Converging Credit-based and Reactive Datacenter Transport using ECN and RTT

Abstract

There are various transport protocols developed to mitigate the congestion caused by burst traffic and thousands of concurrent flows in data center networks (DCNs). These protocols can be classified into two kinds: proactive protocols and reactive protocols. Reactive protocols are common to current DCNs and they deal with the network congestion after congestion arises. However, as the common link speed gradually increases from 10 Gbps to 100 Gbps, reactive protocols, such as DCQCN, encounter a significant problem of slowly responding to the network congestion. In contrast, proactive protocols, especially credit-based congestion control, literally prevent the network from congestion before it occurs, which can also provide zero data loss, fast convergence, and low buffer occupancy.Due to the above advantages, the gradual deployment of credit-based protocols into DCNs is a visible task in the future. However, in the real deployment scenarios, it is hard to guarantee that one protocol can be deployed in every host at one time. Thus, when deploying credit-based protocols into DCNs incrementally, the network will converted to a multi-protocol state and face several fundamental challenges: (i) unfairness, (ii) high buffer occupancy and (iii) heavy tail delay. In this paper, we propose a multi-signal credit-based protocol, called MCP, aiming for converging reactive and credit-based protocols in DCNs. To the best of our knowledge, MCP is the first to leverage both explicit congestion notification and round trip time information to detect the network congestion in data queue and redefine the feedback control of the credit-based protocols. We also compared the performance of feedback control algorithms based on different congestion signals. Our evaluation shows that MCP effectively addresses the unfair link allocation problem and converges with reactive protocols rapidly. In addition to that, MCP achieves high utilization, low buffer occupancy and fast convergence at the same time.