Research On Strategies Of Service Composition And Activities Scheduling For Cloud Workflow

With the deep research of cloud computing and the development of its infrastructure, more andmore powerful, resource-intensive scientific workflows, business processes and collaborativeapplication processes have been developed in the cloud. Cloud computing application processes arealso increasingly more complex, and constrained by the time, cost and resource. Cloud workflow canfacilitate building, implementing, managing and monitoring cloud computing applications flexibility,and can make cloud computing applications run automatically and efficiently as well. The significantdifference between cloud computing and traditional computing environment is on-demand access tocomputing services at any time according to "pay-as-you-go" model. At the same time, owing to thedynamics, distribution, heterogeneity and autonomy of cloud computing, the traditional workflowmethods and technologies cannot deal with the emerging issues of cloud workflow effectively.Cloud computing is a service-oriented paradigm. In this paradigm, hardware, platform orsoftware is provided as a service. The model of "pay-as-you-use" has boosted the rapid growth ofcloud service in types and quantity; these services have laid a solid material foundation for cloudworkflow. There are fruitful research efforts on web service selection, composition and adaptation ingrid environment. But it is still a hot topic in service-oriented computing and cloud workflow thathow to employ cloud service for improving enactment of cloud workflow. Based on the features ofworkflow and considering user preferences, an adaptation-aware web services composition strategyhas been proposed.Scheduling algorithm or strategy is the core component of the cloud workflow engine. Differentfrom grid environment, cloud computing environment is market-oriented, so the objective functionof cloud workflow scheduling algorithm should take not only makespan but also cost into account.Scheduling problem is to discover collections of potential virtual machines which are suitable forrunning the tasks, then select the appropriate subset of virtual machines from these collectionswhich can fulfill the tasks with a minimum objective value, and then map tasks to virtual machineswith satisfied the constraints (usually considering makespan and cost). The scheduling process is anNP-complete problem. Because of the temporal dependencies and causal dependencies, constraintsto consider when schedule the cloud workflow are much more complex than the common cloud computing application.The main research work and originalities are as follows:1) Based on the QoS characteristics of cloud workflow service resources and considering thecriticality of workflow activities, an adaptation-aware cloud workflow services composition strategyhas been proposed which take user preferences into account. The strategy is composed of two stages.During cloud services selection phase, four dimensions of service QoS have been taken into account,they are makespan, cost, reliability and reputation. A multiobjective Ant Colony Optimization hasbeen employed to find Pareto solution set and optimized solutions which are satisfied the user'snonfunctional requirements. During cloud services composition phase, some cloud services havebeen reserved according to user's QoS requirements. When a cloud service fails to finish the task,corresponding reserved cloud service can be woken up to complete the task and the violation can bequickly resolved. Based on user's preferences, the adaptation-aware strategy can decide which taskshould make service reservation and what kind of services should be reserved from the servicereputation, service reliability, criticality perspectives.2) Considering the non-linear characteristics of cloud workflow activities duration, cloudworkflow activities duration prediction model is proposed based on chaotic time series. This modelleverages phase space reconstruction theory and RBF neural network for nonlinear time seriesforecasting. Cloud workflow activities duration is often affected by the activities of systemperformance, network conditions and other factors. Linear time series methods cannot predict thisnonlinear problem effectively. In this paper, the coordinates delay method has been used for chaotictime series phase space reconstruction. The embedding dimension is determined by theGrassberger and Procaccia proposed GP algorithm. The time delay value is determined by theautocorrelation function method. The orthogonal least squares method is employed as the RBFnetwork learning method.3) The cloud workflow scheduling strategy based on intelligent algorithms andadaptation-aware of cloud services composition strategy has been developed to schedule thetwo-level cloud workflow tasks. The tasks of cloud workflow should be run on virtual machineswhich are in the cloud service provider's data center. So, the scheduling process consists of servicelevel scheduling and task level scheduling. During service level scheduling, we suggest to divide thecloud application into subtasks or set of subtasks. These divided tasks are much simple than theoriginal tasks, most of them can be run in parallel and they are packaged as task units. The servicelevel scheduling process is to select suitable cloud service for these task units. We call this methodpackage-based scheduling. The number of the task level scheduling tasks is not too much. Becauseof the virtual machine is shared and many other tasks are running on the virtual machine at the time of allocating tasks, so the number of tasks involved would be considerable. We investigated andredesigned three representative intelligent algorithms (genetic algorithms, ant colony algorithm andparticle swarm optimization) for scheduling optimization. Considering the nature of these threealgorithms, we proposed a workflow activity scheduling policy based on intelligent algorithms.4) The cloud workflow tasks rescheduling strategy is proposed based on temporal consistencymodel and cloud workflow activities duration prediction model to address the recoverable temporalviolations occurred during the run-time phrase. In the run-time phase of the cloud workflow,temporal violation may occur to the assigned tasks on the virtual machines. If the violations are notwell addressed, it may leads to cloud workflow task cannot be fulfilled according to the user qualityof service. This will not only affect the duration of the cloud workflow, but also will affect thereputation of service providers. Most of the temporal violation can be solved through taskrescheduling and the cloud workflow can be completed within the quality of service constraints.Temporal consistency model is used to run statistical analysis of the quality of the workflow, andbased on the basis of recoverable temporal violation the model is defined. Rescheduling strategy isobtained a workflow scheduling which is optimal to the violated sub workflow in QoS constraintsfrom the candidate solutions generated during the scheduling stage, and reallocate resources to theviolated sub workflow. Rescheduling is to search solutions from the candidates, so the time neededis very short and the impact on the system can be ignored. In this sense, the strategy can be used asthe online rescheduling strategy of cloud workflow.This paper has explored and investigated the issues of workflow management in the cloud.Taking workflow management as application background, we have investigated cloud workflowservices resources composition model, activities duration prediction model and the cloud workflowscheduling and rescheduling strategy. This study not only has important practical values in the cloudworkflow application management, but also has great theoretical significance in the cloud computingresearch.