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

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