Agent based Management of Distributed Systems with Variable Polling Frequency Policies

Abstract

Management activities are based on the state of distributed system components, relations of these components, and their behaviour. Since management policies are applied across an abstraction of distributed systems, the quality of decisions is dependent on the representation fidelity of the real system state. Obviously, the data collection process updating the abstract representation of real distributed system components has a major impact on the quality of management decisions. Gathering most topical management data improves the quality of management decisions, but requires a high degree of monitoring activity. This is contradictory to the request for low impact management systems, where the amount of system resources used for management purpose should be as small as possible. In this paper we present a twofold approach to this problem: First a high level management architecture is described where monitoring is performed by distributed agents with generic functionality for filtering and event creation. The distribution of active management agents reduces the amount of management related traffic and avoids a potential bottleneck on a centralized management station. Second an adaptive polling frequency approach is presented which enables the monitoring agents to adapt their polling frequency automatically to different behavioral parameters of managed components. The automatic adaptation reduces the performance impact of the agents significantly while at the same time a high accuracy of management relevant information about critical components is ensured. Implementation aspects of the introduced management architecture in a CORBA environment are also discussed.

Policy-based Management: A Historical Perspective

“…The language is defined using BNF and provides ancestor and descendant links in the definition of a policy in order to link the policy refinement hierarchy. The framework which is implemented in [72] features also the use of monitoring agents and a tool for the graphical representation of chained policies.…”

A scalable monitoring approach based on aggregation and refinement

“…A common approach to reducing monitoring overhead is to vary the polling frequencies based on the state and characteristics of variables being monitored. References [5], [8], and [17] present different approaches to how the polling frequencies can be varied. However, no data aggregation is performed, which limits the overhead reduction achievable.…”

“…Poll the agent for the exact value of flows in ; 3 else 4 Let be the set of segments in each of which contains at least one flow in , ; 5 Ask the agent to refine the segments in some set , where if . The strategies for choosing from in Step 5 vary.…”

“…Similarly, in round 3, flows 6, 7, 9, 11,m and 12 are refined. Finally, at the end of round 3, the ordered list based on upper bound values is [5,9,7,6,1,2,3,4,11,12,13,14,15,16,10,8]. Three more flow delimiter indexes are identified, , , and .…”

Deriving Variable Polling Frequency Policies for Pro-active Management in Networks and Distributed Systems

“…However, we consider that a flexible correlation of the polling frequency with behavioural parameters of a DS components can substantially reduce the management traffic. These behavioural parameters are presented in details by Dini, Bochmann and Boutaba (1996) and Dini (1996), and briefly described in Dini, Bochmann, Koch, and Kramer (1997).…”