Çokişlemcili sistemlerde enerji yönetimi için rastsal modeller

In this thesis, we address the energy management problem that has become animportant concern for computing systems. We concentrate on dynamic voltage scalingenabled multicore systems running soft real-time applications that have been decomposedinto concurrent tasks.We first illustrate on simple examples that how energy management ignorant ofthe statistical nature of task workloads can be significantly suboptimal or not properlyfulfill timing constraints. Based on this observation, we propose stochastic models whichcapture the variability of, and correlations among the workloads of tasks. The stochasticmodels are constructed using the workload data collected from actual application runs.Each stochastic application model suggests an energy management policy whichminimizes the average energy consumption of the system while satisfying the timingconstraints according to the application's tolerance to deadline misses. These policies,which exploit the variability and the correlation information captured by the stochasticapplication models, are obtained by novel optimization formulations. To solve theformulated optimization problems, we present two solution techniques. Our energymanagement policies introduce negligible overhead at runtime, since the associatedoptimization problems are solved offline.We show on a popular multimedia application, MPEG2 video decoding, that therecan be significant variability in and correlations among the workloads of tasks. Using thesame application, we demonstrate how stochastic application models can be constructedand then present experimental results. We compare our energy management policies withtwo other approaches proposed in the literature.