Scholarlink

Due to the ability to provide redundant end-to-end parallel paths thus offering extremely high bisection network bandwidth for data transmission, leaf-spine topology has been widely adopted to construct the underlying physical network of data center. This kind of potential advantage, however, has still not been completely transformed into huge performance enhancement of data transmission for data center. The fundamental reason is that the standard data center load balancing scheme, i.e., Equal Cost MultiPath, always traps in the adverse effect of hash polarization. Consequently, the congestion hotspots in data center network could not be effectively eliminated, while plenty of parallel paths are not fully utilized, ultimately resulting in the waste of bandwidth resources. This paper investigates the impacts of traffic polarization degrees of different types of data center flows on load balancing, and proposes a flow-level data center load balancing scheme, namely FL \(^2\) B, to improve the transmission performance of the data center flows. FL \(^2\) B can, to a certain extent, perceive the flow type. It can also, in combination with the local status of the switch, route flows to those paths that are conducive to improving their transmission performance, thereby achieving efficient utilization of network bandwidth resources. Experimental results of numerous NS3 tests show that FL \(^2\) B not only significantly reduces the overall polarization degree, but greatly improves the transmission performances of data center flows as well.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mitigating Hash Polarization with Flow-Level Load Balancing in Leaf-Spine Data Center Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Siyuan Fan,
  • Ao Zhang,
  • Hui Yu,
  • Tao Zhang,
  • Linfei Dong,
  • Xidao Luan,
  • Hui Yin

摘要

Due to the ability to provide redundant end-to-end parallel paths thus offering extremely high bisection network bandwidth for data transmission, leaf-spine topology has been widely adopted to construct the underlying physical network of data center. This kind of potential advantage, however, has still not been completely transformed into huge performance enhancement of data transmission for data center. The fundamental reason is that the standard data center load balancing scheme, i.e., Equal Cost MultiPath, always traps in the adverse effect of hash polarization. Consequently, the congestion hotspots in data center network could not be effectively eliminated, while plenty of parallel paths are not fully utilized, ultimately resulting in the waste of bandwidth resources. This paper investigates the impacts of traffic polarization degrees of different types of data center flows on load balancing, and proposes a flow-level data center load balancing scheme, namely FL \(^2\) B, to improve the transmission performance of the data center flows. FL \(^2\) B can, to a certain extent, perceive the flow type. It can also, in combination with the local status of the switch, route flows to those paths that are conducive to improving their transmission performance, thereby achieving efficient utilization of network bandwidth resources. Experimental results of numerous NS3 tests show that FL \(^2\) B not only significantly reduces the overall polarization degree, but greatly improves the transmission performances of data center flows as well.