Research Of Dependability Evaluation Of Cluster System

This thesis is focused on the research of the models, approaches and tools of dependability evaluation in computer cluster systems.First, the paper introduces a suitable set of dependability measures for clusters, and describes the formulation of three classes of measures, including system availability, reliability and task completion. The applicability and results of different measures are given and compared through the analysis of a two-node cluster system.Then the paper discusses the theories of cluster dependability modeling. The input parameters such as failure rates, coverage probabilities and recovery rates are investigated with respect to their characteristics and estimation methods. Reliability Block Diagram, Fault tree, Markov Chain and Petri Nets models, commonly used for dependability analysis, are reviewed and their modeling capability and solution complexity are discussed. The main work of the paper is applying Relaibility Block Graph (RBD) to the evaluation of basic cluster dependability, presenting an approach for the dependability analysis of high-availability (HA) clusters, designing a simulation tool NCPN for Coloured Petri Nets (CPN) and a HA cluster dependability evaluation tool ACUTE, and finally studying an instance of HA cluster.Using RBD modeling, the basic dependability of clusters is analyzed. The paper chooses the DWANING 4000a cluster as instance and evaluates its basic dependability measures, which show that the MTTF of a 262-node cluster is only one day.The paper presents an approach for the dependability analysis of HA clusters. Based on Coloured Petri Nets, the approach uses hierarchical modeling to design a submodel library for the characterization of cluster behaviors. The library supports major Peer-to-Peer and Centralized HA cluster architecture, and forms the global model by dynamic modeling. In order to solve the dependability models based on Coloured Petri Nets, a simulation tool called NCPN is developed. The object-oriented mechanism and event-driven simulation are applied in NCPN. A class library is provided to model a wide range of different structures and behaviors of CPN. Benefits of NCPN such as hierarchical modeling and CPN structure extensions are mentioned.To demonstrate the use of ACUTE, a detailed dependability analysis is carried out on a full-system example representative of HA clusters. The system is composed of sixteen nodes, and running several instances of a computing task and Oracle database service. The parameters are retrieved by history profiles and simple fault injection experiments. The paper discusses the improvement of task-completion measures due to checkpoint mechanism, and...