A Fuzzy, Utility-based Approach for Proactive Policy-based Management Christoph Frenzel, Henning Sanneck, and Bernhard Bauer RuleML 2013, July 11 – 13, Seattle, WA, USA Software Methodologies for distributed systems
Policy-based Management (PBM) Software Methodologies for distributed systems Systems Management with Policies Challenge : increase the level of automation The system should make complex decisions guided by operational objectives, by operational objectives, e.g., maximize capacity. The system should act proactively in order to avoid problems
Classical PBM and Its Problems Software Methodologies for distributed systems eliability Problem Threshold Rate Call Setup Success R Re
Classical PBM and Its Problems Software Methodologies for distributed systems eliability Problem Threshold Sharp distinction between acceptable and unacceptable system Rate states Call Setup Success R Re • Reactive behavior • Reactive behavior Policy conflicts are resolved with complex rules that interweave technical knowledge and operational objectives • Costly maintenance
Concept of Proactive PBM Software Methodologies for distributed systems Utility-based Rule System Fuzzy Logic System ECA rule-based Policy system Replace boolean predicates (technical knowledge) with with continuous memberships utility-based conflict resolution to allow reasoning in (business objectives) inaccurate domains Fuzzification of Inference of the Defuzzification by monitoring events to value of actions selecting actions create fuzzy events based on fuzzy rules according to their indicating their weighted with value severity utilities
Concept of Proactive PBM Software Methodologies for distributed systems Utility-based Rule System Fuzzy Logic System ECA rule-based Policy system Replace boolean predicates (technical knowledge) with with continuous memberships utility-based conflict resolution to allow reasoning in (business objectives) inaccurate domains The action value represents the degree of rationality and considers: degree of rationality and considers: 1. Events the action handles 2. Utility of the events‘ treatment 3. Severity of the events Fuzzification of Inference of the Defuzzification by monitoring events to value of actions selecting actions create fuzzy events based on fuzzy rules according to their indicating their weighted with value severity utilities
Behavior of Proactive PBM Software Methodologies for distributed systems eliability Problem Threshold Rate Call Setup Success R Re
Behavior of Proactive PBM Software Methodologies for distributed systems eliability Problem Threshold Replacing sharp thresholds with fuzzy jeopardy zone Rate • Proactive behavior based on fuzzy Call Setup Success R Re event levels event levels Policy conflicts resolved by comparing the action values • Separation of technical knowledge and operational objectives
System Design Software Methodologies for distributed systems Inference of the Defuzzification by Fuzzification of value of actions selecting actions monitoring events to based on fuzzy rules according to their create fuzzy events weighted with value utilities
Fuzzification Software Methodologies for distributed systems � Event level determination › Annotate event with fuzzy event level m e › 3 KPI states: acceptable, l b l e o v r e P unacceptable, and jeopardy L y t t n i l e i b v a E › Memberships can be computed by i l e R any function provided as an event any function provided as an event specification
Fuzzy Inference Software Methodologies for distributed systems � Value computation › Fuzzy rules are technical knowledge: Event IF reliability problem IS raised Condition AND ret available IS true Action THEN action IS ret optimization WITH objective_dcr WITH objective_dcr Utility of Operator Objective Utility of Operator Objective › Objectives are defined using utilities
Fuzzy Inference Software Methodologies for distributed systems � Value computation › Combine expected utilities of rules to overall value » Domain-dependent aggregation, e.g., sum » Maximum of the rules for one objective to avoid double counting
Defuzzification Software Methodologies for distributed systems � Conflict resolution › Resolve action conflicts by selecting actions with higher value Value » Constraint optimization problem Action Action selected
Evaluation Software Methodologies for distributed systems � Scenario taken from mobile networks management › Problem situations and objectives created at random � Fuzzy, utility-based PBMS has best performance › 15% better than fuzzy PBMS › 15% better than fuzzy PBMS › 31% better than classical PBMS 0,4 Average Value 0,3 0,315 0,273 0,2 0,241 0,1 0 Classical PBMS Fuzzy PBMS Fuzzy, utility-based PBMS Classical PBMS Fuzzy PBMS Fuzzy, utility-based PBMS Fuzzy events No Yes Yes Utilities No No Yes
Conclusion Software Methodologies for distributed systems � The presented approach models … › … a Utility-based Policy System with › … a Fuzzy Logic System. � Thus, the system enables … › … automatic control of the system guided by operational objectives encoded as utilities and encoded as utilities and › … proactive actions triggered by fuzzy event levels. � In the future, we are working on… › … include observations, e.g., from ineffective actions › … modeling approach for the operator objectives & technical knowledge › … include stochastic actions and estimate their effectiveness using machine learning
Questions? christoph.frenzel@ds-lab.org Software Methodologies for distributed systems
Recommend
More recommend
