OpenMP

OpenMP is an API for shared-memory parallel programming designed to help developers create scalable applications for multi-core and many-core processors. It targets C, C++, and Fortran and provides a portable model based on compiler directives, runtime library routines, and environment variables to annotate serial code with parallel constructs.

Key features include parallel regions and work-sharing constructs (for, sections, and tasking), data scoping rules (private, shared, firstprivate, lastprivate, reduction), and synchronization mechanisms (barrier, critical, atomic, flush, ordered). In addition to traditional thread-based parallelism, OpenMP supports task-based parallelism (tasks and taskloop) and SIMD-style parallelism through explicit SIMD directives. Scheduling options such as static, dynamic, and guided control how iterations are distributed among threads.

The runtime manages a team of threads within a process, typically using a thread pool that is created at the start of a parallel region and reused across nested regions. Programs can query and control runtime behavior with functions like omp_get_num_threads, omp_get_thread_num, and omp_get_wtime, and influence execution with environment variables (for example OMP_NUM_THREADS and OMP_SCHEDULE).

OpenMP is standardized by the OpenMP Architecture Review Board and has evolved through versions from 1.0 onward, with major additions such as robust tasking and broader SIMD integration in recent releases. It is widely supported by compilers such as GCC, Clang, Intel, IBM, and others, and is commonly used in conjunction with MPI for hybrid shared-memory and distributed-memory applications.

Usage considerations include avoiding data races, managing false sharing, and weighing overhead against parallelism granularity. A typical start involves compiling with OpenMP support (for example, -fopenmp in GCC) and annotating code with parallel directives. Example: #pragma omp parallel for reduction(+:sum) for (int i = 0; i < N; ++i) sum += a[i] * b[i];