Enabling flow-based routing control in data center networks using Probe and ECMP

摘要

Data center networks often use densely interconnected topologies to provide high bandwidth for internal data exchange. In such network, it is critical to employ effective load balancing schemes so that the bandwidth resources can be fully utilized. A simple and widely adopted scheme is equal-cost multi-path (ECMP) routing, which is generally supported by commodity switches and routers. However, research shows that ECMP cannot always ensure even traffic distribution among multiple paths. Consequently, ECMP cannot guarantee optimal resource utilization. We propose a scheme to complement ECMP with per-flow reroute. The basic idea is to perform ECMP-based load balancing by default. When a congestion occurs on a certain link, we dynamically reroute one or a few big flows to alternative paths to alleviate the congestion. The main contribution of our research is to design a scheme that enables per-flow reroute without introducing any modifications to IP switches and routers. All the flow-based operations and reroute functionalities are implemented in software that are installed on end hosts and centralized controllers. We call this scheme PROBE (Probe and RerOute based on ECMP). PROBE uses a traceroute-like approach to discover alternative paths and modifies packet headers to enable flow-based reroute. We show that PROBE is a low cost, low complexity and feasible scheme that can be easily implemented in existing data center networks that consist of commodity switches and routers.