Research On Technologies Of Adaptive Failure Detection In Distributed Systems

Reliable and stable distributed system is used widely in many fields such as military,medic and finance.However,the larger scale and complexity the system becomes,the more likely system breakdown happens.As one of the most basic components for reliable operation of distributed systems,failure detection has important research significance.Aiming at achieving the self-adaptability and expansibility of failure detection,this thesis discuss the adaptive detection protocol based on link fault and the failure detector which can auto adjust itself base on the system and network state.Heartbeat detection is the most widely used technique in failure detection of distributed system.Based on EMA(Exponential Moving Average)and variance ratio,this thesis presents a new adaptive failure detector,DEMA-FD,which is not only more accurate than the traditional heartbeat predictive detector,but also can be adjusted according to the basic evaluation index of the failure detector’s QoS(Quality of Service)to meet requirements of different distributed applications of failure detection.It is proved theoretically that we can implement a failure detector of ◇P-type using DEMA-FD in some synchronization systems.At the end of this thesis,I verified the DEMA-FD algorithm using some experiments.The traditional failure detector simply equates the link failure with the node failure problem,which greatly affects the fault locating and fast repair.This thesis presented a new adaptive detection protocol DLFDA.According to DLFDA,each node has k detectors and detect multiple objects to diagnose the fault type.This protocol can accurately distinguish the fault between link and node.The DLFDA protocol uses a new DA-FD as a direct detection algorithm which uses the center-beat of the DEMA-FD detector to predict algorithm’s adaptability and outputs a time-accumulated decision value based on the exponential function.In the same time,user can adjust the intensity of detection by setting flexibly the threshold.In addition,DLFDA protocol can dynamically adjust the detection structure to increase the coverage of the link detection and use the gossip protocol to release fault diagnosis,reducing the detection load.Finally,the experimental results of DLFDA protocol show that the presented protocol meets the requirements of theoretical design.At the end of this thesis,we design a prototype of a generic and extensible distributed adaptive failure detection system which has three layers including member management,information synchronization and adaptive failure detection.The hierarchical function of the system and the low coupling among the modules can be quickly extended through a unified interface.