Distancer: A host-based distributed adaptive load balancer for datacenter traffic

Abstract

Contemporary datacenter networks are typically organized with multi-rooted tree topologies. To fully utilize the multiple end-to-end paths, effective mechanisms are required to balance traffic across them. However, existing load balancers for datacenters either operate at a coarse granularity, or support little for network failures, or necessitate customized hardware. We propose Distancer, a host-based distributed adaptive load balancer for datacenter traffic, which requires no coordination and modification of switches. Based on a deep investigation of TCP feedback mechanism, we firstly design Congestion Detector (C-Detector), which exploits ACKs to effectively handle network hot-spots and path anomalies in real time; Then we develop Load-Balancer (L-Balancer) to select best paths for both data packets and ACKs. According to our extensive evaluations, Distancer can achieve up to 40% and 20% better average flow completion times (AFCTs) than ECMP and CONGA respectively. Under the presence of path failures, Distancer improves the AFCT up to 400% and 30% over ECMP and CONGA.