Scalable Data Center TCP for Cloud Applications

Many cloud data center applications such as MapReduce and social networking follow the incast traffic pattern, where multiple senders communicate with a single receiver simultaneously. Under this situation, TCP connections suffer throughput collapse due to severe packet losses that occur at the data center switches. Moreover, such packet losses result in relatively long flow completion times, hence poor quality of services. Data Center TCP (DCTCP) is a novel end-to-end solution that addresses this data center network congestion problem, but it is not scalable in terms of the number of concurrent TCP connections. In this paper, we analyze the main cause of such poor scalability and propose a more scalable congestion control algorithm for DCTCP (Scalable DCTCP) based on the analysis. Through ns-3 simulations, we show that the proposed scheme copes with a larger number of concurrent connections, significantly improving TCP performance.