| Virtual machine scheduling and placement is a key technology for data center management.It restricts the total energy consumption and resource utilization of the data center system,and affects the system load balance and service and service level agreement violation index minimization,etc..However,in practical application,the user request task sequence is uncertain,the system resource consumption changing at real-time,the traditional scheduling management method is not well balanced constraint conditions of server resource utilization,energy consumption of the virtual machine,optimizing the goal of load balance,so it cannot meet the demand of virtual machine management of complex dynamic constraints.In view of the above problems,this paper conducts the following research:In order to solve the problem of incomplete scheduling target and insufficient constraint in traditional virtual machine scheduling model,this paper proposes a virtual machine scheduling and placement model with multiple constraints and multiple targets.The model aims to solve the virtual machine placement at a given single moment.The dimensions considered are server CPU resources,memory,bandwidth,energy consumption,virtual machine migration times and load balancing.In order to simulate the scenario that different task requests to access the data center at different time slots in the actual application,a dynamic request simulation model is also established in the thesis,it can simulate lots of tasks changing over time.With this model,we can simulate the virtual machine sequence in the data center,and each task selects a virtual machine with matching performance to hold.In order to solve the problem that the existing genetic algorithm find difficulties when it's facing dynamic virtual machine scheduling problems,this paper proposes a genetic algorithm solver based on time series(T-Gene).The fitness function was designed according to the resource consumption of the physical server and the number of its boot-up,and the mapping relationship between the virtual machine and the physical machine was encoded by chromosome.The time dimension is used into genetic algorithm,and virtual machine dynamic mapping and virtual machine migration are considered to solve the problem.In order to solve the multi-objective optimization in the dynamic scheduling process of virtual machines under multiple constraints problem,this paper designed a Reward function Q table storing rewards under different actions and states based on Q-learning algorithm,and designed the calculation method of Q-value function for different scheduling objectives.According to the objective function and constraint conditions in the scheduling and placement model of multi-constraint multi-target virtual machine,the global Reward value corresponding to each step of the evaluation(virtual machine physical machine placement operation)is designed.This algorithm updates and saves the optimal Q table through model iteration training,and can map out a better virtual machine placement strategy according to the Q table.Finally,this paper develops and implements the T-Gene model and Q-Learning model proposed in this thesis,the heterogeneous data center contains 4 physical machine types.and 4 types virtual machines.We perform simulation verification on different numbers of virtual machine scenarios and compared with Random algorithm and the First Fit Descending(FFD)algorithm.The results show that the startup number of physical machines based on T-Gene algorithm is 14.50% lower on average than that of Random algorithm,and 5.03% lower than that of FFD algorithm.The virtual machine placement algorithm of q-learning can reduce the startup number of physical machines by 5.46% than that of Random algorithm.Compared with Random algorithm,the energy consumption of physical machine and the migration times of virtual machine based on T-Gene algorithm reduce by 16.60% and 18.02% respectively.Compared with FFD algorithm,the energy consumption of physical machine and the migration times of virtual machine are reduced by 4.02% and 9.30% respectively.Q-learning based virtual machine placement algorithm can reduce the physical machine energy consumption and virtual machine migration times by 13.70% and 21.07% respectively compared with Random algorithm.Q-learning algorithm can reduce the physical machine startup energy consumption and virtual machine migration times by 0.684% and 12.70% respectively compared with FFD algorithm.The results show that the proposed T-Gene virtual machine placement algorithm and the virtual machine placement algorithm based on the Q-learning shows its superiority under different dimension objective function in virtual machine model. |
