Hi Kenneth, I took your suggestion, tweaked the super(PicGompi, self) line to be super(OpenMPI, self), and included appropriate imports. I can build software with my custom toolchain after doing this. I've opened https://github.com/hpcugent/easybuild-framework/issues/1903 to see if we can come up with a long-term solution.
Thanks for the help! Tim Slauson On Thu, Sep 8, 2016 at 12:50 AM, Kenneth Hoste <[email protected]> wrote: > Hi Tim, > > OK, I didn't take into account that the toolchain support for OpenMPI > (implemented in easybuild/toolchains/mpi/openmpi.py, which is where the > crash is occuring) may be depending on MPI_MODULE_NAME being non-empty... > > Try including this in your toolchain definition as well (indented properly > to be part of the PicGompi class): > > def _set_mpi_compiler_variables(self): > """Define MPI wrapper commands.""" > > # version-dependent, see http://www.open-mpi.org/faq/? > category=mpi-apps#override-wrappers-after-v1.0 > # assumption here is that OpenMPI version being used is recent > enough (>= 1.7) to provide mpifort > self.MPI_COMPILER_MPIF77 = 'mpifort' > self.MPI_COMPILER_MPIF90 = 'mpifort' > self.MPI_COMPILER_MPIFC = 'mpifort' > > # this needs to be done first, otherwise e.g., CC is set to MPICC > if the usempi toolchain option is enabled > for var, _ in COMPILER_VARIABLES: > self.variables.nappend('OMPI_%s' % var, > str(self.variables[var].get_first()), var_class=CommandFlagList) > > super(PicGompi, self)._set_mpi_compiler_variables() > > > This will effectively override the implementation of > _set_mpi_compiler_variables in openmpi.py, which should hopefully solve > your problem... > Do note that I'm assuming that the OpenMPI version provided by your > external module is sufficiently recent... > > The underlying issue is indeed that get_software_version is not taking > into account that the module could be an external module. > Can you try and reproduce the problem you had with using your external > module name in MPI_MODULE_NAME? > I'll open an issue to try and get this fixed then... > > > regards, > > Kenneth > > > > On 08/09/16 02:27, Tim Slauson wrote: > > Hi Kenneth, > > Thanks for the MPI_MODULE_NAME suggestion. I had earlier tried setting > that to my external module name, but ran into different errors and just > gave up. > > When I set MPI_MODULE_NAME to an empty list, the OpenMPI toolchain module > hits an IndexError when trying to retrieve a software version (see > traceback below). Since I'm specifying that information in an external > module metadata file, shouldn't the software version be taken from there? > > Thanks, > Tim > > > ERROR: Traceback (most recent call last): > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.egg/easybuild/main.py", > line 114, in build_and_install_software > (ec_res['success'], app_log, err) = build_and_install_one(ec, init_env) > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/framework/easyblock.py", line 2363, in build_and_install_one > result = app.run_all_steps(run_test_cases=run_test_cases) > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/framework/easyblock.py", line 2279, in run_all_steps > self.run_step(step_name, step_methods) > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/framework/easyblock.py", line 2158, in run_step > step_method(self)() > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/framework/easyblock.py", line 1554, in prepare_step > self.toolchain.prepare(self.cfg['onlytcmod'], silent=self.silent) > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/tools/toolchain/toolchain.py", line 608, in prepare > self.set_variables() > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/tools/toolchain/compiler.py", line 154, in set_variables > super(Compiler, self).set_variables() > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/tools/toolchain/mpi.py", line 94, in set_variables > self._set_mpi_compiler_variables() > File "/import/usrlocal/easybuild/2.8.2/software/EasyBuild/2.8.2/l > ib/python2.6/site-packages/easybuild_framework-2.8.2-py2.6.e > gg/easybuild/toolchains/mpi/openmpi.py", line 62, in > _set_mpi_compiler_variables > ompi_ver = self.get_software_version(self.MPI_MODULE_NAME)[0] > IndexError: list index out of range > > > On Wed, Sep 7, 2016 at 3:32 AM, Kenneth Hoste <[email protected]> > wrote: > >> Hi Tim, >> >> Welcome to the wonderful world of EasyBuild! >> >> >> On 07/09/16 01:29, Tim Slauson wrote: >> >>> Hello EasyBuilders, >>> >>> I'm trying out EasyBuild for the first time (version 2.8.2), and am >>> having trouble hooking in the vendor-provided MPI modules on one of our >>> clusters. The modules all follow the same naming convention of >>> name/compiler/ver (e.g. openmpi/gnu/1.10.2, mvapich2-psm/intel/2.1, ...). >>> After reading about EXTERNAL_MODULE, I adapted the gompi-2016.06.eb >>> toolchain easyconfig: >>> >>> 3c3 >>> < name = 'gompi' >>> --- >>> > name = 'pic-gompi' >>> 17c17 >>> < ('OpenMPI', '1.10.3', '', ('GCC', gccver)), >>> --- >>> > ('openmpi/gnu/1.10.2', EXTERNAL_MODULE), >>> >>> >>> And wrote external module metadata: >>> >>> [openmpi/gnu/1.10.2] >>> name = OpenMPI >>> version = 1.10.2 >>> prefix = MPI_HOME >>> >>> >>> And defined a new toolchain: >>> >>> from easybuild.toolchains.gompi import Gompi >>> >>> class PicGompi(Gompi): >>> """PIC compiler toolchain with GNU compilers and OpenMPI""" >>> NAME = 'pic-gompi' >>> >>> >>> This allows me to build a toolchain. But when I run this: >>> >>> eb FFTW-3.3.4-gompi-2016.06.eb --try-toolchain=pic-gompi,2016.06 >>> >>> I get this error: >>> >>> ERROR: Build of /tmp/eb-HkZNza/tweaked_easycon >>> figs/FFTW-3.3.4-pic-gompi-2016.06.eb failed (err: "build failed (first >>> 300 chars): List of toolchain dependency modules and toolchain definition >>> do not match (found ['GCC/5.4.0-2.26', 'openmpi/gnu/1.10.2'] vs expected >>> set(['GCC', 'OpenMPI']))") >>> >>> I get similar errors when repeating this process for the other >>> vendor-provided libraries. Can somebody please point me in the right >>> direction? >>> >> >> You're hitting this because EasyBuild verifies the module for the >> toolchain with the definition of the toolchain. >> This is done using module names, see for example >> https://github.com/hpcugent/easybuild-framework/blob/master/ >> easybuild/toolchains/mpi/openmpi.py#L44 . >> >> You should be able to fix your problem by including this in your >> pic-gompi toolchain definition (i.e. reset it to the default): >> >> MPI_MODULE_NAME = [] >> >> This should only affect the toolchain definition verification mechanism, >> nothing else. >> >> >> regards, >> >> Kenneth >> > > > > -- > Tim Slauson > GI/RCS > University of Alaska Fairbanks > > > -- Tim Slauson GI/RCS University of Alaska Fairbanks
