-
Notifications
You must be signed in to change notification settings - Fork 144
Conference call notes 20160111
Kenneth Hoste edited this page Jan 13, 2016
·
9 revisions
(back to Conference calls)
Notes on the EasyBuild conference call dedicated to discussing implementation of RPATH support, Monday January 11th 2016 (5pm CET)
Alphabetical list of attendees (10):
- Xavier Besseron (Uni.lu, Luxembourg)
- Timothy Brown (Uni. of Colorado @ Boulder, US)
- Pablo Escobar (UniBas, Switzerland)
- Elisabeth Fisher (NASA, US)
- Todd Gamblin (LLNL, US)
- Fotis Georgatos (freelancer)
- Kenneth Hoste (HPC-UGent, Belgium)
- Lumír Jasiok (IIT4Innovations, Czech Republic)
- Alan O'Cais (JSC, Germany)
- Ward Poelmans (HPC-UGent, Belgium)
- Robert Schmidt (OHRI, Canada)
- implementation of support for RPATH linking in EasyBuild
- different available options?
- overall approach for implementing option of choice
- who does what?
-
compiler wrapper script (cfr. Spack)
- https://github.com/LLNL/spack/tree/master/lib/spack/env
- may be overkill for only RPATH
-
ld
wrapper script (cfr. Spack, script from Swedish colleagues)- should be sufficient (?)
-
patchelf
- doesn't work on OS X
- 'dummy' values need to be included that are long enough to hold actual path
-
$ORIGIN
- can be used to support relative RPATH'ing
- allows relocating installations later
- advantage of Conda over Nix
- not used in Spack today
-
sanity check?
- match
ldd
&readelf
- Todd has something that works on Linux
- match
-
RPATH support means sort of re-architecting EB design?
- RPATH support is step towards dropping modules as part of build process?
- modules are still useful for users
-
$LD_LIBRARY_PATH
may still be needed- e.g. for compilers (depends on how they were built)
- Intel compilers are patched at LLNL to use RPATH
-
use of RPATH will fix problems caused by loading modules
- loading CMake (which loads ncurses) breaks htop on the system
- also allows using tools built with different toolchains/versions
- linking to two different versions of a library in the same binary may not work though...
- comes close to 'toolchain-neutral' software (cfr. https://github.com/hpcugent/easybuild-framework/issues/570)
- can also helps with keeping size of environment under controle
- important for MPI, job launchers, performance (library lookup)
-
RPATH vs RUNPATH
- if both are set, RUNPATH wins
- if you do RUNPATH,
ld
will actually do both (by default)
- if you do RUNPATH,
- $LD_LIBRARY_PATH wins over RUNPATH, so to much potential issues
- supporting both should be fairly trivial?
- if both are set, RUNPATH wins
-
default on OS X is to use full paths to libraries
- how is that different from RPATH?
- very different from default on Linux
- https://github.com/LLNL/spack/blob/develop/var/spack/packages/bzip2/package.py
-
RPATH & CMake (Todd)
- cmake by default will re-write RPATHs on install.
- so if you use an ld wrapper it will not cover CMake in the standard case
- (ironic b/c cmake has really good rpath support, but it requires a special case if YOU want to implement good RPATH support)
- so you have to set CMAKE_INSTALL_RPATH on your cmake builds to tell it to write the right thing:
- https://github.com/LLNL/spack/blob/develop/lib/spack/spack/build_environment.py#L215
- the reason it does that is b/c it sets the RPATH at build time to the build directory (nice for developers)
- but it rewrites on install so that the install works
- specify in easyconfigs which libraries to RPATH against it?
- --rpath-deps?
- also support to specify "RPATH all the things"
- concern of reproducibility of build where RPATH was used
- which deps were RPATH'ed should be dumped to the easyconfig copied in the install dir/archive
- Spack wrapper: https://github.com/LLNL/spack/blob/develop/lib/spack/env/cc
- see also open PR https://github.com/LLNL/spack/pull/248
- slide @ https://tgamblin.github.io/files/Gamblin-Spack-SC15-Talk.pdf (slide 16)
- ld wrapper by Swedish colleagues
- license?
- Rob
- Ward (help with wrapper script)
- via https://github.com/hpcugent/easybuild-framework/issues/651 & EasyBuild mailing list