Research On Demand Prediction And Allocation Algorithms For Resource Optimization And Management In Cloud Computing

In the era of big data and artificial intelligence,the cloud datacenter,as a service carrier,needs to provide massive physical and virtual resources to support various applications.Optimization and management of massive resources to improve the service and support capabilities of the cloud platform is one of the key research issues of cloud computing.However,with the diversification and complexity of applications,cloud resource demands present stronger diversity,emergency,and batch,which imposes great challenges on the optimization and management of cloud resources When a large number of resource demands emergently appear in the cloud platform,a more accurate prediction method of resource demands is needed to achieve efficient resource allocation.However,the existing prediction methods can not provide high accuracy of the complex characteristics of the current resource demands,which often causes untimely resource allocation and low utilization.Simultaneously,the traditional resource allocation methods can hardly support the emergent mode,which cannot guarantee the fast and optimal resource allocation for massive emergent resource demands.The existing fixed resource quota mechanism cannot continuously meet the users' dynamic resource demands and may make many users occupy high quotas with low utilization for a long time,which makes it difficult to control and optimize resource allocation,and limits the capacity of the cloud platform to provide resource services.Moreover,the existing evaluation methods of resource allocation are difficult to be implemented in practice due to their numerous indexes and complex testing process,which leads to more difficult resource allocation optimization.This dissertation studies the key algorithms,such as resource demand prediction,resource allocation,resource quota allocation,and resource allocation evaluation,to deal with the above problems.The proposed algorithms focus on the optimization and management of cloud resources and are based on the theories of signal frequency feature analysis and multi-objective optimization.The main contributions and innovations are as follows:(1)The short-term prediction algorithms of cloud resource demands and load based on the ensemble empirical mode decomposition(EEMD)are designed and an adaptive prediction algorithm based on dynamic threshold is proposed.First,a short-term prediction algorithm that combines EEMD method and autoregressive integrated moving average(ARIMA)model,EEMD-ARIMA,is designed to meet the emergent and unstable cloud resource demands,which improves the prediction accuracy.Then,a short-term server load prediction algorithm,EEMD-RT-ARIMA,is constructed to solve the problem of long prediction time of the EEMD-ARIMA algorithm through the identification and reconstruction of efficient feature components.Finally,based on the characteristics of the two algorithms,we design the abnormal data processing method,adaptive prediction strategy,and error adjustment method etc.,and propose an adaptive short-term prediction algorithm of cloud resource demands and load,which can adaptively select a high-accuracy method to predict the data with different characteristics.(2)A resource allocation algorithm is proposed for emergent resource demands.First,a proactive prediction strategy and a priority model of resource allocation are respectively designed to deal with the emergency of resource demands,and a resource performance matching model and a resource ratio matching model are built to optimize the allocation and utilization of various resources.Then,a multi-objective optimization model of cloud resource allocation is established.Finally,a multi-objective optimized algorithm of resource allocation is built by improving the non-dominated sorting genetic algorithm with elite strategy(NSGAII),which accelerates the solving speed and enhances the quality and the distribution of solution sets by parallelly computing the fitness functions,deleting the repeated individuals,and selecting the adjacent excellent individuals.Consequently,the timeliness and optimization of resource allocation are ensured,while the traditional resource allocation method cannot support the emergent mode and respond quickly to the emergent resource demands.(3)A credit factor based resource quota allocation algorithm is proposed to optimize the resource control and fair use,where the users' credit factor and weight models are established based on the historical utilization of their resource quotas.The users' resource quotas are dynamically allocated and periodically adjusted according to the credit factors.The punitive allocation of resource quotas is carried out for users occupying high resource quotas with low utilization for a long time and the penalty effect is accumulated.This algorithm controls the number of users' resources,enhances the fairness of users' resource usage,optimizes the resource utilization,and improves the resource service capacity of the cloud platform.(4)An evaluation method based on event sequences for cloud resource allocation is proposed.Aiming at the problems of many evaluation indexs and complicated testing process of resource elastic allocation,a minimum evaluation index set is designed and a metamorphic testing method based on event sequences is established.This method not only defines the metamorphic properties and metamorphic relations between event sequences,but also constructs a quantitative measurement method of metamorphic relation priorities that can estimate the fault detection capabilities of metamorphic relations in the design phase.Experiment results show that it not only realizes the simple and effective testing of resource elastic allocation,but also has strong suitability to the system testing in other fields with large test volume and complicated process.