Data parallel language extensions for exploiting locality in irregular problems

Abstract

Many large-scale computational applications contain irregular data access patterns related to unstructured problem domains. Examples include finite element methods, computational fluid dynamics, and molecular dynamics codes. Such codes are difficult to parallelize efficiently with current HPF compilers. However, most of these problems exhibit spatial locality. This property is exploited by our approach.

In the sequential program, unstructured domains are accessed via indirection arrays. We introduce a new directive that serves to identify indirection arrays and the boundaries of the associated domains. The data domains are distributed using Multiple Recursive Decomposition (MRD), a pseudo-regular distribution, which combines efficient implementation with good load balancing and communication behavior. Indirection arrays are aligned with the data arrays. Using the information provided in the directive, the compiler can produce a target program with significantly better performance than an approach based on indirect distributions and the inspector/executor paradigm.

This work was supported by the Ministry of Education and Science (CICYT) of Spain under project TIC95-0942-C03 and by the Human Capital and Mobility program of the European Union (ERB4050P1921660)