Research On Adaptive Cloud Resource Scheduling Based On Reinforcement Learning

As a new type of computing mode,cloud computing has been widely applied in the market.It changes the service model of traditional network computing and provides dynamic service model for users through the Internet.According to this model,users can access configurable resources,such as networks,storage,applications and services,at anytime,anywhere.In order to provide high-quality services,cloud service providers need to manage resource scheduling problems among different users1 requests under the condition of rational utilization of resource pools,for the purpose of allocating resources to users on demand.Therefore,it is benefit for cloud service providers and users to study resource scheduling.With the continuous expansion of market demand,the uninterruptedly growth of the number of users and the diversity of user task requirements,as well as the randomness of task arrival time,coupled with the dispersion,heterogeneity and uncertainty of computing resources,how to allocate resources reasonably and flexibly to meet the different requirements of tasks has become a very challenging problem in resource scheduling.Therefore,we will focus on the adaptive resource scheduling to achieve lower task response time,higher resource utilization and lower energy consumption in heterogeneous cloud environments.Aiming at the problem of resource scheduling in complex cloud environnent,we focus on two aspects:cloud resource scheduling model and cloud resource scheduling strategy.The main work of this paper are;(1)we analyze the rules of users' arrival and server services in cloud environment,we model the resource scheduling system of cloud computing based on M/M/S queuing model to allocate task to the idle resource pool(physical machine),which avoids the phenomenon of local queue overlong in heterogeneous cloud environment,reduces the waiting time of tasks,and achieves the optimal allocation of resources to random tasks.(2)In order to further reduce the response time of tasks,improve resource utilization and achieve energy saving during scheduling process,task length,deadline and waiting time are taken into consideration to preprocess tasks dynamically.And according to the current execution status of physical machines and virtual machines in the system,such as CPU utilization,resource availability,load variation characteristics and energy consumption characteristics,etc.,the task is scheduled to the virtual machine through the scheduling strategy {state-action-feedback} of reinforcement learning to achieve adaptive resource scheduling.(3)We make a theoretical analysis and conduct extensive experiments on the scheduling model,and the scenario of random task arrival in heterogeneous cloud environment is constructed on CloudSim simulation platform.The proposed adaptive scheduling mechanism compared with other classical methods,and the experimental results demonstrate that our method is effective and accurate.