BenchmarkConfigSweeps

BenchmarkConfigSweeps.jl can be used for running benchmarks over various --nthreads settings, environment variables, and julia versions.

Each configuration is specified as a combination of configuration specifiers such as BenchmarkConfigSweeps.nthreads, BenchmarkConfigSweeps.env, and BenchmarkConfigSweeps.julia. The set of configurations to run benchmarks is simply an iterable of (tuples of) such configuration specifiers:

using BenchmarkConfigSweeps

nthreads_list = [1, 4, 8, 16]
configs = Iterators.product(
    zip(
        BenchmarkConfigSweeps.nthreads.(nthreads_list),
        BenchmarkConfigSweeps.env.("OPENBLAS_NUM_THREADS" .=> nthreads_list),
    ),
    BenchmarkConfigSweeps.env.(
        "JULIA_PROJECT" .=> ["baseline", "target"],
        "JULIA_LOAD_PATH" => "@",  # not varied
    ),
    BenchmarkConfigSweeps.julia.(["julia1.5", "julia1.6"]),
)

It can be then passed to BenchmarkConfigSweeps.run:

BenchmarkConfigSweeps.run("build", "benchmarks.jl", configs)

where "build" is the directory to store the result and "benchmarks.jl" is the Julia script that defines the variable SUITE :: BenchmarkGroup at the top-level.

The result can be loaded using

sweepresult = BenchmarkConfigSweeps.load("build")

The sweepresult object implements the Tables.jl interface. For example, it can be easily converted into a DataFrame by

using DataFrames
df = DataFrame(sweepresult)

Note that the default table conversion is rather too DWIM-y in that it tries to guess the meaning of BenchmarkGroup keys. For more information, see BenchmarkConfigSweeps.flattable for how it works. Use BenchmarkConfigSweeps.simpletable for programmatic processing.