Research On Parallel Computing Universal Programming Model And System Architecture

Data and Information is enormously valuable for all industries. However, it is quite difficult to process massive data and extract useful information in time. In the past decade, the amount of data has grown rapidly along with the enhancement of informationization in many industries. In order to meet the demands of massive data processing and analyzing, the parallel computing has been introduced into more and more fields. In the past five to eight years, the research and application of parallel computing programming model has been extended from professional field to many information industries, such as IT and e-commerce.It is not a novel technology to process data in parallel. The concept of parallelization has been put forward for decades. In many professional fields, there is a long history of research on parallel computing and has got many achievements. However, as the application environment has been different, the scenario and demands has changed greatly. So far, there is a lack of research achievement of general parallel computing programming model and system architecture based on cluster.Along with the popularization of parallel computing technique in more and more industries, the demands of general parallel computing programming model are increasing. This is an opportunity as well as challenge for researchers. In recent years, the research of general parallel computing has taken shape, and a great amount of general programming models have been proposed as well as related implementations, such as MapReduce, Dryad and so on. However there are still lots of issues need to be investigated, such as follows:(1) The commonality of programming model and related implementation:Most of the programming models and related implementations are designed to meet certain kind of demands, and only a few kinds of data analyzing jobs can be processed directly and efficiently. Meanwhile, as processing modes are integrated in the implementations, the program designs are lack of flexibility.(2) The scalability of system:The general parallel computing systems are built on large-scale clusters in usual. However, the system architecture is lack of scalability. As the cluster rapidly expanded and the task size constantly increased, the control unit has been hard to manage the system.(3) The design demands of common framework:Although the separation of application management and resource management may enhance the commonality of clusters, it has little benefit to specific programming model. If general process of application management can not be abstracted, the usage of common framework will be limited to resource management.(4) The applicability of programming model:As the general parallel computing has done very well in massive data processing, more jobs are trying to be done in parallel. However, the parallel processing is not suitable for all jobs. We should take account of the applicability of the programming models.This thesis focuses on above issues and carries out the following researches:(1) The research on general parallel computing programming model and system. We analyzed the existing models and summarized the advantages and shortages. Then a new programming model was proposed which integrated three parallel modes on different logical layers, including function parallel, time parallel and data parallel. Meanwhile, we redesigned the architecture of related system. As the process control of application is a special task in the new design, the program design is more flexible. Such design will enhance the commonality of programming model and related system.(2) The research on the scalability of general parallel system. Through analyzing the reasons causing bottleneck in scalability, we proposed a new system architecture with distributed multiple masters. In the new system, the distributed management replaced the centralized management; the resource signaling and related processing mechanism were optimized; the managements of concurrent applications would be assigned to different masters. All these new designs may enhance the scalability of system.(3) The research on common framework. We proposed a common sustainable scalable framework for general parallel computing system. In the new framework, we not only took account of the scalability of system, but also considered the demands of general parallel computing programming model. In the new framework, we separated application management from resource management, and used hierachical structure to construct scalable management modules. Meanwhile, general functions of application management were integrated into the framework, and the definition of specific processing flow could reload from different programming models.(4) The research on the application of general parallel computing is focus on processing data of network states in parallel.We analyzed general parallel computing system and summarized the characteristics of the problems which could be processed in parallel. We also designed a parallel algorithm to solve the problem of traffic congestion. The parallel algorithm might accelerate the process by using parallel system. The target of this reseach is to find a way of analyzing network states in parallel.