Boost.JSON Benchmarks

To run the benchmarks first run clone.sh to fetch the third party repositories. Then run the bench program with no arguments for a list of command line options.

Usage: bench [options...] <file>...

Options:  -t:[p][s]            Test parsing, serialization or both
                                 (default both)
          -i:[b][d][r][c][n]   Test the specified implementations
                                 (b: Boost.JSON, pool storage)
                                 (d: Boost.JSON, default storage)
                                 (u: Boost.JSON, null parser)
                                 (s: Boost.JSON, convenient functions)
                                 (o: Boost.JSON, stream operators)
                                 (r: RapidJSON, memory storage)
                                 (c: RapidJSON, CRT storage)
                                 (n: nlohmann/json)
                                 (default all)
          -n:<number>          Number of trials (default 6)
          -b:<branch>          Branch label for boost implementations
          -m:(i|p|n)           Number parsing mode
                                 (i: imprecise)
                                 (p: precise)
                                 (n: none)
                                 (default imprecise)
          -f                   Include file IO into consideration when testing parsers

When building with b2, it is possible to create several different copies of the bench program for different build properties (toolset, build variant, etc.). Rather than figuring out where those programs are located from b2 output you can simply run them directly from the build system. For this, use the bench//run target. For example:

b2 variant=release toolset=gcc,clang bench//run

You can set bench options using the flag bench.option. By default the target uses all files from the bench/data directory as benchmark data. If you want to use just a few of those files or use different files altogether, use the flag bench.file. Finally, when running benchmarks it is often necessary to use a wrapper program that sets up the process for better result stability. This can be achieved with the flag bench.launcher. In this case also consider telling b2 to not run several jobs at a time with -j1, as they would interfere with each other. Combining everything together we get:

b2 -j1 variant=release toolset=gcc,clang bench//run bench.option=-t:p bench.option=-n:10 bench.option=-i:b bench.launcher="nice -n -20 taskset -c 1" bench.file=mydata.json

Which is the equivalent of

nice -n -20 taskset -c 1 path/to/gcc/version/of/bench -t:p -n:10 -i:b mydata.json
nice -n -20 taskset -c 1 path/to/clang/version/of/bench -t:p -n:10 -i:b mydata.json

Alternatively, you can copy the version of bench that you want into the bench directory by requesting the target bench//bench-local and then run it manually:

b2 variant=release toolset=gcc bench//bench-local
nice -n -20 taskset -c 1 bench/bench-local -t:p -n:10 -i:b mydata.json

The benchmarked files were sourced from the simdjson repository.