hybrid-bench

Dependencies:

• MPL w/ hybrid scheduling
• CUDA (version ?)
• Futhark (preferably most recent: 0.25.16)
• smlfut
• smlpkg
• For dmm benchmark: OpenBLAS
• For benchmarking scripts and plotting: jq, Python 3, matplotlib, numpy, etc.

Required input datasets:

• For sparse-mxv benchmark: stokes (SuiteSparse)
• For quickhull benchmark: [TODO: instructions for points file generated by pbbsbench]
• For bfs benchmark: [TODO: instructions for rmat graph generated by pbbsbench]

Setup:

• Edit bench/config.mk for your paths (MPL, OpenBLAS, etc.)
• Edit scripts/gencmds for your paths (input datasets) [TODO: make this nicer]
• run cd dep && smlpkg sync

To run a benchmark:

• cd bench/<BENCHMARK>; make; ./main.mpl.bin <ARGS>
• You can run scripts/gencmds to see preferred configuration
• Alternatively, to run a batch: scripts/gencmds | jq 'select(<QUERY>)' | scripts/runcmds --compile
  • For example: scripts/gencmds | jq 'select(.tag == "quickhull" and .procs == 80 and .gpus == 4 and .impl == "hybrid")' | scripts/runcmds --compile