Coarse Grain Parallelisation

• Section1 (source):
  distribution of several tasks to threads:

  • SECTIONS/END SECTIONS
  • SECTION
  combination of PARALLEL block and task distribution:

  • PARALLEL SECTIONS
  problem: number of tasks is fixed (can't be adapted to number of processors)

  • too many threads: some wait
  • not enough threads: some get several SECTIONS
• Section2 (source):
  parallelise loop in smooth with PARALLEL DO

  • nested parallelism
  result: only 3 of 6 requested threads work

  • default: inner PARALLEL blocks get only one thread
  • set environment variable OMP_NESTED to TRUE
• Section3 (source):
  explicit control of nesting:

  • omp_set_nested
  • (overrules environment variable)
  info functions:

  • omp_get_num_threads()	number of threads
    omp_get_num_procs()	number of cpus
    omp_get_thread_num()	number of current thread (0 .. N-1)

  program output (with OMP_NUM_THREADS=6):

  • Zahl der Threads in parallel sections:  6
    Zahl der Threads in inner parallel loop:  1
    Zahl der Threads in inner parallel loop:  1
    Zahl der Threads in inner parallel loop:  1
    

  • outer block: 6 threads in PARALLEL block (SECTIONS only distributes)
  • inner block: 1 thread per block, 3 blocks (one for each SECTION)
  standard allows: nested blocks are simply ignored!

  • (f.i. guide 3.6 or f90 under IRIX 6.5)