Hybrid analytical-simulation model for performance prediction of distributed systems

Abstract

Distributed systems are widely adopted as a response to the increasing demand of computational resources and also to facilitate integration of remote components. Such systems are at risk of being useless, since design flaws often lead to critical performance issues at deployment, when usually they are too costly to fix. Software Performance Engineering is concerned with modeling and prediction of system non-functional properties early in the software development process, leading to informed design decisions. This thesis proposes a new hybrid analytical-simulation approach regarding performance model solving. The UML model of the analyzed system is transformed into a performance model, whose structure resembles a simulation model, hierarchically decomposed into layers, in order to be able to address submodels. Simulation results for these submodels are later used in the higher level submodels by the analytical solver. In case hierarchical decomposition is not possible, a sequential breakdown is implemented. Method evaluation is performed with a tool called PHYMSS (Performance Hybrid Model Solver and Simulator). This tool is developed by the thesis author in C# and allows hybrid analysis, pure simulation and pure analytical calculus. The pure simulator uses an improved multithreaded version of a coroutine-based simulation model, model which is extended to support nested calls, additionally to sequential actions. Several case studies are described, in order to prove the efficiency of the new hybrid model and solver, compared to pure analytical and simulation approaches. Where it was possible, measurements were performed on a distributed online system to provide more confidence in the analysis results. The method proved to converge much faster than simulation and to yield more accurate results than analytical calculus.