A Compiler Driven Execution Model for Irregular Applications

Current parallelizing compiler technology does not handle irregular applications effectively. Approaches in the past have either focused on specific irregular applications to generate good code, or relied on fall-back generic techniques that are less efficient. This work proposes a compiler-driven approach that makes use of a novel runtime execution model for a cluster of computers. The model consists of data-flow style workers to enable loadbalancing, and HPF-style data-blocking (or tiling) to divide computations into units of work. The run-time system is loosely based on Space-Time Memory abstraction developed at Compaq’s Cambridge Research Lab, which enables blocks to be addressed totally independently of their physi-