Coarrays

Coarrays are a language feature used to express a partitioned global address space for parallel programming, enabling direct remote access to data stored on different processing elements called images. In coarray-enabled languages, a variable can be declared with an attached coarray dimension, so each image holds its own local data plus a global view of corresponding data on other images. This arrangement allows one-sided communication and shared-memory-like access patterns without explicit message passing.

In practice, coarrays rely on a global address space that is logically distributed across images. Each image executes the same program, and synchronization primitives coordinate operations across images as needed. A typical Fortran example declares a coarray as real :: A(10)[*], creating a 10-element real array whose elements are replicated across all images. Access to a remote element uses a syntax such as A(i)[j], addressing the i-th local element on image j. Intrinsics such as this_image() and num_images() reveal the current image’s index and the total number of images, respectively. Synchronization constructs, like sync all and sync images, ensure memory consistency and orderly progress across images.

Coarrays support both data-parallel and task-parallel styles by combining local computation with remote data access. They are a central feature of the partitioned global address space (PGAS) programming model and are most closely associated with Fortran 2008/2018, though similar concepts appear in other PGAS languages. The approach aims to simplify scalable parallel programming by providing direct access to distributed data while preserving explicit control over synchronization and data locality. Potential benefits include simpler programming models for distributed memory systems, balanced with considerations about data locality and contention.