SFTB: Scheduling Flows to Quickly Fit Traffic Burst in Data Center Networks: A Traffic Balancing Framework Based on Congestion Feedback

Abstract

A modern data center may host tens of thousands of machines, mixing with hundreds of thousands flows. Large number of flows concurrently traverse in a data center network may frequently cause traffic burst and unbalance, which may further induce congestion, packet losing, and therefore low efficiency. Generally speaking, there are three large families of scheduling algorithms to address this problem: (i) stochastic load balancing, (ii) optimizing traffic distribution by VM migration, (iii) scheduling flows to spread in different paths. But all of them suffer from some shortages. Stochastic load balancing is load-agnostic, it may encounter local congestion, and VM migration is time-consuming, making it unable to fit burst quickly. Meanwhile, because the number of concurrent flows is large, the third category may also fall into inefficiency. An alternative approach is only scheduling those large flows, but usually we don't know how large it is until a flow has finished. We propose SFTB, a flow scheduling framework to quickly fit the traffic burst in data center networks. In fact, SFTB belongs to the third family, but by intelligently leveraging the Explicit Congestion Notification (ECN), SFTB can quickly respond to the burst and congestion. Especially, STFB performs in a fully distributed manner and require no traffic matrix information, making it suitable for any traffic patterns. We evaluate SFTB via large scale simulation. The results show that our method outperform single path TCP and ECMP from more than 10% to more than 70% in average throughput.