Benchmark

多进程软件锁

1. Peterson算法(filter算法)
2. lamport_bakery算法
3. McGuire算法

硬件锁:

1. pthread_mutex

---

代码:

• 以peterson为例,其他类同

peterson_mutex_t peterson_mutex;

void* peterson(void* arg) {
  int i = *(int*)arg;
  // sleep(1);
  peterson_mutex_lock(&peterson_mutex, i);
  test();
  peterson_mutex_unlock(&peterson_mutex, i);
}

double peterson_main() {
  peterson_mutex_init(&peterson_mutex);

  pthread_t tid_list[n];
  int index[n] = {0};

  double start = omp_get_wtime();
  for (int i = 0; i < n; ++i) {
    index[i] = i;
    pthread_t tid;
    pthread_create(&tid, NULL, peterson, index + i);
    tid_list[i] = tid;
  }
  for (int i = 0; i < n; ++i) {
    pthread_join(tid_list[i], NULL);
  }
  double end = omp_get_wtime();
  printf("peterson_sum %d\n", sum);
  return end - start;
}

---

结果:

• 说明: 在默认测试状态下,一般只有peterson会产生竞态问题!

Average delay in 3 times with 200 processes |||  peterson: 0.547366 , pthread_mutex: 0.019146 , lamport_bakery: 0.019223, McGuire: 0.018633

---

为了使进程陷入竞态,暂且使用sleep(1)

• 由于软件锁都是忙等待模式,所以200进程的占用cpu忙等待有时候会卡死电脑,这里使用100进程

Average delay in 3 times with 100 processes |||  peterson: 1.035813 , pthread_mutex: 1.016462 , lamport_bakery: 1.640425, McGuire: 1.936620

• 应改成用alarm()统一唤醒

---

---

• 硬件锁pthread_mutex_lock,使用cmpxchgl保证原子性,阻塞而不是自旋
• 可以尝试使用__sync_fetch_and_and

---

• 经测试,发现bakery算法存在问题,但是懒得改了.