Readme_OMP.md

Threading: openMP

void MyCoordination::calculate() {
  double ncoord = 0.;
  auto pos = getPositions();
  const unsigned nn = pos.size();
  unsigned nt = OpenMP::getNumThreads();
  Vector distance;
#pragma omp parallel num_threads(nt)
  {//start parallel scope
#pragma omp for 
    for (unsigned int i0 = 0; i0 < nn; ++i0) {
      for (unsigned int i1 = 0; i1 < nn; ++i1) {
        if (i0 == i1) {continue;}
        distance = delta(getPosition(i0), getPosition(i1));
        ncoord += (distance.modulo() < R_0) ? 1 : 0;
      }
    }
  }//end parallel scope
  setValue(ncoord);
}

This is nearly identical to the serial version. With few differences: We are asking the system if we want to run with more threads with unsigned nt = OpenMP::getNumThreads(), (if is the first time that it is called stores and return the number in the environmental variable PLUMED_NUM_THREADS, otherwise it will return the stored number).

Then with that information we open a parallel environment with #pragma omp parallel num_threads(nt) in which we explicilty specify the number of threads (otherwise il will use the number in the environmntal variable OMP_NUM_THREADS). Note that the pragma is prepended to a scope delimited by curly braces: the code within them will be distribuited between the threads.

Then we have the next pragma instruction: #pragma omp for: this pragma will tranform the for in order to make it work in the multithreaded environment.

SPOILER

There is at least a race condition here: can you spot it?

ANSWER

ncoord is the race condition: each time two threads or more threads execute the += simultaneusly they will increment the ncoord from the same value, meaning the the number saved in memory has been incremented only once instead of twice or more.

The solution is to append reduction(+:) to for pragma: #pragma for reduction(+:)

Plumed tools used:

• tools/OpenMP.h contains some function that are useful in working with openMP.
  • OpenMP::getNumThreads() to get the number of threads from the environmental variable `PLUMED_NUM_THREADS

back

See also: serial openMP MPI Cuda