Abstract

Fault-tolerance is an important requirement for long running parallel applications. Many approaches are discussed in various literatures about providing fault-tolerance for parallel systems. Most of them exhibit one or more of these shortcomings in delivering fault-tolerance: non-specific solution (i.e., the fault-tolerance solution is general), no separation-of-concern (i.e., the application developer's involvement in implementing the fault tolerance is significant) and limited to inbuilt fault-tolerance solution. In this thesis, we propose a different approach to deliver fault-tolerance to the parallel programs using a-priori knowledge about their patterns. Our approach is based on the observation that different patterns require different fault-tolerance techniques (specificity). Consequently, we have contributed by classifying patterns into sub-patterns based on fault-tolerance strategies. Moreover, the core functionalities of these fault-tolerance strategies can be abstracted and pre-implemented generically, independent of a specific application. Thus, the pre-packaged solution separates their implementation details from the application developer (separation-of-concern). One such fault-tolerance model is designed and implemented here to demonstrate our idea. The Fault-Tolerant Parallel Architectural Skeleton (FT-PAS) model implements various fault-tolerance protocols targeted for a collection of (frequently used) patterns in parallel-programming. Fault-tolerance protocol extension is another important contribution of this research. The FT-PAS model provides a set of basic building blocks as part of protocol extension in order to build new fault- tolerance protocols as needed for available patterns. Finally, the usages of the model from the perspective of two user categories (i.e., an application developer and a protocol designer) are illustrated through examples.