A word of caution regarding Intel MPI 2019: They changed a lot of things under the hood, and we have seen lots of issues in our systems on relatively large jobs (1.5K+ MPI processes). Basically most collective algorithms don't make it through. We have seen that on 2 different InfiniBand systems, I am unsure about OmniPath, or about the underlying cause (maybe libfabric issues?), but with Intel MPI 2018 we haven't seen them.
Best, Damian On 07.01.19, 16:37, "[email protected] on behalf of Kenneth Hoste" <[email protected] on behalf of [email protected]> wrote: Dear EasyBuilders, By tradition at the start of the year, I have started looking at updating the 'foss' and 'intel' common toolchains, currently for the 2019a update. The plan is to include these in the upcoming EasyBuild v3.8.1 release, which I hope to release in a couple of weeks. Current proposals are: * foss/2019a: (see https://github.com/easybuilders/easybuild-easyconfigs/pull/7371) - GCC 8.2.0 + binutils 2.31.1 [LATEST for both] - OpenMPI 3.1.3 (latest 3.x) - OpenBLAS 0.3.5 [LATEST] + ScaLAPACK 2.0.2 [LATEST] - FFTW 3.3.8 [LATEST] => There also is OpenMPI 4.0.0, but we generally stay away from new major version of a toolchain component for the common toolchains. Does anyone have a detailed view on how OpenMPI v4.0.0 compares to v3.1.3, and whether or not we should consider v4.0.0? * intel/2019a (see https://github.com/easybuilders/easybuild-easyconfigs/pull/7372) - Intel C/C++/Fortran compilers 2019.1.144 [LATEST] (a.k.a. 2019 update 1, a.k.a. 19.0.1.144 according to "icc -V") on top of GCC 8.2.0 + binutils 2.31.1 - Intel MPI 2019.1.144 (a.k.a. 2019 update 1) [LATEST] - Intel MKL 2019.1.144 (a.k.a. 2019 update 1) [LATEST] => Any positive/negative experiences with the latest version of the Intel tools, or any remarks that may be relevant to the version we pick for intel/2019a? I'm currently testing these foss/2019a and intel/2019a proposals with a bunch of existing easyconfigs using the */2018b equivalent. A couple of minor compilation problems popped up when compiling with GCC 8.2.0 (for HDF5, NASM, ...), but they were resolved trivially by simply updating to the latest version of these dependencies. The 2019a toolchain definitions will also be discussed during the EasyBuild conf call this Wednesday (Jan 9th). regards, Kenneth ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------ 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 ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------
