OpenMP

• Multi-threading
• Shared mem.
• Incremental parallelization

fork-join model

Task distribution structure:

• do / for
• section
• single

for loop scheduling:

• static: suitable for balanced workload
• dynamic
• guided: k tasks for the 1st time; k/2 tasks for the 2nd time; ...
• • better working balance but more cost than static
  • not balanced very well as dynamic

sections

#pragma omp parallel num_threads(2)
{
  #pragma omp sections
  {
    #pragma omp section
    {
      // section1
    }
    #pragma omp section
    {
      // section2
    }
  }
}

Sync

critical

atomic

Loosen Consistency

• value of a shared variable is not guaranteed to be the latest

• need "flush"

• auto flushed commands:

  • in/out: parallel, critical, ordered
  • out: for, sections, single

• manually flush:

  #pragma omp flush [acq_rel / release / acquire] [var1, ...]
  

---

Last update: March 29, 2022

Authors: Co1lin