Resource Scheduling In Cloud And Edge Environments

The core problem in cloud edge environment is how to manage resources and accept tasks,that is,resource scheduling problem.In the distributed architecture of edge sites,a large number of users move around and initiate business requests irregularly.There are a lot of service arrivals,service fluctuations,security uncertainties,uneven resource allocation and deployment among edge sites and finiteness of edge resources in the network junction layer.The uncertainty of service arrival leads to the problem of dynamic multi-workflow offloading and scheduling in MECC(Mobile edge cloud computing)environment,and the key performance evaluation indexes of edge are required to be optimized.However,the uncertainty of service load fluctuation,unbalanced resource allocation and provisioning among edge sites,and limited number of edge resources lead to the requirement of resource provisioning and periodic adjustment of resource pools.Computation offloading avoids the information leakage caused by the remote cloud of some data uploaded,but it still has the risk of being attacked by the side channel,which leads to the privacy protection requirement in computation offloading.These problems come from a wide range of practical scientific computing and other application scenarios,which constitute the scientific research topic of resource scheduling in MECC environment.On the basis of integrating the computing requirements of agile connectivity,content optimization,security and privacy protection,this paper considers the offloading and scheduling problem of workflow under the edge resource deployment based on workload prediction.The privacy protection of user data is enhanced by the priority setting and isolation method based on security entropy constraint,and the overall solution is guaranteed to be low risk.The optimization objective considers the makespan and cost,and the constraints include: soft deadline constraint,task partial order relation constraint,task affinity constraint,entropy constraint and overall risk constraint,with emphasis on dynamic multiple workflows.The innovative work of this paper is mainly reflected in:(1)Offloading and scheduling dynamic multi-workflows: Considering the coupling requirements of offloading and scheduling,a heuristic algorithm is proposed to offload workflows one by one,and the priority sequence of internal tasks of each workflow is constructed.When multiple workflow tasks are scheduled on a specific cloud edge platform,an improved fair algorithm is proposed to define the priority order of workflow according to the urgency of the tasks.Considering the uncertainty of edge task arrival,the method of offloading and sending to target platform is proposed.Incoming workflows are received and allowed to join in the schedule to ensure that the processing of the workflow meets the deadline.Considering the key performance evaluation indexes,the duration and cost are adopted as the optimization objectives.This coupling of computational offloading and resource scheduling improves the optimization at a finer granularity.Through a lot of experiments,it adapts and eliminates the uncertainty of workflow dynamic arrival,and save at least 23.8% on makespan,8.14% on cost,and 28.7% on missed deadlines.(2)The affinity tasks are computed and offloaded based on dynamically adjusting the resource pool of distributed units: The number of service arrivals in a MECC environment is uncertain.When service is scarce,distributed unit needs to be shut down to save energy.When a service overflow occurs,some distributed computing resources in the resource pool should be enabled in advance to respond in a timely manner.Resources on the edge are limited,and resource provisioning and allocation are unbalanced.The periodic provisioning/deployment on the edge needs to be adjusted dynamically and periodically.Resource hardware takes time to boot,hence periodic resource prediction can be done in a variety of ways,but accuracy is important,as well as when the initial data set is small is introduced to determine future workloads and resource requirements.Prediction can be done in a variety of ways,but accuracy is important,as well as when the initial data set is small.We use long short-term memory and dynamic Bayesian network(LSTM-DBN)model to predict periodic workload data and adjust computing resources accordingly.Specific scenarios generate special types of affinity tasks.Since the task characteristics can be sensed by the controller,content-based offloading is introduced.The specific offloading strategy for affinity tasks can reduce the transmission cost of offloading,improve the degree of parallelism,and reduce the degree of resource competition.This solution covers resource provisioning,computing offloading,and resource allocation,and is a feasible overall solution to the critical problem of resource scheduling.It effectively deals with the imbalance of resource allocation and provisioning,the uncertainty of traffic fluctuations,and the limitation of edge resources,and fills the gap of previous research.LSTM-DBN can predict with less data and improve the prediction performance using probabilistic inference and optimal estimation.The deep reinforcement learning offloading algorithm with resource adjustment(relaxation of resource constraints)is compared with the existing methods,and the effectiveness and efficiency of the proposed algorithm are verified.(3)Privacy protection and edge offloading based on security entropy constraint: The previous edge offloading schemes follow certain policy rules,which are easy to reveal the individual characteristics of users and face the risk of side channel attack.We use entropy-constrained privacy protection scheme to determine the security level and eliminate the offloading of individual features to avoid this risk.The system architecture where data surface and control surface are together is prone to attacks or security vulnerabilities.We introduce the global assignment system architecture of centralized controller to separate the data surface from the control surface,which reduces the burden of edge server and enhances the security of offloading policy.The edge produces a huge number of tasks,which is unreasonable if users prioritize themselves(all users think they have the highest priority).Manually determining the security level of big data is also an arduous process,so the author upload the user task characteristics to the controller,and the controller determines the security priority according to the exponential probability of the Poisson distribution on the task arriving.Terminals are told by the central controller which tasks are offloaded to the cloud and need encryption,and which tasks are offloaded to the edge and do not need encryption.It is necessary to strike a balance between privacy protection and the offloading operation to optimize the cost/duration of the task.We propose an intelligent optimization method based on MOEA/D for workflow offloading.The central controller tries to find a balance between the optimization goal and the privacy protection by offloading a large number of workflows,assigning priority according to Poisson distribution,and dealing with the entropy constraint of security level.Our research combines the optimization of computational offloading and privacy protection,which is proved to effectively deal with the security uncertainty through a large number of experiments verification,and fills the gap of previous research.