Four Counters One Line

Four threads, each pinned to its own socket, run this. Each thread gets a distinct id in 0..3.

struct counter { long value; };
static struct counter counters[4];   /* 32 bytes total */

void *worker(void *arg) {
    int id = *(int *)arg;
    for (long i = 0; i < 500000000L; i++)
        counters[id].value++;
    return NULL;
}

One thread finishes in 1.6 s. Four threads take 18 s. No thread ever touches another thread's element. More than one of the following removes the slowdown. Which do?

Change value to _Atomic long and increment it atomically.

Allocate counters with malloc instead of making it a static array.

Pad struct counter to 64 bytes and align counters to a 64-byte boundary.

Accumulate into a local long inside worker and store it into counters[id].value once, after the loop.

Declare value as volatile long.