Premium
SDN‐based dynamic multipath forwarding for inter–data center networking
Author(s) -
Wang YaoChun,
Lin YingDar,
Chang GueyYun
Publication year - 2018
Publication title -
international journal of communication systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.344
H-Index - 49
eISSN - 1099-1131
pISSN - 1074-5351
DOI - 10.1002/dac.3843
Subject(s) - computer science , computer network , multipath propagation , scalability , latency (audio) , data center , software defined networking , network topology , load balancing (electrical power) , multipath routing , routing (electronic design automation) , routing protocol , telecommunications , operating system , channel (broadcasting) , geometry , mathematics , grid , link state routing protocol
Summary Equal‐cost multipath (ECMP)–based traffic engineering (TE) methods are commonly used in intra–data center (DC) networks to improve the transmission performance for east‐west traffic (ie, traffic from server to server within a DC). However, applying ECMP on inter‐DC wide area network (WAN) offers limited performance enhancement as a result of irregular network topology. Since TE can be intelligently and efficiently realized with software‐defined networking (SDN), SDN‐based multipath becomes a popular option. However, SDN suffers from scalability issue caused by limited ternary content‐addressable memory (TCAM) size. In this paper, we propose an SDN‐based TE method called dynamic flow‐entry‐saving multipath (DFSM) for inter‐DC traffic forwarding. DFSM adopts source‐destination–based multipath forwarding and latency‐aware traffic splitting to reduce the consumption of flow entries and achieve load balancing. The evaluation results indicate that DFSM saves 15% to 30% of system flow entries in practical topologies and reduces the standard deviation of path latencies from 10% to 7% than do label‐switched tunneling, and also reduces average latency by 10% to 48% by consuming 6% to 20% more flow entries than do ECMP in less‐interconnected topologies. Note that the performance gain may not always be proportional to flow entry investment, with the interconnectivity between nodes being an important factor. The evaluation also indicates that per‐flow provision consumes several times the flow entries consumed by DFSM but reduces latency by 10% at most. Besides, DFSM reduces the standard deviation of path latencies from 14% to 7% than do even traffic splitting.