Increasing pthreads thread count has no effect on speed

Question

In the following program, increasing the number of threads does not yield any speed-up benefits at all (measured with the time command under linux). I have run on it on the following processors:

• Intel i5 M520
• Intel Xeon X5650

It seems to me that the logic of dividing the work among the threads is correct. I have even tried removing the lock, which obviously gives the wrong result, but still no increase in speed. Any ideas?

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

double sum;
pthread_mutex_t mutex;
typedef struct {
    int start;
    int end;
}sumArg;

void *sumRoots(void *arg) {
    sumArg s = *((sumArg *) arg);
    int i = s.start;
    double tmp;
    while(i <= s.end) {
        tmp = sqrt(i);
        pthread_mutex_lock(&mutex);
        sum += tmp;
        pthread_mutex_unlock(&mutex);
        i++;
    }
    free(arg);
}

int main(int argc, char const *argv[]) {
    int threadCount = atoi(argv[1]);
    int N = atoi(argv[2]);
    if (N < 1 || threadCount < 1) printf("Usage: ./sumOfRoots threads N\n");

    pthread_t tid[threadCount];
    pthread_attr_t attr;
    pthread_attr_init(&attr);
    pthread_mutex_init(&mutex, NULL);

    sumArg *s;
    int i = 0;
    while(i < threadCount) {
        s = (sumArg *) malloc(sizeof(sumArg));
        s->start = ((N/threadCount) * i) + 1;
        s->end = (N/threadCount) * (i + 1);
        pthread_create(&tid[i], &attr, sumRoots, s);
        i++;
    }

    i = 0;
    while(i < threadCount) pthread_join(tid[i++], NULL);
    printf("sum: %f\n", sum);
    return 0;
}

Edit: Here are some runs of the program, all running on an i5 M520:

time ./sumOfRoots 1 1000000000
sum: 21081851083600.558594

real    0m21.933s
user    0m21.268s
sys     0m0.000s

time ./sumOfRoots 2 1000000000
sum: 21081851083600.691406

real    0m21.207s
user    0m21.020s
sys     0m0.008s

time ./sumOfRoots 4 1000000000
sum: 21081851083600.863281

real    0m21.488s
user    0m21.116s
sys     0m0.016s

time ./sumOfRoots 8 1000000000
sum: 21081851083601.777344

real    0m21.432s
user    0m21.092s
sys     0m0.020s

I believe the variation in the sum is caused by floating point precision loss.

Answer 1

The reason why the timing remains virtually unchanged is that it it dominated by synchronization. On my computer a single-thread solution was even faster!

Changing the code as follows brings the timing in line with expectations:

void *sumRoots(void *arg) {
    sumArg s = *((sumArg *) arg);
    int i = s.start;
    double tmp = 0;
    while(i <= s.end) {
        tmp += sqrt(i++);
    }
    pthread_mutex_lock(&mutex);
    sum += tmp;
    pthread_mutex_unlock(&mutex);
    free(arg);
    return 0;
}

Now your thread runs for a while without synchronization, and then synchronizes only once during the addition.

The timing I see on my system is as follows:

> time ./a.out 1 1000000000
sum: 21081851083600.558594
real    0m13.220s
user    0m13.098s
sys 0m0.009s

> time ./a.out 2 1000000000
sum: 21081851083600.863281

real    0m6.613s
user    0m12.930s
sys 0m0.027s