Telcon: 2015 07 23

Participants

• Todd Gamblin (LLNL)
• Jim Galarowicz (LLNL)
• Mike Collette (LLNL)

Contributions

• Mark Miller contributing better HDF5 package, other FastMath packages
• zsh from Tom Scogland
• CBTF from Krell Institute folks

Bugs

• Can't pass input to Spack Executable class for commands that need it on stdin (add input/output parameters), e.g.:

  cat = which('cat')
  cat(input='foo.txt')
  

• Large RPATH expression in CBTF build breaks CMake

  -- Installing: /usr/global/tools/openspeedshop/oss-dev/new_spack_fork/spack/opt/spack/chaos_5_x86_64_ib/gcc-4.4.7/openspeedshop-2.1-cbtf-ofxen7abe4ahtejzpoeinzfe3oj5xx3m/lib64/libopenss-message.so
   RegularExpression::compile(): Expression too big.
  

  • Spack currently adds entire DAG to RPATH. Probably overkill.

    • Should we only add direct dependencies to RPATHs?

    • Suggested implementation in package.py:

      @property                                                                                                    
      def rpath(self):                                                                                             
          """Get the rpath this package links with, as a list of paths."""                                         
          rpaths = [self.prefix.lib, self.prefix.lib64]                                                            
          rpaths.extend(self.spec[d].prefix.lib for d in self.spec.dependencies                                    
                        if os.path.isdir(self.spec[d].prefix.lib))                                                      
          rpaths.extend(self.spec[d].prefix.lib64 for d in self.spec.dependencies                                  
                        if os.path.isdir(self.spec[d].prefix.lib64))                                                    
          return rpaths  
      

    • Relies on dependency libraries to have RPATHS for their deps (reasonable)

  • Other options:
    • Should we shorten Spack install paths and use links for descriptive hierarchy?

Discussion Items

• Mark Miller proposes investigating common variant names for common options:
  • shared/static
  • debug/optimized
  • fortran variants
  • 32/64 bit
    • Todd: should probably be an architecture thing, not a variant)
      • should we support linking 64-bit binaries with 32-bit binaries?
      • does that even work?
    • Mike: this probably refers to length of variables, NOT arch.
      • e.g.: Hypre has a 64-bit int compile option.
      • Todd: ok I understand this better now.
  • Todd: 2-d, 3-d variants
• Mark Miller wants a way to understand a compiler's Fortran name mangling
  • Suggest deferring this until we get into guarantees about ABI compatibility bt/w compilers.
  • Spack should be able to have rules about consistency, pick reasonable compiler/arg combinations.
• RPM generation a big issue for TOSS 3.
  • Want to use Spack to generate RPMs for potentially many locations, including /usr/*** and /opt
  • How to generate RPMs for many ecosystems like this?
    • How to account for dependency RPMs and how to know what's on the system?
• Mike C having issues with Boost
  • Jim, too.
  • Missing library in boost build.
  • Mike says there is a file that needs modifying in the boost build so that b2 will use different compilers.
    • Todd: consider using Boost-cmake?
• Mark suggested changing names of spack compilers to spack-cc, spack-cxx, etc.
  • Mike doesn't want to have his packages disrupted too much
  • Suggest gradually phasing in spack-cc, etc. and keeping cc in the meantime
• Mike having issues with clang + Python on BG/Q
  • other 8 configurations are fine
  • breaks the first time with no particular error msg (just after patching?)
  • running build a second time makes it work.