A high-level framework for efficient computation of performance - energy trade-offs in Markov population models

Internet scale applications such as search engines and social networks run their services on large-scale data centres consisting of tens of thousands of servers. These systems have to cope with explosive and highly variable user demand and maintain a high level of performance. At the same time, the energy consumption of a data centre is one of the major contributors to its operational cost. This embodies the performance-energy trade-off problem. We need to find configurations which minimise the energy consumed in running important applications in complex environments, but which also allow those applications to run reliably and fast. In this thesis, we develop a general performance--energy analysis framework that can be used to express complex behaviour in communicating systems and provide a rapid analysis of performance and energy goals. It is intended that this framework can be used both at design time to predict long-run performance and energy consumption of an application in a large execution environment; and at run time to make short-term predictions given current conditions of the environment. In both cases the rapid model analysis permits detailed what-if scenarios to be tested without the need for expensive experiments or time-consuming simulations. The major contributions of this thesis are: (i) development of the Population Continuous-Time Markov Chain (PCTMC) representation as a low-level abstraction for very large performance models, (ii) development of rapid ODE analysis techniques to compute performance-based Service Level Agreements (SLA) and reward-based energy metrics in PCTMCs, (iii) hybrid extension of PCTMCs that allows models to incorporate continuous variables such as temperature and that permits the specification of systems with time-varying workloads, (iv) an extension of the GPEPA process algebra that can support session-based synchronisation between agents and that can be mapped to PCTMCs, thus giving access to the rapid ODE analysis. We support the framework with a software tool GPA, which implements all the described formalisms and analysis techniques.Open Acces