Approximate Mean Value Analysis based on Markov Chain Aggregation by Composition Abstract.pdf

Approximate Mean Value Analysis based on Markov Chain Aggregation by Composition Dorina C. Petriu, C. Murray Woodside Department of Systems and Computer Engineering Carleton University, Ottawa, Canada, K1S 5B6 email: {petriu|cmw}@sce.carleton.ca Abstract Markovian performance models are impractical for large systems because their state space grows very rapidly with the system size. This paper derives an approximate Mean Value Analysis (AMVA) solution for Markov models that represent a com- position of subsystems. The goal is robust scalable analytical approximation. The approach taken here is to create approximate aggregated Markov chain submod- els, each representing a view of the Markov chain for the entire system from the perspective of a selected set D of tagged components, and to derive mean value equations from them. The analytic solutions of submodels are then combined using system-level relationships, which must be identified for each system; this is not au- tomatic but is usually straightforward. The first point of novelty is the method used to create the aggregate submodels for different sets D, building up each submodel by composition of the components in D rather than by aggregating the entire state space. Another point of novelty is the use of partitioned Markov models to obtain analytic solutions. Key words: Performance models, Software performance, Compositional modeling, Markov Chains, Aggregation by composition, Mean Value Analysis 1 Introduction Markovian performance models based on system states and transitions are impractical for large systems because of the very rapid increase of the state space with system size, also known as state explosion. Different approaches have been identified to circumvent state explosion, such as: ? hierarchical decomposition into smaller submodels, linked by a high-level model, or by coordination relationships at their boundaries; the solution is iterated among the submodels Preprint submitted to Elsevier Science 24 F