Research On User Experience Oriented Resource Management Techniques In Cloud Systems

With the development of Internet and virtualization technologies,cloud computing develops rapidly and is becoming a mainstream computing model.Cloud computing provides multi-tenant elastic services for Internet services,enterprise applications,scientific research,and so on.However,during the process of applying cloud computing,we are still facing many key scientific problems and technical challenges.In particular,the challenge of simultaneously satisfying resource utilization and user experience has become increasingly prominent,making how to optimize the system resource management while guaranteeing user experience becomes a key problem to be solved urgently.This thesis studies the user experience oriented resource management techniques in cloud systems,focusing on two types of problems which are user experience oriented resource scheduling and resource usage effectiveness evaluation.The detailed research content and main contributions are summarized as follows.Firstly,this thesis presents an user experience oriented resource scaling approach that provisions resources for long-running services in a fine-grained manner while guaranteeing the tail latency Service Level Objective(SLO).Specifically,this approach firstly employs a convolution based time series analysis to mine the spatio-temporal patterns of long running services.On this basis,it proposes top-K based collaborative filtering and wavelet based clustering algorithms to allocate resources to long-running services in a fine-grained manner.Moreover,it designs an online reprovisioning mechanism to deal with the potential tail latency SLO violations.The experimental results show that the proposed approach can significantly improve the resource utilization for long-running services while meeting the tail latency SLO.Secondly,this thesis presents an user experience oriented resource time-multiplexing approach for long-running services and batch jobs that can guarantee the tail latency SLO of long-running services while providing predictable performance for batch jobs.Specifically,this approach firstly estimates the reclaim probabilities of transient resources(i.e.,temporary idle resources)of long-running services,using historical resource usage patterns.On this basis,it proposes a cardinality-constrained portfolio model to select optimal transient resources for each batch job.Moreover,it designs a partial checkpointing mechanism to minimize the re-computation overheads of batch jobs in the case of reclaiming.The experimental results show that the proposed approach can significantly reduce the performance loss of batch jobs running on transient resources and improve the system resource utilization,while meeting the tail latency SLO for long-running services.Lastly,this thesis presents an user experience oriented resource usage effectiveness evaluation approach.Specifically,this approach firstly proposes a new metric called Resource Productivity by combining tail latency and resource utilization.On this basis,it adopts a Markov-Modulated Poisson Process(MMPP)to properly characterize the behavior of workloads.Moreover,it leverages the Stochastic Network Calculus(SNC)and Utilization Law to derive the Resource Productivity mathematically.The experimental results show that the proposed approach can effectively evaluate the resource usage effectiveness of cloud systems.