[gmx-users] Install GROMACS 2018 with CUDA features in dynamic linking way

Tue Dec 11 22:14:24 CET 2018

AFAIK the right way to control RPATH using cmake is:
https://cmake.org/cmake/help/v3.12/variable/CMAKE_SKIP_RPATH.html
no need to poke the binary.

If you still need to turn off static cudart linking the way to do that
is also via a CMake feature:
https://cmake.org/cmake/help/latest/module/FindCUDA.html
The default is static.

--
Szilárd
On Tue, Dec 11, 2018 at 10:45 AM Jaime Sierra <jaisiero at gmail.com> wrote:
>
> I'm trying to rewrite the RPATH because shared libraries paths used by
> GROMACS are hardcoded in the binary.
>
> ldd /nfs2/opt/APPS/x86_64/APPS/GROMACS/2016/CUDA/8.0/bin/gmx
> linux-vdso.so.1 =>  (0x00007ffddf1d3000)
> libgromacs.so.2 =>
> /nfs2/opt/APPS/x86_64/APPS/GROMACS/2016/CUDA/8.0/bin/../lib64/libgromacs.so.2
> (0x00007f0094b25000)
> libcudart.so.8.0 => not found
> libnvidia-ml.so.1 => /lib64/libnvidia-ml.so.1 (0x00007f0094500000)
> libz.so.1 => /lib64/libz.so.1 (0x00007f00942ea000)
> libdl.so.2 => /lib64/libdl.so.2 (0x00007f00940e5000)
> librt.so.1 => /lib64/librt.so.1 (0x00007f0093edd000)
> libpthread.so.0 => /lib64/libpthread.so.0 (0x00007f0093cc1000)
> libstdc++.so.6 => /lib64/libstdc++.so.6 (0x00007f00939b7000)
> libm.so.6 => /lib64/libm.so.6 (0x00007f00936b5000)
> libgomp.so.1 => /lib64/libgomp.so.1 (0x00007f009348f000)
> libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007f0093278000)
> libc.so.6 => /lib64/libc.so.6 (0x00007f0092eb7000)
> libcudart.so.8.0 => /nfs2/LIBS/x86_64/LIBS/CUDA/8.0/lib64/libcudart.so.8.0
> (0x00007f0092c50000)
> /lib64/ld-linux-x86-64.so.2 (0x00007f0097ad2000)
>
> ldd /nfs2/opt/APPS/x86_64/APPS/GROMACS/2018/CUDA/8.0/bin/gmx
> linux-vdso.so.1 =>  (0x00007fff27b8d000)
> libgromacs.so.3 =>
> /nfs2/opt/APPS/x86_64/APPS/GROMACS/2018/CUDA/8.0/bin/../lib64/libgromacs.so.3
> (0x00007fcb4aa3e000)
> libstdc++.so.6 => /lib64/libstdc++.so.6 (0x00007fcb4a71f000)
> libm.so.6 => /lib64/libm.so.6 (0x00007fcb4a41d000)
> libgomp.so.1 => /lib64/libgomp.so.1 (0x00007fcb4a1f7000)
> libgcc_s.so.1 => /lib64/libgcc_s.so.1 (0x00007fcb49fe0000)
> libpthread.so.0 => /lib64/libpthread.so.0 (0x00007fcb49dc4000)
> libc.so.6 => /lib64/libc.so.6 (0x00007fcb49a03000)
> libcudart.so.8.0 => /nfs2/LIBS/x86_64/LIBS/CUDA/8.0/lib64/libcudart.so.8.0
> (0x00007fcb4979c000)
> libcufft.so.8.0 => /nfs2/LIBS/x86_64/LIBS/CUDA/8.0/lib64/libcufft.so.8.0
> (0x00007fcb4094e000)
> libdl.so.2 => /lib64/libdl.so.2 (0x00007fcb40749000)
> librt.so.1 => /lib64/librt.so.1 (0x00007fcb40541000)
> libfftw3f.so.3 =>
> /nfs2/LIBS/x86_64/LIBS/FFTW/3.3.3/SINGLE/lib/libfftw3f.so.3
> (0x00007fcb401c8000)
> libmkl_intel_lp64.so =>
> /nfs2/LIBS/x86_64/LIBS/mkl/l_mkl_11.1.0.080/composer_xe_2013_sp1.0.080/mkl/lib/intel64/libmkl_intel_lp64.so
> (0x00007fcb3faa4000)
> libmkl_intel_thread.so =>
> /nfs2/LIBS/x86_64/LIBS/mkl/l_mkl_11.1.0.080/composer_xe_2013_sp1.0.080/mkl/lib/intel64/libmkl_intel_thread.so
> (0x00007fcb3ea0a000)
> libmkl_core.so =>
> /nfs2/LIBS/x86_64/LIBS/mkl/l_mkl_11.1.0.080/composer_xe_2013_sp1.0.080/mkl/lib/intel64/libmkl_core.so
> (0x00007fcb3d4dc000)
> libiomp5.so =>
> /nfs2/LIBS/x86_64/LIBS/mkl/l_mkl_11.1.0.080/lib/intel64/libiomp5.so
> (0x00007fcb3d1c2000)
> libmkl_gf_lp64.so =>
> /nfs2/LIBS/x86_64/LIBS/mkl/l_mkl_11.1.0.080/composer_xe_2013_sp1.0.080/mkl/lib/intel64/libmkl_gf_lp64.so
> (0x00007fcb3caa0000)
> /lib64/ld-linux-x86-64.so.2 (0x00007fcb4d785000)
>
> See the differences between the 2016 & 2018 version.
