OpenMP parallelisation of pi calculation is either slow or wrong

问题

I'm having trouble parallelising my monte carlo method to calculate pi. Here is the parallelised for-loop:

#pragma omp parallel for private(i,x,y) schedule(static) reduction(+:count)
  for (i = 0; i < points; i++) {
    x = rand()/(RAND_MAX+1.0)*2 - 1.0;
    y = rand()/(RAND_MAX+1.0)*2 - 1.0;

    // Check if point lies in circle
    if(x*x + y*y < 1.0) { count++; }
  }

The problem is, it underestimates pi if I use schedule(static), and its slower than the serial implementation if I use schedule(dynamic). What am I doing wrong? I've tried other ways to fix it (like this: Using OpenMP to calculate the value of PI) but it's still much slower than the serial implementation.

Thanks in advance

回答1:

Assuming you're using the C library rand function, that function is not reentrant or thread-safe. POSIX provides a rand_r function, but (to quote the glibc documentation):

POSIX.1 extended the C standard functions to support reproducible random numbers in multi-threaded programs. However, the extension is badly designed and unsuitable for serious work.

In particular, the seed must be an unsigned int, which doesn't have enough bits for a good PRNG. They recommend using the SVID random number functions, of which nrand48_r is probably what you're looking for.

Alternatively, you could use a different library.

回答2:

One thing you have to take into account when doing such things in parallel, is that there might be different rounding errors, caused due to different ways of doing the calculation.

Example:

((A+B) + (C+D)) where (A+B) and (C+D) will be computed in parallel might vary from the serial approach (((A+B) + C) + D).

来源：https://stackoverflow.com/questions/8991119/openmp-parallelisation-of-pi-calculation-is-either-slow-or-wrong

标签

c++

parallel-processing

openmp