6 Work Sharing

So far, the code within a parallel region is executed by all the threads of the team. Nevertheless, each thread might execute parts of the computations.

      do I = 1, N
         A(I) = ...
      end do

!$omp parallel private (INODE, NNODE, LB, UB)
      NNODE = omp_get_num_threads ()
      INODE = omp_get_thread_num()
      LB  = I * (N / NNODE) + 1
      UB  = INODE * (N / NNODE)
      do LB = UB
         A(I) = ..
      end do
!$omp end parallel

!$omp do
      do I = 1, N
         A(I) = ...
      end do

New Clauses first met in work-sharing directives are:

• lastprivate (list) - when threads join the variables in list take on the value they would normally have after the last loop iteration of sequential execution;
• schedule(static|dynamic|guided|runtime)
• ordered - iterations are performed in the same order as in sequential execution. The ordered clause binds to the dynamically enclosing do directive.

• schedule(static,chunk) - round-robin allocation of work to threads in ``chunks''. Default chunk size is $n/nthread$;
• schedule(dynamic,chunk) - dynamic allocation to threads. Blocks of work of size chunk. Default chunk size is 1;
• schedule(guided,chunk) - this is fully explained in the documentation.
• schedule(runtime) - allows the system to select the scheduling at run time using an environment variable, e.g. OMP_SCHEDULE=''guided,4''.