Towards a Model for Shared Data Abstraction with Performance

This paper demonstrates the utility of typed shared data abstractions as an effective high-level means of structuring and coordinating parallel programs. Access to data shared by concurrent processes is expressed through operations of Shared Abstract Data Types (SADTs). SADTs abstract low-level concerns of communication and synchronisation. The exposition addresses two major challenges: mismatches between representations and actual patterns of usage; and over-specified coherence. A prototype library of SADTs provides a set of implementations for each SADT, supporting stereotypical usage patterns and allowing exploitation of weakened coherence protocols. The efficacy of this approach is demonstrated on both a network of workstationsand a dedicated massively parallel computer. A fine-grained irregular parallel computation obtained a speedup of 156 on a 256-processor Cray T3D, showing that the high levels of abstraction afforded by SADTs are compatible with efficient and scalable implementation.