Research On Power Control And Performance Optimization For Large Scale Network Service Systems

As the demand on internet services dramatically increases, the number and scale of internet service provide systems have been skyrocketing, as is the electricity con-sumption. Higher electricity consumption not only results in boosted operation costs, but also restricts the system performance improvement and produces more carbon e-mission. There is no doubt that energy consumption of network service systems is a pressing factor in system design both from an economic and environmental point of view. Of all the internet services, multimedia application has been a dominant source of internet traffic and still keeps rapid growth. Hence, power management and per-formance optimization for multimedia service system have important academic and practical significance.This work focuses on the power control and performance optimization for large s-cale multimedia service systems, including multimedia service clusters and multimedia gateways. Using the proportional computing idea, this dissertation gives the mathemat-ical models to describe system dynamic evolution with the consideration of practical engineering background, and derives the optimal control policy to achieve the adap-tion between system resource and actual workload. The analytical approach and model possess generality and could be applied to other internet services. The main work is as follows:For the power control problem of video on demand (VoD) server clusters, a hier-archical optimization mechanism for system structure is adopted to achieve power con-trol and performance optimization, i.e. increasing the profit rate per system resource in cluster and applying client adaptation technology to compensate performance degrada-tion due to cluster adjustment in client. In cluster side, we use a bottom-up approach to analyze the system evolution and propose a two-level Markov switching state-space control processes model to describe the hierarchical dynamic structure of the cluster. This model has two levels, the lower level is named as QoS level corresponding to in-crease the QoS and the upper level is named as power level corresponding to decrease the system power consumption. The optimization object is to maximize the ratio of performance to power consumption. Then we propose a hierarchical cooperation algo-rithm to search the optimal policy. In client side, we employ adaptive palyout control technology to combat cache outage caused by network jitter or limited cluster resource. Firstly, an statistical model is established to estimate the underflow time of the cache. Based on the estimation, an adaptive playout algorithm with dual thresholds to adjust the frame rate is presented. This approach adjusts the threshold dynamically accord-ing to the status of both network and cache, and thus regulates the frame rate timely. Moreover, the algorithm possesses scene aware characteristic, since it adjusts the play-out rate according to the motion intensity of the playout scenes to reduce the underflow probability of cache and provide better quality of experience (QoE).For the power control problem of popular time shifted TV cluster system, we set up a time shifted TV cluster architecture with power management function. The core function of this architecture is to monitor system status of the utilization of system re-source, etc., and execute the control policy to adjust the number of system available resource dynamically by power controller. Based on this system structure, we con-sider system optimization with stationary determined policy and stationary stochastic policy respectively. When the stationary determined policy is adopted, the system re-configuration problem is modeled as a Markov switching state space control process to dynamically powering on or off server optimally. Based on the theory and online estimation approach of performance potential, an online adaptive policy iteration algo-rithm is presented. The algorithm solves the optimal policy based on a sample path and the solving procedure does not depend on any prior knowledge of system parameters. When the stationary stochastic policy is adopted, we firstly group the system resource to form different configuration schemes, and quantize the connection number and channel number to different levels. In this way, the dimension of state space is reduced largely. Then the same control model is adopted to describe the system evolution and the prob-lem of power conservation is posed as a constrained stochastic optimization problem with the goal of minimizing the average power consumption subject to the constraint on the average blocking ratio. Applying lagrange approach and online estimation of the performance gradient, a policy iteration algorithm is proposed to search the optimal policy online. The convergence of the algorithm is theoretically analyzed.For the power control problem of multimedia gateway, we employ adaptive link rate technology to adjust the bandwidth of gateway so as to reduce the power con-sumption. Firstly, we analyzed the multimedia network traffic, including concurrent connection number, sojourn time and the characteristics of multimedia data. Then the Markov Modulated Poisson Process (MMPP) model is introduced to model the mul-timedia traffic and the power control of multimedia gateway is modeled as a semi-Markov decision process. Then we derive the performance sensitivity formulas based on embedded Markov chain. And the performance gradient is estimated through re-generative cycles on the sample path. The online optimization algorithm is presented in the final. Unlike the conventional optimization algorithm to search optimal policy, such as linear programming or infinitesimal generator based performance optimization, the sample path needed in the algorithm can be generated by semi-Markov kennel ma-trix directly, hence this optimization mode is equivalent to search the optimal based on semi-Markov kernel matrix, avoiding the drawback of infinitesimal generator based optimization algorithm.