Want to avoid one call to sempost for eight threads

Question

Scenario is:

    nthreads = 8
    for j=0 = nthreads
        pthread_created (thread_func)


    for 1=0 to 1000
    {
        // Some work.

        for j=0 = nthreads
          sempost(mutex1)

        // thread_func()

        for j=0 = nthreads
          semwait(mutex2)

       // Some work.

    }

    thread_func()
    {
        while(1)
        {
            semwait(mutex1)

            // Thread funcionality.
            sempost(mutex2)
        }
    }

Que: I want a more efficient synchronization technique than current one. I want to remove one call to sempost (total 8, one for one thread) at least. At the same time I want every thread to start their work at the same time.

How can it be achieved?

---

doing barrier or sempost/wait by each thread would involved the same expense. I want to remove one set of sempost/wait or one barrier in your case

Answer 1

What you want is a barrier, which blocks waiting threads until all threads have entered barrier_wait(). Depending on your threading implementation, you might already have a barrier primitive (pthread_barrier for instance).

The other option is a conditional wait (cond_wait), which would be ideal for your "launch" option.

To synchronize, you need to use one of these primitives. Synchronization has overhead as you have determined - either make larger workloads, synchronize less, or use another concurrency scheme (STM)