spack/linux-centos7-x86_64/intel-18.0.3/intel-mpi-2018.2.199-2k65vtkd3pfhdoq3ds5p23utfcr7qv7o/compilers_and_libraries/linux/mpi/i
ntel64/bin/mpiicc
35 checking whether the C compiler works... no
>> 36 configure: error: in `/tmp/sajid/spack-stage/spack-stage-tdyo12k_/hdf5-1.10.2':
>> 37 configure: error: C compiler cannot create executables
38 See `config.log' for more details
See build log for details:
/home/sajid/packages/spack/var/spack/stage/hdf5-1.10.2-x5u6v4jqn252hr6rdt5ljjrxfgjq2klt/hdf5-1.10.2/spack-build.out
Information on your system
This includes:
- linux-centos7-x86_64, intel@18.0.3
About this issue
- Original URL
- State: closed
- Created 8 years ago
- Comments: 29 (29 by maintainers)
@s-sajid-ali You may have interference between how you integrated the Intel-18.0.3 compilers for Spack and the
intel-mpipackage.Since 18.0.3 is neither in #7469 (I was on travel), nor in
develop, I take it you saw this issue without that PR applied.FWIW, I get (of course with #7469 applied):
I’ll integrate 18.0.3 and will revisit.
@s-sajid-ali To address the issue in code may need some work – I filed #8632 for it.
For you to merely get going, you could try two avenues:
Use an alternative
intel18modulefile with MPI excised:MODULEPATHthat you control.I_MPI_ROOT*/mpiin it.Forego
modules:incompilers.yamland use full paths only.Abstractly, this may work, but be warnded that when I tried that in practice, HDF5 would compile only if I also provided a
cflags:setting.@s-sajid-ali
Bingo! – Looks like you integrated the composer edition for Spack with a
modules: [intel18]tag, which is a reasonable thing to do. The module setsI_MPI_ROOTto suit the MPI runtime it provides. That might be tolerable, but not if the compiler module gets loaded after Spack has setI_MPI_ROOTto suitintel-mpi, which is just what happens:So far, so good.
That pulls the rug out from under
intel-mpi. 💣… and then things go awry. 💥
Given that Spack demands full paths for compilers anyway, one might think that
modules:could be left empty or just select agcc. Alas, one would be wrong:Outside of Spack, I’d load compiler stuff first, then MPI. I have to believe that there are reasons why Spack does it the other way around. I have to think about the best angle of attack.
@s-sajid-ali It appears your
intel-mpipackage was installed whilespackwas checked out on another branch.intel-prefixesbranch version. If differences are indeed due to a branch change, that will be pretty obvious.intel-mpiwhilespackis on theintel-prefixesbranch. This may affect dependents – I trust you know how do deal with that.hdf5.