01-02-2013, 04:36 PM
The Vision of Autonomic Computing
ABSTRACT
An autonomic computer system is a system which is, to a certain degree capable of self-management. As most of today’s computing infrastructures require high availability, the need for self-managing systems becomes apparent when fault resolution is required. System administrators need to be able to resolve system problems quickly despite the complexity of the systems they manage. Having systems that can provide pointers to possible causes/solutions is therefore desirable.
For a computer system to be capable of self-management; it needs to possess four attributes. These are self-awareness,self-situation, self-monitoring and self-adjustment. Self-awareness refers to the system’s ability to be informed about its internal state. To achieve self-awareness, the system needs to be able to differentiate between its different internal states.
Previous work has leveraged on either the collection of system performance metrics or system logs as sources of data, which could be analyzed by a system to make it aware of its internal state. System logs are usually good indicators of system state as they contain reports of events that occur on the several interrelated components of complex systems. To use system logs to characterize internal state, it is necessary to compute the correlated message type sequences in the log.
Message types are textual templates, which abstract the natural language messages in system logs. In this work, we explore the use of entropy-based information content clustering of system log partitions as a means of discovering system state. Such a system log partition will contain log data from a single source on the network over a unit period of time.
Our results show that such partitioning of system logs leads naturally to partitions, which contain correlated message types, a previously unknown property of system logs. We demonstrate how these correlated message types can easily be discovered through the grouping of the log partitions into conceptual clusters using their entropy-based information content scores. Conceptual clusters are clusters where objects in a cluster can be described by a concept, not just based on their distance from each other.