Gerência de redes distribuída e confiável baseada em grupos de agentes

Abstract

Network management systems are essential when parts of the network are non-operational. Particularly, fault management applications must be able to work correctly in the presence of network faults. Access to the management data of a crashed or unreachable network element may help to determine why it is faulty. However, most network monitoringsystems only allow the examination of managed objects of fault-free agents. This work presents a strategy for the construction of highly available network management systems. The strategy employs data replication, a distributed and hierarchical organizational model,and the clustering approach, which allows a logical division of networks, in order to reduce the overhead of messages exchanged among network elements. The rst contribution of this thesis is the denition of an agent clustering architecture for object replication. The architecture is structured in three layers. The lower layer corresponds to typical agents at the network elements, which keep management objects at their local MIB's (Management Information Base). The middle layer corresponds tomanagement entities called cluster managers that have the task of monitoring agent's managed objects and replicating them in other clusters. The upper layer corresponds to the manager entity that denes each cluster of agents as well as the relationship among clusters. A cluster of agents provides fault-tolerant object functionality. In this way,replicated managed objects of a crashed or unreachable agent that belongs to a given cluster may be accessed through its cluster manager or one of its peer cluster managers. The second contribution of this thesis is an SNMP agent clustering framework for the Internet community. This SNMP framework describes a set of management objects that supports the replication of managed objects. The MIB called ReplicMiB specieshow to dene cluster members, replicated objects, and peer clusters of a given cluster. Furthermore, it introduces the compliance statements for the SNMP manager and cluster manager entities, i.e. which management objects need to be implemented in these SNMPentities. An example of the framework usage is introduced along with the description of the MIB objects.The third contribution of this thesis is a fault management tool based on the SNMP agent clustering framework. The tool extends the functionalities of SNMP agents to object replication and enables the access to management data replicated in the fault-free SNMP agents. The tool was built using the NET-SNMP package and the Ensemble group communication toolkit. Changes in the internal structure allow the SNMP agents to play the role of cluster managers. A group application called mcluster provides the insfrastructure for reliable communication among cluster managers and ensures the consistency of replicated managed objects. An extensive evaluation of the tool deployed at a local area network was carried out. The evaluation consisted of a resource consumption analysis, a performance analysis, and a brief study of the availability of managed objects in failure situations.