Dynamic Optimization Of Performance And Energy Consumption In Cloud Computing

The enormous energy consumption and carbon emissions of datacenters have become a pressing issue for cloud service providers seeking low operational costs and sustainable development. With the rapid advancement of cloud computing (CC) technologies, cloud services are no longer exclusive to desktop users, but available to mobile users as well. A new computing paradigm called mobile cloud computing (MC2) hence emerges. In MC2, mobile devices and clouds are connected via mobile networks, and energy-efficient data transmission between these two ends has become a key issue. This dissertation aims to optimize the performance and energy consumption of CC and MC2. Dynamic optimization theory has been applied to modeling the energy-conserving mechanisms and designing online control algorithms for optimizing both energy consumption and performance. The main contributions of this dissertation is summarized as follows:(1) Energy and performance optimization of resource auto-scaling in cloud datacenters. First, the Markov decision process (MDP) theory is employed to formulate the resource auto-scaling problem. In light of the absence of an environment model, a resource auto-scaling algorithm (RASA) is then proposed using reinforcement learning and approximate dynamic programming (ADP) approaches. The RASA algorithm utilizes the dynamic frequency scaling mechanism of server CPUs to match workload and resource allocation, thereby optimizing datacenter energy consumption and user performance. Simulations are conducted to verify the effectiveness of the RASA algorithm.(2) Cost-optimal request routing and virtual machine scheduling in geo-distributed SaaS clouds. First, this dissertation formulates the cost and performance management problem of the geo-distributed SaaS clouds as a discrete-time stochastic dynamic program (DT-SDP). The DT-SDP aims to minimize the time average cloud operational costs including energy cost, carbon taxes and bandwidth cost, subject to request queue stability constraints. Then, this dissertation takes advantage of Lyapunov stochastic optimization techniques to design a distributed online algorithm called GREEN. The proposed GREEN algorithm provably achieves arbitrary tradeoffs between cost optimality and queue stability. Real-world trace-driven simulations prove the efficacy and adaptivity of GREEN in non-stationary environments.(3) Throughput and energy optimization of link selection and data transmission scheduling in MC2. First, the uplink and downlink data transmission between mobile devices and clouds is formulated as a SDP. Then this dissertation uses ADP to design an online learning algorithm called eLean. Taking advantage of different links’diversity in energy efficiency and the delay-tolerance property of some mobile applications, the eLean algorithm optimizes both energy consumption and throughput of mobile devices via dynamic link selection and data transmission scheduling. Finally, simulations are conducted to validate the convergence of eLean, and demonstrate its superior energy-efficiency over alternative link selection algorithms.