Research Of Multi-level Parallelism Programming Pattern For Hybrid Parallel Computing Environment

Parallel computing becomes more and more popular because of two important technology renovations and people's endless demands for computing power. Especially the CLUSTER system has become popular. CMP-CLUSTER is one kind of CLSUTER and involves distributed memory and shared memory parallel architectures. So it is referred to a hybrid parallel computing system, HPCS. However, it is very difficult to programming in this HPCS environment because there is hierarchical parallel structure including inter-node, inter-CMP-socket and intra-CMP-socket three-level parallelism. Therefore, it is a hot area of research to provide specific methods of parallel programming, programming models, environments and tools to simplify parallel programming in HPCS environment. The work of this thesis is done in this context and the main work including:This dissertation analyzed and proposed the hierarchical communication model and memory model of the CMP-CLUSTER system. For the communication model, the dissertation pointed out the operating costs and impacts on every layer of hierarchical model of communication. For the memory model, the dissertation analyzed the shared cache architecture feature of CMP system. And then the method how to programming in the CMP-CLUSTER environment was given and a suitable CMP-CLUSTER parallel processing pattern was proposed with the aim to put forth the respective advantages of the each parallel architecture in the CMP-CLUSTER system.According to the features of the conceptual model of CMP-CLUSTER, this dissertation proposed a top-down hierarchical and gradually refined parallel programming model which had five layers: task allocation pattern layer, task computation pattern layer, design-pattern based parallel programming framework layer, middle framework code layer, task deployment and mapping layer. And each layer has a counterpart in the CMP-CLUSTER parallel processing pattern.Some aided methods had been studied and provided to each layer in the CMP-CLUSTER hierarchical parallel programming model in order to simplify parallel programming and improve programming efficiency and reduce errors introduced by manual factors. Two pattern-based parallel programming frameworks had been designed and implemented. Users can directly invoke the high-level application program interface which the framework system provided, and the other work such as data division and parallel computing will be done by the framework system automatically. A array-based data parallel design pattern FJRR and a task interaction graph based parallel design pattern had been proposed and implemented through the parallel programming framework. They can greatly reduce the complexity of parallel programming.