A Ternary Detection Multi-level Hashed Asynchronous Traffic Scheduling Method Based on DDQN
In Data Center Networks(DCNs),short message services related to Web access and distributed computing dominate the traf-fic.To address the long tail effect of short messages in data centers,we propose a receiver-driven traffic scheduling algorithm based on reinforcement learning and dynamic priorities.Leveraging DDQN(Double Deep Q-Network)reinforcement learning,the algorithm es-tablishes a dynamic scheduling threshold,performs dynamic priority allocation,and effectively reduces the tail delay of low-priority long messages through differential polling in the sending queue.Additionally,it directly sends short packets with lengths below the scheduling threshold from the sending end.At the receiving end,priority queues are dynamically assigned to packets longer than the scheduling threshold based on triplet detection information,eliminating preemption delay and ensuring high link occupancy and low transmission delay.Experimental results demonstrate that the proposed algorithm can reduce the completion time slowdown ratio of short message streams with less than Unscheduled Bytes by 85％under high loads exceeding 95％of the link capacity.
traffic schedulingdata center networkDDQN reinforcement learningin-network priorityternary detection
万方数据
维普
