Hi Mikael, On 12 April 2018 at 16:53, Mikael Öhman <[email protected]> wrote:
> Hi everyone, > > It seems to be that there is a consensus that moving the path extension to > iccifort would be the solution for the core issue here? > I.e. > 1. Move the path extension to iccifort > Yes, this is what we do. > 2. Have the intel module load iccifort instead of icc and ifort > independently > We also do this BUT you must also have icc and ifort included as dependencies otherwise things currently fail. I can't remember the details of why but I know that there is an implicit requirement for icc/ifort to be present in the intel easyconfig. > > I don't think hiding icc or ifort would be required as long as they don't > have the path extension. > I think this could be introduced into the standard HMNS without harming > backwards compatibility at all. > You can't introduce a change into the intel easyconfig that is just for HMNS, we would have to branch off for that. The real solution would be to bundle icc/ifort at install time, I really can't see a good reason to keep them separate (after X years there's still no toolchain that mixes Intel C with a fortran compiler from another source). With that approach iccifort becomes a compiler rather than a toolchain and all our troubles are over. We'd have to vote on that one though... Alan > > Existing modules will remain for those users who are currently loading > ifort + icc + impi. Newly installed toolchain versions will start giving > correct dependency information. > For those who wish to fix older versions, a module-only reinstall of the > intel toolchain should suffice. > > Best regards, Mikael > > > On Thu, Apr 12, 2018 at 4:35 PM, Alan O'Cais <[email protected]> > wrote: > >> Sorry guys, I can't believe I didn't remember this (given that I wrote >> it) but our custom MNS fixes this problem: >> https://github.com/easybuilders/JSC/blob/master/Custom_MNS/ >> 2017b/custom_hierarchical_mns.py >> It ignores path extensions for icc/ifort (which are hidden for us anyway) >> and instead does the path extension for iccifort (which it also renames to >> Intel). It also puts the MPI implementations into a separate directory so >> that we can nicely present the tree. >> >> Alan >> >> On 12 April 2018 at 15:47, Alan O'Cais <[email protected]> wrote: >> >>> Yeah it is sorry, I was writing without checking! >>> >>> On 12 April 2018 at 15:44, Mikael Öhman <[email protected]> wrote: >>> >>>> Hi Alan, >>>> >>>> However, I don't understand how you can run into this problem. Normally >>>>> for any package the toolchain components are included as dependencies in >>>>> the final module: loading Caffe would also mean loading the missing ifort >>>>> module. I guess you must have a configuration that does not include the >>>>> toolchain deps in the final module? Removing this option would also solve >>>>> the problem (but you probably decided on that setting for a reason). >>>>> >>>> >>>> Isn't this the default behavior with HMNS? >>>> I suppose the rationale is that with HMNS you'll have to load the >>>> compilers first to have access to the modules in the first place (though, >>>> clearly, there is a mismatch between the modulepath and dependencies in >>>> this case). >>>> >>>> Best regards, Mikael >>>> >>>> >>>> >>> >>> >>> -- >>> Dr. Alan O'Cais >>> E-CAM Software Manager >>> Juelich Supercomputing Centre >>> Forschungszentrum Juelich GmbH >>> 52425 Juelich, Germany >>> >>> Phone: +49 2461 61 5213 >>> Fax: +49 2461 61 6656 >>> E-mail: [email protected] >>> WWW: http://www.fz-juelich.de/ias/jsc/EN >>> >> >> >> >> -- >> Dr. Alan O'Cais >> E-CAM Software Manager >> Juelich Supercomputing Centre >> Forschungszentrum Juelich GmbH >> 52425 Juelich, Germany >> >> Phone: +49 2461 61 5213 >> Fax: +49 2461 61 6656 >> E-mail: [email protected] >> WWW: http://www.fz-juelich.de/ias/jsc/EN >> >> >> ------------------------------------------------------------ >> ------------------------------------ >> ------------------------------------------------------------ >> ------------------------------------ >> Forschungszentrum Juelich GmbH >> 52425 Juelich >> Sitz der Gesellschaft: Juelich >> Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 >> Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher >> Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender), >> Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt, >> Prof. Dr. Sebastian M. Schmidt >> ------------------------------------------------------------ >> ------------------------------------ >> ------------------------------------------------------------ >> ------------------------------------ >> >> > -- Dr. Alan O'Cais E-CAM Software Manager Juelich Supercomputing Centre Forschungszentrum Juelich GmbH 52425 Juelich, Germany Phone: +49 2461 61 5213 Fax: +49 2461 61 6656 E-mail: [email protected] WWW: http://www.fz-juelich.de/ias/jsc/EN
