Data abstractions for portable parallel codes

Development of parallel software is a very complex task. Many details, such as domain type, partition method, or mapping scheme, should be considered during the implementation phase. However, they are usually taken into account in the design phase because the developers do not have the appropriate abstractions. This causes that small conceptual changes, for example a modi-fication in the distribution policy, imply the developing of an application from scratch. The result is that it is not always possible to obtain portable and efficient parallel codes. In this paper, we present a programming methodology that decouples data representation, partitioning, and mapping from the algorithmic and parallel strategy decisions. This allows to define explicitly all the design details with the appropriate abstractions. Postponing the low level parallel decisions to the implementation. To test this methodology, we have developed a library that offers these abstractions allowing to generate highly portable code. Our experimental results show that the use of this library allows to achieve similar performance as carefully-implemented manual code while reducing programming effort