Simulator Prototyping Through Graphical Dependency Modeling

Abstract

Given a complex system, simulating the behavior of the system under various conditions and inputs is a common requirement in different domains. Designing a simulator by non-domain expert is not an easy task. However, the vast amount of performance related parameters that can be monitored on any physical system can help in probabilistic modeling of the system. Accurate modeling of a complex system requires the identification of dependency among the individual components. These dependencies can be deduced by probabilistic analysis of the performance parameters generated by the physical system. We propose constrained Hill Climbing for structure learning of a Bayesian Network to learn the dependency among internal components of the system. Dependency graph, thus generated will act as a basis for developing a prototype simulator. Multivariate Adaptive Regression Splines are compared with the proposed constrained hill climbing to deduce the mathematical relation between the interdependent components. Two prototype simulators are designed, one for a complex large-scale storage system and another for a group of computer servers. Behavior of both the physical systems is tested on workload traces. The normalized mean absolute error for simulation of storage systems and server was 4.6% & 3.75% respectively. Results indicate that the proposed simulation prototyping can be a useful and unique way of understanding complex systems without the help of a domain expert.