Benchpark diff

[pearce8]

Given 2 runs of an experiment (same machine or different machines), what is different in:

• systems (spec diff?)
• build (ignore system software - compilers (+cmake etc.), mpi, math libraries
• run parameters (ignore scheduler, launcher)
• performance (we use caliper+thicket to diff these by establishing common call tree and parameters/metadata)

Spack diff notes: spack find -l # returns package and its hash ls spack location -I /hash_here/.spack/spec.json. # finds spec.json for the package spack spec # outputs a list of dependencies, one per line spack providers # outputs a list of "virtual packages": interfaces that can be provided by other packages, # e.g., mpi, blas, lapack # hip is amd's package name for rocm etc. https://github.com/spack/spack/pull/41711 # added ability to ignore dependencies in Spack diff

How to specify what to ignore? System sw (compilers, mpi, math libraries) but its a longer list:

./spack diff --ignore=gcc-runtime --ignore=hsa-rocr-dev --ignore=rocprim --ignore=rocrand --ignore=rocsparse --ignore=rocthrust --ignore=llvm-amdgpu --ignore=cmake --ignore=gmake --ignore=hip --ignore=blas --ignore=lapack --ignore=mpi /Users/pearce8/Documents/spack_diff/tioga-gtl-spec.json /Users/pearce8/Documents/spack_diff/magma-spec.json

produces:

• amg2023 amdgpu_target gfx90a

• amg2023 openmp True

• amg2023 openmp False

• amg2023 rocm False

• amg2023 rocm True

• hypre fortran True

• hypre amdgpu_target gfx90a

• hypre openmp True

• hypre fortran False

• hypre rocm False

• hypre openmp False
• hypre rocm True