>
> I'm using Cmake 3.13.1.
>
> ~/cmake-3.13.1-Linux-x86_64/bin/cmake .. -DGMX_BUILD_OWN_FFTW=ON
> -DREGRESSIONTEST_DOWNLOAD=ON -DGMX_GPU=ON
> -DCUDA_TOOLKIT_ROOT_DIR=/nfs2/LIBS/x86_64/LIBS/CUDA/8.0
> -DCMAKE_INSTALL_PREFIX=/nfs2/opt/APPS/x86_64/APPS/GROMACS/2018/CUDA/8.0
> -DCUDA_USE_STATIC_CUDA_RUNTIME=OFF -DBUILD_SHARED_LIBS=ON
> -DCUDA_NVCC_FLAGS=--cudart=shared -DGMX_PREFER_STATIC_LIBS=OFF
> -DEXTRA_NVCCFLAGS=--cudart=shared
>
> I think I've tried almost everything.
>
> Regards.
>
> El lun., 10 dic. 2018 a las 16:09, Szilárd Páll (<pall.szilard at gmail.com>)
> escribió:
>
> > On Sat, Dec 8, 2018 at 10:00 PM Gmail <jaisiero at gmail.com> wrote:
> > >
> > > My mistake! It was a typo. Anyway, this is the result before executing
> > > the chrpath command:
> > >
> > > chrpath -l $APPS/GROMACS/2018/CUDA/8.0/bin/gmx
> > > $APPS/GROMACS/2018/CUDA/8.0/bin/gmx: RPATH=$ORIGIN/../lib64
> > >
> > > I'm suspicious that GROMACS 2018 is not being compiled using shared
> > > libraries, at least, for CUDA.
> >
> > First of all, what is the goal, why are you trying to manually rewrite
> > the binary RPATH?
> >
> > Well, if the binaries not linked against libcudart.so than it clearly
> > isn't (and the ldd output is a better way to confirm that -- a library
> > can be linked against gmx even without an RPATH being set).
> >
> > I have a vague memory that this may have been the default in CMake or
> > perhaps it changed at some point. What's your CMake version, perhaps
> > you're using an old CMake?
> >
> > >
> > > Jaime.
