This is a synthetic benchmark that runs at varied levels of arithmetic intensity.

Building

Run `make` to build the program.

Run `make NOGEOPM=1` to build the program with no GEOPM dependencies.

Running

The Makefile will generate 3 executables: bench_sse, bench_avx2, bench_avx512, which run with the vector instruction types SSE, AVX2 and AVX512 respectively.

Help

$ ./bench_? -h
Run a benchmark with configurable arithmetic intensity and MPI imbalance
Usage: ./bench_? [OPTIONS]

Options:
  -h,--help                   Print this help message and exit
  --slowdown UINT=1           When imbalance is present, this specifies the amount of work for slow ranks to perform, as a factor of the amount of work the fast ranks perform.
  --slow-ranks UINT=0         The number of ranks to run with extra work for an imbalanced load.
  --segments UINT=67108864    The number of 64-byte-sized segments in the problem array.
  -v,--verbose                Run in verbose mode
  -s,--single-precision       Run in single-precision mode
  -i,--iterations UINT=5      The number of times to run each phase of the benchmark
  -b,--benchmarks UINT ...    List of benchmark intensity variants to run (all are run if this option is not specified)

Execution

The program is an MPI app. It uses as many ranks as are available.

For example, use the following command to launch 44 ranks with AVX512 instructions.

geopmlaunch srun -N 1 -n 44 ./bench_avx512

If compiled with NOGEOPM option, use the following command to do the same as above.

srun -N 1 -n 44 ./bench_avx512