Benchmarking

#| id: meson-executables
executable('ls-bench', 'src/benchmark.cpp',
    dependencies: [libeigen, benchmark],
    include_directories: incdir)

We use Google's benchmark library to setup some benchmarks for both run-time and memory use.

//| file: src/benchmark.cpp
#include <benchmark/benchmark.h>
#include <random>
#include "measurements.hpp"

using namespace ls_bench;

template <typename M>
static void BM_sparse_qr(benchmark::State &state) {
    // state.PauseTiming();
    std::mt19937_64 r;
    auto coef = random_coef(r, state.range(0), 10, 10);
    auto [input, value] = mock_measurements<M>(r, coef, state.range(1), 0.1);
    // state.ResumeTiming();
    for (auto _ : state) {
        auto result = MatrixTraits<M>::solve_qr(input, value);
        benchmark::DoNotOptimize(result);
    }
}
BENCHMARK(BM_sparse_qr<Sparse>)->Args({10, 1000});
BENCHMARK(BM_sparse_qr<Dense>)->Args({10, 1000});

BENCHMARK_MAIN();

Memory¶

The benchmark library can tell us a lot about compute efficiency, but it doesn't naturally measure memory usage. There is a utility in benchmark to measure memory use, but it requires creating a custom allocator and then forcing all of the code to use that allocator. An easier method is to use Valgrind and its massif tool to track memory usage from the outside. Simply run:

valgrid --tool=massif ./build/debug/least-squares -n 10 -m 100000 -dense

to measure the memory use of the least-squares test program.