Abstract: As the amount of data increases in data center networks, the traffic load of some links is excessive. The traditional ECMP mechanism is no longer applicable to the data center network because it does not consider the link status and traffic characteristics. Meanwhile ECMP mechanism may hash some large flows to the same path. It may cause large flows collision and bring the problem of link bottleneck when scheduling flows. Based on the Software Defined Network（SDN） architecture, this paper proposes a dynamic Load Balancing scheme based on Flow Classification（LBFC） for the Fat-Tree topology, and considers link state information and traffic characteristics. LBFC mechanism dynamically adjusts the flow classification threshold to classify flows into the large and small flows. LBFC selects the forwarding path for the large and small flows in different ways to meet the different transmission bandwidth requirements of the large and small flows. The simulation results show that the LBFC can dynamically determine the large flows and small flows according to the network link status and traffic characteristics and achieve load balancing. Compared with ECMP, GFF and DLB algorithms, the LBFC mechanism improves network throughput and link utilization. The transmission delay is also reduced.

Key words: Software Defined Network（SDN）, load balancing, data center, fat-tree topology

摘要： 随着数据中心内的数据流量不断增加，导致网络中部分链路负载过重。传统的ECMP机制由于没有考虑链路状态以及流量特征，因此不再适用数据中心网络。同时ECMP可能会将多条大流映射到同一条路径上，造成大流映射冲突，导致链路瓶颈问题。基于SDN（Software Defined Network）架构提出一种面向Fat-Tree拓扑的动态流量负载均衡机制（Load Balancing based on Flow Classification，LBFC），同时考虑了链路状态信息与流量特征进行负载均衡。LBFC机制动态调整流分类阈值来判定大流和小流，采用不同的方式为大流和小流选择转发路径，以满足大流和小流不同的传输性能需求。仿真结果表明LBFC机制能够根据网络链路状态以及流量特征动态地判定大流和小流并实现负载均衡，与ECMP、GFF和DLB算法相比，LBFC机制提高了网络吞吐量以及链路利用率，降低了传输时延。

关键词: 软件定义网络, 负载均衡, 数据中心, 胖树拓扑