Research On Optimized Execution Of Cloud Workflow Services

With the development of cloud computing,increasing number of applications are de-ployed into cloud environment,which are provided to users in the form of cloud services.Workflows,as a collaborative tool,can also be a public service in the cloud environment.Hence,this thesis investigates the optimized execution of WFaaS(WorkFlow as a ser-vice).Specifically,this thesis focuses on three representative types of cloud applications:the service-oriented applications,the data analysis-oriented applications,and the machine learning applications.The major contributions are summarized as follows:First,a three-layers service architecture,named "Engine manager-Engines-Actors",is proposed for cloud workflow execution service.Under this architecture,the workflow engines are responsible for the control-flow and data-flow between tasks,and the actors are responsible for running the tasks.The decoupling of engines and actors makes the execution of the cloud workflow services more flexible and universal than existing cloud workflow systems.Two different granularity deployment modes of the cloud workflow engines are also devised,which are workflow-level deployment mode and task-level deployment mode.Meanwhile,considering the dynamic nature of the cloud requests,the author further proposes a hybrid rule based engine elasticity policy by combining the workload prediction algorithm,queuing theory and elasticity rules.Experimental results suggest that with the proposed elasticity approach,cloud workflow engines can capture the trend of dynamic workload and adjust themselves adaptively.Second,the issue of cloud workflow execution optimization problem for service coordination is addressed in this thesis.With the quick development and spreads of cloud computing,the number of various types cloud services in the Internet is increasing rapidly.Workflow can be applied to coordinate multiple cloud services to fulfill complex tasks.Considering the global distribution of Web services,the author proposes a general-ized methodology for service-oriented workflow scheduling in geographically distributed clouds.Experimental results show that the proposed method has shorter makespans and SLR(Scheduling Length Ratio)while dealing with large-scale Web service tasks.Third,the execution optimization of cloud workflow service for streaming data anal-ysis tasks is studied in this thesis.With the advent of big data era,the data analytical applications have sprung up which can be modeled using workflow.Such analytical workflow applications are subject to continuously arriving requests and have a rigid re-quirement on throughput or response time.When running streaming analytical workflows on a cloud platform,one of the critical questions which arise is how to provision resources so that the monetary cost can be reduced while still guaranteeing system throughput.The author proposes two cost effective resource provisioning algorithms which can guarantee system throughput and response time,respectively.Last,the author devises a machine learning workflow to coordinate the parallel execution of machine learning algorithms using cloud workflow service.By analyzing the parallel execution process of machine learning tasks,the author proposes the concept of machine learning workflow,which is characterized by a large number of independent parallel branches.At the same time,each branch has a single cloud server running independently.When using cloud workflow services to perform machine learning work-flows,the objective is to run the machine learning tasks with a minimum completion time while monetary cost as little as possible.First,the running time of the machine learning algorithms is estimated.Then,the author solves the problem for two situations:homogeneous servers and heterogeneous servers.Theoretical analysis and extensive experimental results show the proposed methods well capture the nature of various applications,and improve the performance notably.