Research On Energy Efficient Load Balancing Algorithm For Data Centers

With the rapid growth of cloud services,the size and number of data centers need to continue to expand.However,the consequent huge energy consumption discourages the widespread deployment of data centers and significantly increases the costs for cloud service operators.Therefore,energy-efficient technologies for data centers are particularly important.Load balancing technologies based on workload allocation can achieve the goal of energy efficient by scheduling the utilization of computing resources of data center servers.However,traditional workload allocation algorithms,which focused on using accurate models to build optimization problems with complex constraints,lacked the adaptability of the complex and highly dynamic server environment,and lacked the mechanism of autonomic learning and updating policies.Deep Reinforcement Learning(DRL)is a method that could adapt to the environment and update policies through autonomic learning.However,the existing energy-efficient load balancing algorithm based on DRL Deep QLearning Network(DQN)algorithm does not consider the influence of normal workloads on server dynamics when allocating computation-intensive workloads.At the same time,it is vulnerable to the fluctuation of the policy caused by the change of the server state,so the learning effect needs to be improved.To solve the above problems,this paper proposes an energy efficient load balancing algorithm based on DRL actor-critic(AC)algorithm for high dynamic server scenarios.First,this paper constructs a highly dynamic system model in which server utilization is affected by computation-intensive and normal types of workloads,and establishes an optimization problem with the goal of minimizing long-term energy consumption under the constraints of quality of service.And then,the deep neural network is used to generate the actor network in the actor-critic algorithm,and the server state is constructed by the server's utilization,the number of idle physical cores,the energy consumption,and the state information,which is input into the actor network.The policy function in the actor network then outputs actions composed in terms of computation-intensive workload allocation decisions.At the same time,this thesis utilizes the deep neural network generate the critic network,and inputs the state information and the actions output by the actor network into the critic network.The critic uses the evaluation function to evaluate the actor's actions and directs the actor to adjust the policy function in the direction of maximizing the long-term reward,which consists of a reduction in the amount of energy consumed by the server.The stability of the actors' network policy function is guaranteed by the way that the critic evaluates the actors' output actions,and the policy fluctuation caused by environmental changes is avoided.In this paper,Alibaba 2018 Cluster Trace data set is used to build the data center operating environment,and the data of real scenes are used to simulate the energy efficient load balancing algorithm based on the actor-critic algorithm.The results show that the proposed algorithm in this paper reduces more data center energy consumption than the round-robin algorithm,the best fit algorithm and the DQN algorithm.