rand.c

// gist for showing that rand() with global lock is slower ~60x times
// (came to this for librdkafa random partitioning with producer per group, and eventually program has ~1200 threads)
//
// $ g++ -DRAND -o rand -O3 -pthread rand.cpp
// $ g++ -DRAND_R -o rand_r -O3 -pthread rand.cpp
//
// $ time ./rand
//
// real    0m2.336s
// user    0m2.685s
// sys     0m15.588s
//
// $ time ./rand_r
//
// real    0m0.041s
// user    0m0.106s
// sys     0m0.042s

#include <cstdlib>
#include <thread>
#include <vector>
#include <stdexcept>

static constexpr size_t MESSAGES = 10'000'000;
static constexpr size_t THREADS = 1000;
static constexpr size_t MESSAGES_PER_THREAD = MESSAGES/THREADS;
thread_local unsigned int seed = 0xdeadbeaf;

// from google benchmark (just in case, the gen_rand() does not optimized anyway)
inline __attribute__((always_inline)) void DoNotOptimize(int const &value) {
  asm volatile("" : : "r,m"(value) : "memory");
}

int gen_rand()
{
#ifdef RAND
	return rand();
#elif defined(RAND_R)
	return rand_r(&seed);
#else
#error No rand/rand_r seleted
#endif
}

int main()
{
	std::vector<std::thread> threads;
	for (int i = 0; i < THREADS; ++i) {
		threads.emplace_back([]() {
			for (int i = 0; i < MESSAGES_PER_THREAD; ++i) {
				DoNotOptimize(gen_rand());
			}
		});
	}
	for (auto &thread : threads) {
		thread.join();
	}
	return 0;
}