Towards the enrichment of parallel systems skeletons

点击翻译按钮获取中文摘要

Despite the advancements of hardware technologies and consequently the availability of low-cost commodity clusters and high-performance computers, complexity of parallel application development remains one of the major obstacles. As one of the techniques for easing this complexity, researchers are actively investigating the patterns-based approaches to parallel programming. As re-usable components, patterns are intended to ease the design and development phases of a parallel applications, as compared to the traditional approaches like MPI, PVM and sockets. PAS (Parallel Architectural Skeleton) is one such patterns-based parallel programming model and tool, originated at the University of Waterloo and currently being actively researched at the Concordia University, which generically defines the architectural aspects of parallel computational patterns. Ongoing research has contributed to an extensible PAS (EPAS) model and system that supports extensibility and skeleton composition, enabling design and incorporation of new skeletons and creation of new composite skeletons from existing repertoire. In the course of this research, some of the other significant requirements and research issues for PAS have been identified, better understanding and incorporation of which could significantly enhance its usefulness and appeal to the parallel computing community. More specifically, the original intention of PAS (and most other patterns-based approaches) was towards programmability. However, being pure architectural abstractions which are decoupled from an application and programmer, PAS could encapsulate many useful systems-specific requirements which are otherwise rather difficult and time-consuming to implement from scratch. The proposed research in this direction addresses the following: (i) Dependability: dependable skeletons with reliable composition and communication-synchronization abstractions; pattern-specific strategic fault-tolerance; support for avoidance/detection of possible atomicity violation; maintenance.(ii) Performance: adaptable and re-configurable architectural skeletons; architectural skeletons with cost-performance models and support for performance profiling and re-structuring.