ACENET Summer School - Directive-Based Parallel Programming with OpenMP and OpenACC: Glossary

Shared-memory parallel programs break up large problems into a number of smaller ones and execute them simultaneously

OpenMP programs are limited to a single physical machine

OpenMP libraries are built into all commonly used C, C++, or Fortran compilers

Pragmas are directives to the compiler to parallelize something

Thread number is typically controlled with an environment variable, OMP_NUM_THREADS

Order of execution of parallel elements is not guaranteed.

If the compiler doesn’t recognize OpenMP pragmas, it will compile a single-threaded program. But you may need to escape OpenMP function calls.

The PARALLEL FOR (or PARALLEL DO) pragma makes a loop execute in parallel

A single variable accessed by different threads can cause wrong results

The PRIVATE clause makes a copy of a variable for each thread

Race conditions can be avoided by using the omp critical or the omp atomic directives

The best option to parallelize summation is to use the reduction directive

Reduction operators handle the common case of summation, and analogous operations

OpenMP can manage general parallel sections

You can use ‘pragma omp single’ to have a single thread execute something

OpenMP can manage general parallel tasks

tasks now allow the parallelization of applications exhibiting irregular parallelism such as recursive algorithms, and pointer-based data structures.

Different loop scheduling may compensate for unbalanced loop iterations

Different loop scheduling may compensate for unbalanced loop iterations

Different loop scheduling may compensate for unbalanced loop iterations