Multi-threaded random_r is slower than single threaded version

Question

The following program is essentially the same as the one described here. When I run and compile the program using two threads (NTHREADS == 2), I get the following run times:

real        0m14.120s
user        0m25.570s
sys         0m0.050s

When it is run with just one thread (NTHREADS == 1), I get run times significantly better even though it is only using one core.

real        0m4.705s
user        0m4.660s
sys         0m0.010s

My system is dual core, and I know random_r is thread safe and I am pretty sure it is non-blocking. When the same program is run without random_r and a calculation of cosines and sines is used as a replacement, the dual-threaded version runs in about 1/2 the time as expected.

#include <pthread.h>
#include <stdlib.h>
#include <stdio.h>

#define NTHREADS 2
#define PRNG_BUFSZ 8
#define ITERATIONS 1000000000

void* thread_run(void* arg) {
    int r1, i, totalIterations = ITERATIONS / NTHREADS;
    for (i = 0; i < totalIterations; i++){
        random_r((struct random_data*)arg, &r1);
    }
    printf("%i\n", r1);
}

int main(int argc, char** argv) {
    struct random_data* rand_states = (struct random_data*)calloc(NTHREADS, sizeof(struct random_data));
    char* rand_statebufs = (char*)calloc(NTHREADS, PRNG_BUFSZ);
    pthread_t* thread_ids;
    int t = 0;
    thread_ids = (pthread_t*)calloc(NTHREADS, sizeof(pthread_t));
    /* create threads */
    for (t = 0; t < NTHREADS; t++) {
        initstate_r(random(), &rand_statebufs[t], PRNG_BUFSZ, &rand_states[t]);
        pthread_create(&thread_ids[t], NULL, &thread_run, &rand_states[t]);
    }
    for (t = 0; t < NTHREADS; t++) {
        pthread_join(thread_ids[t], NULL);
    }
    free(thread_ids);
    free(rand_states);
    free(rand_statebufs);
}

I am confused why when generating random numbers the two threaded version performs much worse than the single threaded version, considering random_r is meant to be used in multi-threaded applications.

Answer 1

A very simple change to space the data out in memory:

struct random_data* rand_states = (struct random_data*)calloc(NTHREADS * 64, sizeof(struct random_data));
char* rand_statebufs = (char*)calloc(NTHREADS*64, PRNG_BUFSZ);
pthread_t* thread_ids;
int t = 0;
thread_ids = (pthread_t*)calloc(NTHREADS, sizeof(pthread_t));
/* create threads */
for (t = 0; t < NTHREADS; t++) {
    initstate_r(random(), &rand_statebufs[t*64], PRNG_BUFSZ, &rand_states[t*64]);
    pthread_create(&thread_ids[t], NULL, &thread_run, &rand_states[t*64]);
}

results in a much faster running time on my dual-core machine.

This would confirm the suspicion it was meant to test - that you are mutating values on the same cache line in two separate threads, and so have cache contention. Herb Sutter's 'machine architecture - what your programming language never told you' talk is worth watching if you've got the time if you don't know about that yet, he demonstrates false sharing starting at around 1:20.

Work out your cache line size, and create each thread's data so it is aligned to it.

It's a bit cleaner to plonk all the thread's data into a struct, and align that:

#define CACHE_LINE_SIZE 64

struct thread_data {
    struct random_data random_data;
    char statebuf[PRNG_BUFSZ];
    char padding[CACHE_LINE_SIZE - sizeof ( struct random_data )-PRNG_BUFSZ];
};

int main ( int argc, char** argv )
{
    printf ( "%zd\n", sizeof ( struct thread_data ) );

    void* apointer;

    if ( posix_memalign ( &apointer, sizeof ( struct thread_data ), NTHREADS * sizeof ( struct thread_data ) ) )
        exit ( 1 );

    struct thread_data* thread_states = apointer;

    memset ( apointer, 0, NTHREADS * sizeof ( struct thread_data ) );

    pthread_t* thread_ids;

    int t = 0;

    thread_ids = ( pthread_t* ) calloc ( NTHREADS, sizeof ( pthread_t ) );

    /* create threads */
    for ( t = 0; t < NTHREADS; t++ ) {
        initstate_r ( random(), thread_states[t].statebuf, PRNG_BUFSZ, &thread_states[t].random_data );
        pthread_create ( &thread_ids[t], NULL, &thread_run, &thread_states[t].random_data );
    }

    for ( t = 0; t < NTHREADS; t++ ) {
        pthread_join ( thread_ids[t], NULL );
    }

    free ( thread_ids );
    free ( thread_states );
}

with CACHE_LINE_SIZE 64:

refugio:$ gcc -O3 -o bin/nixuz_random_r src/nixuz_random_r.c -lpthread
refugio:$ time bin/nixuz_random_r 
64
63499495
944240966

real    0m1.278s
user    0m2.540s
sys 0m0.000s

Or you can use double the cache line size, and use malloc - the extra padding ensures the mutated memory is on separate lines, as malloc is 16 (IIRC) rather than 64 byte aligned.

(I reduced ITERATIONS by a factor of ten rather than having a stupidly fast machine)

Answer 2

I don't know if this is relevant or not - but i just saw a very similar behavior (order of magnitude slower with 2 threads than with one) ... I basically changed a:

  srand(seed);
  foo = rand();

to a

  myseed = seed;
  foo = rand_r(&myseed);

and that "fixed" it (2 threads is now reliably almost twice as fast - e.g. 19s instead of 35s).

I don't know what the issue could have been -- locking or cache coherence on the internals of rand() maybe? Anyway, there is also a random_r() so maybe that would be of use to you (a year ago) or someone else.