Preemptive SDN Load Balancing With Machine Learning for Delay Sensitive Applications

被引:37
作者
Filali, Abderrahime [1 ]
Mlika, Zoubeir [2 ]
Cherkaoui, Soumaya [3 ]
Kobbane, Abdellatif [4 ]
机构
[1] Univ Sherbrooke, Elect & Comp Engn, Sherbrooke, PQ J1K 2R1, Canada
[2] Univ Sherbrooke, Fac Genie, Sherbrooke, PQ J1K 2R1, Canada
[3] Univ Sherbrooke, Fac Genie, Elect & Comp Engn, Sherbrooke, PQ J1K 2R1, Canada
[4] UM5R ENSIAS, BP 713, Rabat, Morocco
关键词
Control systems; Load management; Load modeling; Predictive models; Delays; Computer architecture; Prediction algorithms; Load balancing; machine learning; migration; multi-access edge computing; predictions; reinforcement learning; software defined networking; SOFTWARE-DEFINED NETWORKING; CONTROL PLANE SCALABILITY; SWITCH MIGRATION;
D O I
10.1109/TVT.2020.3038918
中图分类号
TM [电工技术]; TN [电子技术、通信技术];
学科分类号
0808 ; 0809 ;
摘要
SDN is a key-enabler to achieve scalability in 5G and Multi-access Edge Computing networks. To balance the load between distributed SDN controllers, the migration of the data plane components has been proposed. Different from most previous works which use reactive mechanisms, we propose to preemptively balance the load in the SDN control plane to support network flows that require low latency communications. First, we forecast the load of SDN controllers to prevent load imbalances and schedule data plane migrations in advance. Second, we optimize the migration operations to achieve better load balancing under delay constraints. Specifically, in the first step, we construct two prediction models based on Auto Regressive Integrated Moving Average (ARIMA) and Long Short-Term Memory (LSTM) approaches to forecast SDN controllers load. Then, we conduct a comparative study between these two models and calculate their accuracies and forecast errors. The results show that, in long-term predictions, the accuracy of LSTM model outperforms that of ARIMA by 55% in terms of prediction errors. In the second step, to select which data plane components to migrate and where the migration should happen under delay constraints, we formulate the problem as a non-linear binary program, prove its NP-completeness and propose a reinforcement learning algorithm to solve it. The simulations show that the proposed algorithm performs close to optimal and outperforms recent benchmark algorithms from the literature.
引用
收藏
页码:15947 / 15963
页数:17
相关论文
共 36 条
[31]   An Efficient Online Algorithm for Dynamic SDN Controller Assignment in Data Center Networks [J].
Wang, Tao ;
Liu, Fangming ;
Xu, Hong .
IEEE-ACM TRANSACTIONS ON NETWORKING, 2017, 25 (05) :2788-2801
[32]   ON THE EQUAL-SUBSET-SUM PROBLEM [J].
WOEGINGER, GJ ;
YU, ZL .
INFORMATION PROCESSING LETTERS, 1992, 42 (06) :299-302
[33]   Dynamic switch migration with noncooperative game towards control plane scalability in SDN [J].
Wu, Guowei ;
Wang, Jinlei ;
Obaidat, Mohammad S. ;
Yao, Lin ;
Hsiao, Kuei-Fang .
INTERNATIONAL JOURNAL OF COMMUNICATION SYSTEMS, 2019, 32 (07)
[34]   A survey on software defined networking with multiple controllers [J].
Zhang, Yuan ;
Cui, Lin ;
Wang, Wei ;
Zhang, Yuxiang .
JOURNAL OF NETWORK AND COMPUTER APPLICATIONS, 2018, 103 :101-118
[35]   Elastic Switch Migration for Control Plane Load Balancing in SDN [J].
Zhou, Yang ;
Zheng, Kangfeng ;
Ni, Wei ;
Liu, Ren Ping .
IEEE ACCESS, 2018, 6 :3909-3919
[36]  
Zhu L., 2019, Advances in Neural Information Processing Systems, P32