Unified framework for sparse and dense SPMD code generation(preliminary report)

Abstract

We describe a novel approach to sparse {\em and} dense SPMD code generation: we view arrays (sparse and dense) as distributed relations and parallel loop execution as distributed relational query evaluation. This approach provides for a uniform treatment of arbitrary sparse matrix formats and partitioning information formats. The relational algebra view of computation and communication sets provides new opportunities for the optimization of node program performance and the reduction of communucation set generation and index translation overhead.