mpidgemmio.docopt

mpidgemmio. Stress CPU cores with computation (dgemm), and I/O (sequential writes)
Each core can be dedicated to either computation or I/O. By default all cores run computation.
The balance between computation and I/O cores can be adjusted using the --numio parameter.
The affinity of threads can be adjusted using the --affinity parameter.

Usage:
    mpidgemmio (-h | --help)
    mpidgemmio --version
    mpidgemmio --blocks=<blocks> --outer=<outer> [--inner=<inner>] [--numio=<numio>] [--compute_outfile=<comp_file>] [--io_outfile=<io_file>] [--tmpdir=<outdir>] [--gflops] [--ioboost] [--quiet] [--affinity=<policy>] [--read]

Options:
    -h --help                       Show help.
    -q --quiet                      Suppress all stdout output.
    --blocks=<blocks>               Number of 4K blocks to write per I/O thread.
    --outer=<outer>                 Number of outer loops (for both compute and I/O threads).
    --inner=<inner>                 Number of inner loops (compute threads only) [default: 400].
    --numio=<numio>                 Number of I/O threads [default: 0].
    --compute_outfile=<comp_file>   Output file for compute thread data [default: compute.txt].
    --io_outfile=<io_file>          Output file for I/O thread data [default: io.txt].
    --tmpdir=<outdir>               I/O threads write files to this directory [default: /tmp/mpidgemmio].
    --gflops                        For compute threads, output GFLOPS instead of execution time.
    --ioboost                       Use dirty.ko boost feature.
    --read                          Do sequential read() instead of write().
    --affinity=<policy>             Process affinity policy. Policies are described next. [default: rank]
                                        "rank" policy: Set affinity to cpu number that corresponds to rank number. #cpu <- #rank.
                                        "rank_inv" policy: Like "rank" but in reverse order.
                                        "interleave" policy: interleave compute and I/O.
                                        "random" policy: set affinity randomly.
                                        "none" policy: Don't set affinity.
                                        "iolowtemp" policy: Schedule I/O threads to cores with lower temperature.
                                        "complowtemp" policy: Schedule compute threads to cores with lower temperature.