UFLB: A Unified Framework for Modeling and Analyzing Load Balancing Methods in DCNs

Abstract

Data centers usually employ scale-out network topologies to provide sufficient network bandwidth for applications. The traditional equal-cost multi-path (ECMP) routing method is proposed to tackle the serious load imbalance problem across all links. However, it does not achieve the desired performance and still incurs low network throughput. Consequently, researchers recently redesigned some load balancing mechanisms for data center networks (DCNs) from different design dimensions. However, it remains open to systematically measure and evaluate their performance in various settings. It is impractical for evaluators to implement or simulate involved load balancing mechanisms. In this paper, we propose a unified framework, UFLB, which can well model and emulate representative load balancing mechanisms for data center networks in a lightweight way. This framework has overcome three significant challenges: model traffic distribution in the symmetry as well as asymmetry data center networks, characterize mainstream load balancing methods, and systematically combine them with high accuracy. We evaluate the effectiveness of our model under not only general settings of data center networks but also some special settings, such as various link failures and asymmetric topologies. The results indicate that the deviation rate of UFLB is within 15% against the implementation of load balancing mechanisms, such as ECMP, CONGA, DRILL, HERMES, PRESTO, in NS2, while it can be several orders of magnitude faster.