CAFT: Congestion-Aware Fault-Tolerant Load Balancing for Three-Tier Clos Data Centers

Abstract: Production data centers operate under various workload sizes ranging from latency-sensitive mice flows to long-lived elephant flows. However, the predominant load balancing scheme in data center networks, equal-cost multi-path (ECMP), is agnostic to path conditions and performs poorly in asymmetric topologies, resulting in low throughput and high latencies. In this paper, we propose CAFT, a distributed congestion-aware fault-tolerant load balancing protocol for 3-tier data center networks. It first collects, in real time, the complete congestion information of two subsets from the set of all possible paths between any two hosts. Then, the best path congestion information from each subset is carried across the switches, during the Transport Control Protocol (TCP) connection process, to make path selection decision. Having two candidate paths improve the robustness of CAFT to asymmetries caused by link failures. Large-scale ns-3 simulations show that CAFT outperforms Expeditus in mean flow completion time (FCT) and network throughput for both symmetric and asymmetric scenarios.