> > >
> > >
> > > On 8/12/18 21:39, Mark Abraham wrote:
> > > > Hi,
> > > >
> > > > Your final line doesn't match your CMAKE_INSTALL_PREFIX
> > > >
> > > > Mark
> > > >
> > > > On Sun., 9 Dec. 2018, 07:00 Jaime Sierra <jaisiero at gmail.com wrote:
> > > >
> > > >> Hi pall,
> > > >>
> > > >> thanks for your answer,
> > > >> I have my own "HOW_TO_INSTALL" guide like:
> > > >>
> > > >> $ wget ftp://ftp.gromacs.org/pub/gromacs/gromacs-5.1.4.tar.gz
> > > >> $ tar xzf gromacs-5.1.4.tar.gz
> > > >> $ cd gromacs-5.1.4.tar.gz
> > > >> $ mkdir build
> > > >> $ cd build
> > > >> $ export EXTRA_NVCCFLAGS=--cudart=shared
> > > >> $ export PATH=$APPS/CMAKE/2.8.12.2/bin/:$PATH
> > > >> $ cmake .. -DCMAKE_INSTALL_PREFIX=$APPS/GROMACS/5.1.4/CUDA8.0/GPU
> > > >> -DGMX_FFT_LIBRARY=fftw3 -DCMAKE_PREFIX_PATH=$LIBS/FFTW/3.3.3/SINGLE/
> > > >> -DGMX_GPU=ON -DCUDA_TOOLKIT_ROOT_DIR=$LIBS/CUDA/8.0/
> > > >> $ make -j $(nproc)
> > > >> $ make install
> > > >> $ chrpath -r '$ORIGIN/../lib64' $APPS/GROMACS/5.1.4/GPU/bin/gmx
> > > >>
> > > >> that works until GROMACS 2016, I couldn't make it work for GROMACS
> > 2018.
> > > >>
> > > >> Regards,
> > > >>
> > > >> Jaime.
> > > >>
> > > >> El vie., 7 dic. 2018 a las 15:49, Szilárd Páll (<
> > pall.szilard at gmail.com>)
> > > >> escribió:
> > > >>
> > > >>> Hi Jaime,
> > > >>>
> > > >>> Have you tried passing that variable to nvcc? Does it not work?
> > > >>>
> > > >>> Note that GROMACS makes up to a dozen of CUDA runtime calls (kernels
> > > >>> and transfers) per iteration with iteration times in the range of
> > > >>> milliseconds at longest and peak in the hundreds of nanoseconds and
> > > >>> the CPU needs to sync up every iteration with the GPU. Hence, I
> > > >>> suspect GROMACS may be a challenging use-case for rCUDA, but I'm very
> > > >>> interested in your observations and benchmarks results when you have
> > > >>> some.
> > > >>>
> > > >>> Cheers,
> > > >>> On Fri, Dec 7, 2018 at 10:39 AM Jaime Sierra <jaisiero at gmail.com>
> > wrote:
> > > >>>> Hi,
> > > >>>>
> > > >>>> my name is Jaime Sierra, a researcher from Polytechnic University of
> > > >>>> Valencia, Spain. I would like to know how to compile & install
> > GROMACS
> > > >>>> 2018 with CUDA features with the "--cudart=shared" compilation
> > option
> > > >> to
> > > >>>> use it with our rCUDA software.
> > > >>>>
> > > >>>>
> > > >>>> We haven't had this problem in previous releases of GROMACS like
> > 2016,
> > > >>>> 5.1.4 and so on.
> > > >>>>
> > > >>>>
> > > >>>> Regards,
> > > >>>>
> > > >>>> Jaime.
> > > >>>> --
> > > >>>> Gromacs Users mailing list
> > > >>>>
> > > >>>> * Please search the archive at
> > > >>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > >>> posting!
> > > >>>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > > >>>>
> > > >>>> * For (un)subscribe requests visit
> > > >>>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users
> > or
> > > >>> send a mail to gmx-users-request at gromacs.org.
> > > >>> --
> > > >>> Gromacs Users mailing list
> > > >>>
> > > >>> * Please search the archive at
> > > >>> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > >>> posting!
> > > >>>
> > > >>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > > >>>
> > > >>> * For (un)subscribe requests visit
> > > >>> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users
> > or
> > > >>> send a mail to gmx-users-request at gromacs.org.
> > > >>>
> > > >> --
> > > >> Gromacs Users mailing list
> > > >>
> > > >> * Please search the archive at
> > > >> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > > >> posting!
> > > >>
> > > >> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > > >>
> > > >> * For (un)subscribe requests visit
> > > >> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > > >> send a mail to gmx-users-request at gromacs.org.
> > > --
> > > Gromacs Users mailing list
> > >
> > > * Please search the archive at
> > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > posting!
> > >
> > > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> > >
> > > * For (un)subscribe requests visit
> > > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > send a mail to gmx-users-request at gromacs.org.
> > --
> > Gromacs Users mailing list
> >
> > * Please search the archive at
> > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> > posting!
> >
> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >
> > * For (un)subscribe requests visit
> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> > send a mail to gmx-users-request at gromacs.org.
> --
> Gromacs Users mailing list
>
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.