Flow distribution-aware load balancing for the datacenter

Leveraging centralized and distributed load balancing algorithm to jointly schedule long and short flows attracts much attention since by classifying flows, they could reduce the completion times of flows without losing the scalability. However, existing centralized-distributed joint algorithms generally classify flows according to a static threshold and are unable to adapt to traffic dynamics. More importantly, through over investigation, the overheads of the controller caused by classifying flows and handling flows can be further reduced. In this paper, we present a Flow Distribution-Aware Load Balancing (FDALB) mechanism to reduce flow completion times and achieve high scalability. In FDALB, flows are split into short flows and long flows according to a threshold. The traffic of short flows and long flows are balanced by distributed and centralized algorithms respectively. To adapt to traffic dynamics, we propose a simple yet effective scheme adaptively adjusting the splitting threshold. To reduce the overheads of classifying flows, FDALB leverages end-hosts to tag long flows, which requires no changes in networking hardware. To further reduce the overheads of handling flows, we proposed a new centralized algorithm without requiring flow rates information. Using realistic datacenter workloads, we show that FDALB reduces the average FCT of flows by up to 47% over ECMP while achieves higher scalability than the state-of-art load balancing mechanism Mahout. (C) 2017 Elsevier B.V. All rights reserved.