On Data Locality in Supernode Transformation

摘要

Supernode transformation is used to achieve better performance in parallel programs. Iterations are grouped together to form a tile that is executed as a unit to reduce communication overhead in supernode transformed programs. Cache issues in the supernode transformations are studied in this paper. Data accessed in a program can be regularly spaced or irregularly spaced. The spacing of the data will have an impact on the cache performance. In this paper, a method is proposed to prearrange data according to the data reference pattern of the program so that contiguous referenced data are stored in contiguous memory locations. This way, the data locality is increased and cache performance is improved for irregularly spaced data.