Dear All, I am wondering if this error is associated with the bug reported by intel with glibc. https://software.intel.com/en-us/articles/intel-compiler-not-compatible-with-glibc-224-9-and-newer
Code compiled with newer versions of the Intel compilers may result in unexpected numerical values when used under newer versions of glibc such as the one in CentOS 7.4, due to a bug in the Intel compiler. Best wishes, Victor H. Rusu, PhD CSCS Swiss National Supercomputing Centre Via Trevano 131 6900 Lugano Switzerland On 20.11.17, 09:30, "[email protected] on behalf of Åke Sandgren" <[email protected] on behalf of [email protected]> wrote: On 11/20/2017 09:15 AM, Kenneth Hoste wrote: > Hi Åke, > > Can we somehow integrate these FFTW tests in the sanity check that > EasyBuild runs for FFTW? Those tests are already part of the FFTW build sanity check. That's why it fails to build. > > regards, > > Kenneth > > On 20/11/2017 07:56, Åke Sandgren wrote: >> 2018.0.128 miscompiles FFTW on KNL (i.e. AVX512) and i suspect it will >> miscompile FFTW on skylake too. >> Since that is a fairly well tested code and all the previous versions >> (and 2018.1.163) has no problems with FFTW I would advise people to stay >> away from it. >> >> And more so since 2018.0 and older also have problems with the glibc >> from CentOS 7.4, which 2018.1 has fixed. >> >> >> (Damian wanted some more details so here goes) >> >> As for FFTW, Intel 2018.0.128, intermittently segfaults on the single >> precision non-threaded basic test, and always produces incorrect >> numerical results on the single precision 2-thread basic test. >> >> This is with the default -O2 -xHost flags that the easybuild recipe >> uses. Also verified with a manual build. >> >> Just -O2 alone doesn't show any problems, but then it would be fairly >> useless since FFTW is heavy on the AVX512 side. >> >> And since it causes the numerical errors (and the intermittent >> segfaults) I would classify it as a brown-paperbag-release. I.e. don't >> ever use it. >> >> I never bothered testing double precision. >> >> On 11/20/2017 07:23 AM, Alvarez, Damian wrote: >>> Hey Ake, >>> >>> Could you elaborate on that? I mean, I do expect the compiler to have >>> bugs, but we haven’t seen anything particularly bad so far, just >>> these 2 things: >>> >>> -The issue with glibc in CentOS 7.4 >>> -An internal compiler error when optimizing certain loops in PETSc >>> with complex numbers. >>> >>> Is there any particular bug we should be aware of? >>> >>> Damian >>> >>> On 19/11/17 00:23, "[email protected] on behalf of Åke >>> Sandgren" <[email protected] on behalf of >>> [email protected]> wrote: >>> >>> I strongly advise against using the intel 2018.0.128 compiler. >>> It has bugs! >>> >>> 2018.1.163 works much better. >>> >>> On 11/18/2017 10:13 AM, Kenneth Hoste wrote: >>> > Hi all, >>> > >>> > On 17/11/2017 11:29, Holger Angenent wrote: >>> >> On 13.11.2017 13:28, Joachim Hein wrote: >>> >>> Hi Holger >>> >> Hi Joachim and Kenneth. >>> >>> >>> >>> Thanks for this. >>> >>> >>> >>> Kenneth, what are chances to get an emergency EB 3.4.1b or >>> EB 3.4.2 >>> >>> that contains perhaps intel 2018 only. We are running slurm >>> and >>> >>> currently rebuilding the software with intel 2018 or moving >>> to a >>> >>> newer libc as suggested by Alexandre from FZJ seem our best >>> bets to >>> >>> get out of the issues. >>> >>> >>> >>> If we go for a rebuilding marathon, it would be a lot >>> easier, if >>> >>> intel 2018 would sit in EB and not in a development branch >>> or PR. >>> >> By the way, we are speaking about this PR: >>> >> https://github.com/easybuilders/easybuild-easyconfigs/pull/5129 >>> > >>> > I opened a PR for the 2018 update 1 versions of Intel >>> compilers, Intel >>> > MPI & Intel MKL a couple of days ago, see >>> > >>> https://github.com/easybuilders/easybuild-easyconfigs/pull/5345; it's >>> > currently awaiting review/merging. >>> > >>> > I'm up for doing a new EasyBuild release in the coming days, >>> if I can >>> > find the time for it that is, things are quite crazy the last >>> couple of >>> > weeks... >>> > >>> > >>> > From the Intel article [1], there is a workaround that already >>> works >>> > with intel/2017b (which includes the 2017 update 4 compilers), >>> i.e. to >>> > build with -fPIC, or in EasyBuild terms, enable the 'pic' >>> toolchain >>> > option in the easyconfig file. >>> > This workaround is also required with the initial 2018 release >>> > (intel/2018.00), but no longer necessary with 2018 update 1 >>> > (intel/2018.01, see PR #5129). >>> > >>> > The -fPIC workaround has proven to circumvent this issue in a >>> recent >>> > pull request for ABINIT, see >>> > https://github.com/easybuilders/easybuild-easyconfigs/pull/5251. >>> > >>> > >>> > regards, >>> > >>> > Kenneth >>> >>> >>> >>> Best wishes >>> >>> Joachim >>> >> Best, >>> >> Holger >>> >>> >>> >>>> On 13 Nov 2017, at 01:17, Holger Angenent >>> >>>> <[email protected] >>> >>>> <mailto:[email protected]>> wrote: >>> >>>> >>> >>>> On 12.11.2017 20:59, Joachim Hein wrote: >>> >>>>> We got bitten by: >>> >>>>> >>> https://software.intel.com/en-us/articles/inconsistent-program-behavior-on-red-hat-enterprise-linux-74-if-compiled-with-intel >>> >>> >>>>> >>> >>>>> >>> >>>>> We are running CentOS 7.4. Many of our intel build apps >>> are not >>> >>>>> working any longer. >>> >>>>> >>> >>>>> Best wishes >>> >>>>> Joachim >>> >>>>> >>> >>>>> Sent from my nanoPad >>> >>>> >>> >>>> We are also aware of this issue. Since Intel 2018 should >>> solve this >>> >>>> bug, we are building as many modules as we can for the >>> intel-2018 >>> >>>> toolchain. (see for example >>> >>>> >>> https://github.com/easybuilders/easybuild-easyconfigs/pull/5291) As >>> >>>> far as I know, they will be part of the 3.5.0 EasyBuild >>> Release. >>> >>>> >>> >>>> The Travis builds are failing at the moment, but this >>> should only be >>> >>>> the case because intel-2018 is not in the official repo, >>> yet. We are >>> >>>> already using those modules in production. >>> >>>> >>> >>>> Best regards, >>> >>>> >>> >>>> Holger >>> >>>> >>> >>>> >>> >>> >>> >> >>> >> -- >>> >> Westfälische Wilhelms-Universität Münster (WWU) >>> >> Zentrum für Informationsverarbeitung (ZIV) >>> >> Röntgenstraße 7-13 >>> >> 48149 Münster >>> >> +49-(0)251-83 31569 >>> >> [email protected] >>> >> www.uni-muenster.de/ZIV >>> > >>> >>> -- >>> Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden >>> Internet: [email protected] Phone: +46 90 7866134 Fax: +46 >>> 90-580 14 >>> Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se >>> >>> >>> >>> >>> ------------------------------------------------------------------------------------------------ >>> >>> ------------------------------------------------------------------------------------------------ >>> >>> 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 >>> ------------------------------------------------------------------------------------------------ >>> >>> ------------------------------------------------------------------------------------------------ >>> >>> > -- Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden Internet: [email protected] Phone: +46 90 7866134 Fax: +46 90-580 14 Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se
