A Failure Detection Solution For Multiple QoS In Data Center Networks

As the core infrastructure of cloud computing and virtualization,a data center is the center of data transmission,computing and storage,which integrates multiple hardware and software resources.Large-scale data center networks are very complicated because they can contain up to hundreds of thousands devices and links.Failures in data center networks are very common,these failures of devices or links can lead to user-perceived service interruptions.Thus automated failure detection plays a very important role in maintaining the reliability of data centers.Many researches have been conducted on failure detection.Among these researches,some of them focus on the failure detection in the environment of distributed systems,they try to design a scalable failure detection architecture in the application layer without considering network topology;others try to solve the problem of failure localization in networks.However,these researches rarely identify that there exist different failure detection quality of service(QoS)among different kinds of devices.Aiming at the shortage of previous work,this thesis proposes a failure detection solution for data center networks,which can provide different failure detection QoS for different kinds of devices respectively.First of all,this thesis divides network devices into two categories: realtime devices and non-realtime devices.The main difference between them is that realtime devices have strict timing constraints on QoS.Subsequently,this thesis concludes the failure detection QoS for these two kinds of devices respectively.In order to ensure the QoS,this thesis proposes a codetection approach named K-detectors to detect the failures of realtime devices,K-detectors leverages multiple detectors to enhance the reliability and fault tolerance of detection;for the failure detection of non-realtime devices,this thesis analyzes network traffic changes over time and utilizes a Feedforward Neural Network based approach to compute failure probability for each device.At last,to evaluate the effectiveness of the proposed approach,this thesis sets up a simulated network with ns-3 and implements the proposed work and all the other related work.The experimental results show that the failure detection QoS behind various experimental metrics improves more or less,for both realtime devices and non-realtime devices.