文摘
Workload distribution has high impact on several performance indices,such as delay and energy costs in computing systems. Typically,average delays at servers and data centers are seen as Service Level Agreement SLA) at infrastructure levels. However,they are not explicitly computed and dependent on various factors such as statistics of arrival and service processes,flow controls,dispatching policies and others. In addition,energy costs that are incurred by these servers become one of important issues in large-scale computing systems. As a result,we investigate how to distribute incoming loads,so that the SLA requirement is achieved and power consumption is optimized. When the SLA requirement is given,dispatching and scheduling controls are utilized to bound the average delays and to optimize the energy costs. Guaranteeing the SLA can raise system performances and revenues of service providers,because it is a binding contract between customers and service providers. In general,penalty costs will be highly incurred by the SLA violations. If the SLA requirements are achieved and optimized under some constraints,quality for computing services can be ensured at a certain degree. In brief,our contributions are listed as follows. First,stability conditions and feasibility regions of dispatching policies are shown by using stability and feasibility analysis. Relationships between classes of policies are also determined in term of delays. Second,static policies are investigated for bounding the average delays. Theoretical analysis of waiting time is provided,and a LP method is proposed for determining a feasible set of controls to bound delays. NLP optimization can be formulated to optimize the waiting time at queues when the static policy is used. Third,dynamic policies are considered for bounding the average delays. These policies can be categorized into ideal and implementable cases. Stability conditions and convergence properties of dynamic policies are demonstrated by using stability and convergence analysis. Performances of these policies are evaluated via simulation,and an adaptive dynamic policy is proposed to bound the average delays. Lastly,scheduling controls for incoming requests with complex structures are investigated. Analytical results related to the optimal execution time and the bound of systems capacity are provided. Even a simple scheduling can be shown to achieve the throughput close to this bound. In addition,a scheduling method for optimizing energy and congestion costs is discussed for tasks with complex structures.