Research On Reliability And Failure Detection Of Distributed Storage System Based On P2P

Storage has become one of the cores IT infrastructure in information age which center is data. Data storage, which directed the network into a data-centric era, has become another wave of technology after Internet blooming. Distributed storage system based on P2P (Peer-to-Peer) provides users with large-capacity storage service by organizing idle storage resource into a logically unified view. P2P network organizes multiple nodes on Peer-to-Peer. User can make use of storage space and independently find other node to backup data and exchange storage space through this platform. P2P storage is easy to maintain, scalable, strong self-configuration features. In particular, P2P storage can be formed storage system combining with a large number of joining PC and servers. Therefore, it is greatly important to study P2P storage system.Reliability is one of the basic connotations of dependability, and it is also the key of P2P Storage Systems study. To P2P storage, reliability is seen from system standpoint and availability lays more concern on data itself. For example, when a node is temporarily offline, the stored data is temporarily inaccessible at this moment which reduces the availability of data. But data stored in the node is not loss. Reliability of the system has not changed unless temporarily off-line nodes are later found to be leave permanently. Fault detection is key technology to improve the reliability of P2P storage systems. Existing fault detections maybe use a static method, which is proofed lack of flexibility and high rate of false fault detection. That detection time fixed is of susceptible to equipment or network link failures. Another fault detection using dynamic mode need to do a certain probability of hypothesis for heartbeat information and network load, which is not meet the actual network conditions. That designing those fault detection algorithm which meet the character of P2P network and the demand of storage system is extraordinary key for constructing highly reliable storage system.The main contributions of this dissertation includes,(1) The current study of storage system reliability is based on systems which node is high available and isomerous. It generally sets up model which is revolved around relationships between the requirements of data reliability measurement and the number of copies. To address this, the relationship between redundancy and reliability is analysised by using fault tree and stochastic Petri nets (SPN) model. The paper identified the purpose and basic principles of studying system reliability from theory. P2P storage system improves system reliability through file or data redundancy. So, it is important to measurement of data redundancy, copies of data consistency maintenance, max nodes of group and availability of single node. In order to improve accuracy of data query and distribution efficiency of data copies, what structure underlying overlay network is and how to organize and manage nodes in the system all need to be addressed on reliability studies. Model analysis method can help to optimize the system design on early.(2) This paper implemented a self-feedback dynamic fault detection algorithm. P2P storage system has the basic characteristics of a distributed system. Host which is organize together through the distributed file system distributed in the WAN environment. In order to keep in touch with each other, they need to send a great of messages. Sometimes the link is not reliable. For example, a normal working nodes has been judged as failure by system. P2P storage system nodes are highly dynamic. That node temporarily frequent offline is false detection will increase the system unnecessary wastage. This paper provided a self-feedback dynamic fault detection algorithm. Experimental results show that just under the premise in a small amount increase of the calculation it reduces the false rate of fault detection and fault detection time compared to Chen's algorithm.(3) Designed a semi-structured P2P storage system architecture, which effectively reduces the problem of overlay network not matching the location of physical network. Semi-structured P2P overlay network which combines the characteristics of structured and unstructured has good scalability and manageability. This paper assigns group according to the response time of detection message and selects supernode according to the reputation value which is computed through combining on-line rate and response time of node. The results of simulation show that the designed structure reduces join and exit number of messages each node sent and the search delay of information. It achieved good customer satisfaction.(4) Implemented a supernode elect services which included four basic components: The registration center module, group membership maintenance module, fault detectors and supernode election algorithm. Supernode is the most important part of the semi-structured P2P storage systems. On the one hand, supernode needs to manage the common node, as well as to respond to a user's data query request; On the other hand, supernode also needs to maintain contact with the outside world, which play a supportive role to global data transfer and routing. Registry is responsible for node joining and leaving. Group membership maintenance module is responsible for the management of node and information table maintenance. Fault detector is responsible for monitoring node status and collecting historical information. A supernode election algorithm is responsible to re-select a new node as a super-node when the super-node fails. The results show that average service restoration time and average error rates of a supernode election service presented by this paper are less than Toueg's S1, and the availability of service close to or slightly lower than the S3. It achieves the desired effect of the experiment. In addition, this realization of the supernode service mechanism can be deployed flexibly according to user needs and can provide convenience for the future implementation of solutions.