A congestion control method of SDN data center based on reinforcement learning. (14th September 2018)

• Record Type: Journal Article
• Authors: Jin, Rong; Li, Jiaojiao; Tuo, Xin; Wang, Weiming; Li, Xiaolin
• Is Part Of: International journal of communication systems. Volume 31:Number 17(2018)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2018-09-14
• Subjects: congestion control -- data center network -- reinforcement learning -- software‐defined network; Telecommunication systems -- Periodicals; 621.382
• DOI: 10.1002/dac.3802
• Languages: English
• ISSNs: 1074-5351

Abstract

Summary: With the development of cloud computing and big data, the internal communication business in data center has increased dramatically, and then the traffic in data center has also significantly increased. The bandwidth of data center is difficult to meet the bandwidth requirements of those intensive applications, and data center is facing a risk of network congestion. Under the background of the development of network intelligence, software‐defined network (SDN) should demonstrate its intelligence as a future network architecture. In this paper, we introduce reinforcement learning into the SDN data center to implement congestion control based on flow. We improve the Q ‐learning and Sarsa algorithms and propose two methods of congestion control based on the algorithms. Test results show that these two congestion control methods can control congestion effectively. And Sarsa method has a better performance of link utilization. The average link utilization of the Sarsa method is 2.4% higher than the Q ‐learning method and is 4.48% higher than the on‐demand method. Abstract : We introduce reinforcement learning into the SDN data center to implement congestion control based on flow, improve the Q ‐learning and Sarsa algorithms, and propose two methods of congestion control based on the algorithms. Test results show that these two congestion control methods can control congestion effectively, and Sarsa method has a better performance of link utilization. The average link utilization of the Sarsa method is 2.4% higher than the Q ‐learning method and is 4.48% higher than the on‐demand method.