Skip to content

Latest commit

 

History

History
59 lines (53 loc) · 2.05 KB

BuildValidation.md

File metadata and controls

59 lines (53 loc) · 2.05 KB

NWChemEx Build Validation Policy

It is impractical to test all functionality in NWChemEX on every permutation of software/hardware platforms for every pull request. We should identify a core test set which executes on every pull request in order to catch low-hanging bugs, and a more comprehensive extended test set to run before releases.

  • Core test set

    • Adequatly cover core functionality such as Hartree-Fock, DFT, MP2, CC for a small range of molecules/basis sets.
    • Hard execution time limit for the complete set (1 hour?).
    • While the core test set might not explicitly test the features contributed in the pull request, the core test set is executed in addition to the new tests included with the pull request.

  • Extended test set

    • Run before each public release, or periodically.
    • A hard execution time limit makes less sense here.
    • In addition to covering more of the NWChemEx code,

  • NWChemEx should be validated for most or all of the hardware/softare environments in use at major supercomputing centers. Covering all permutations is impractical, but the most important combinations should be identified and updated as platforms are introduced or retired.

    • Hardware:
      • Intel Xeon CPU
      • AMD CPU
      • IBM PowerX CPU
      • Intel Xeon Phi
      • Nvidia GPU
    • Software:
      • Compilers
        • Intel
        • GCC
        • PGI
        • clang
      • Math Libraries
        • MKL
        • ATLAS
        • OpenBLAS
      • MPI
        • MVAPICH
        • IMPI
        • OpenMPI
      • Other Software
        • Linux kernel (from experience, this is important)
          • Current Ubuntu and Redhat LTS?
      • Just cover kernels used on target supercomputers?
        • Different versions of CUDA, OpenCL, OpenACC, etc.

  • Continuous integration software

    • TravisCI is convenient for the core test set triggered on each PR. It does not offer an adequate hardware/software environment to validate code for the full range of use cases.
    • What are some alternative CI frameworks which offer more control over these parameters?