Data conversion control of virtual network devices in cloud computing: A deep reinforcement learning approach

Abstract

As the response time of virtual space task requests is affected by different data conversion processes, it becomes more serious to solve the load imbalance problem between the underlying nodes and the underlying data links. The traditional least square method needs to distinguish each node for the remapping of the original virtual space. After the trajectories in different spaces and times are mapped again, it is easier to cause new resource bottlenecks. Aiming at the necessary virtual space remapping cycle of cloud service nodes, this paper selects the nodes with the highest upper and lower limits of load rate according to the dynamic critical value of each node's load rate. It migrates part of the virtual network space on the overloaded service nodes to the non-important nodes with a lower load rate, which is to avoid the gradual formation of new resource bottlenecks after remapping. The data transformation model based on reinforcement learning can further improve the acceptance rate of virtual space requests and reduce the computational overhead caused by space remapping. The experiments compare the performance evaluation of different algorithm models on the simulation platform data conversion tasks, including complex instructions test, virtual network performance test, average remapping link overhead, and root mean square error of the underlying node load rate test. The test results indicate that the calculation accuracy of the proposed method is 7.25% higher than other comparison methods, and the computational overhead is reduced by 15.22%. The overall performance is improved by 10.74%. It is verified that the technology in this paper improves the data conversion control method of virtual network devices in the cloud computing environment, and has practical significance for virtual network remapping and data services in cloud computing.