Throughput-Optimal Load Balancing for Intra Datacenter Networks

TL;DR

This paper introduces a new in-network load-balancing algorithm for datacenter networks that is throughput-optimal, supports link-capacity sharing, and improves performance over traditional ECMP methods.

Contribution

A novel load-balancing algorithm that achieves throughput optimality, enables link-capacity sharing, and is practical for implementation with TCP traffic.

Findings

Outperforms ECMP in simulations

Supports link-capacity sharing

Provides fairness similar to weighted fair queueing

Abstract

Traffic load-balancing in datacenters alleviates hot spots and improves network utilization. In this paper, a stable in-network load-balancing algorithm is developed in the setting of software-defined networking. A control plane configures a data plane over successive intervals of time. While the MaxWeight algorithm can be applied in this setting and offers certain throughput optimality properties, its bang-bang control structure rewards single flows on each interval and prohibits link-capacity sharing. This paper develops a new algorithm that is throughput-optimal and allows link-capacity sharing, leading to low queue occupancy. The algorithm deliberately imitates weighted fair queueing, which provides fairness and graceful interaction with TCP traffic. Inspired by insights from the analysis, a heuristic improvement is also developed to operate with practical switches and TCP flows. Simulations from a network simulator shows that the algorithm outperforms the widely-used equal-cost multipath (ECMP) technique.