iModel

Abstract

Deriving analytic performance models requires detailed knowledge of the architecture and behavior of the computer system being modeled as well as modeling skills. This detailed knowledge may not be readily available (or it may be impractical to gather) given the dynamic nature of production computing environments. This article presents a framework, called iModel , for automatically deriving and parameterizing analytic performance models for multi-tiered computer systems. Analytic performance models consist of a workload model and a system model. iModel uses system logs and configuration files to generate a high-level characterization of the system; e.g., open queuing network (QN) model versus closed QN model. By harvesting more information from the system logs and configuration files, iModel generates a workload model by inferring user-system interaction patterns in the form of a Customer Behavior Model Graph (CBMG) and generates a system model by discovering system components and their interaction patterns in the form of a Client-Server Interaction Diagram (CSID). iModel includes a library of well-known single-queue and QN models and their solutions stored in an XML-based repository. The generated workload model and system model are compared to the model repository to determine which model in the repository best matches the system’s observable behavior and architecture. This article also presents a black-box optimization approach that is used to derive analytic model parameters by observing the input-output relationships of a real system. This optimization approach can be used in any computer system (multi-tier or not) that can be modeled by single queues or QNs. The important question is whether the automatically generated and parameterized performance model has predictive power, i.e., can the derived model predict the output values that would be observed in the real system for different values of the input? The results presented in this article demonstrate that the analytic performance models derived by iModel are relatively robust and have predictive power over a wide range of input values.