Controlling TCP Incast Congestion in Data Centre Networks

The rapid increase in data centre communication in recent years has led to a wave of interest in the problems that affect communications in the data centre environment. Many data centre applications rely on TCP/IP protocols for reliable data transport, which while successful for many Internet applications, does not integrate with and perform seamlessly in the data centre environment. One of these problems arises in data centre setups where many servers communicate simultaneously, and effectively in parallel, with one client through a single switch. In this scenario, the servers could experience a large drop in throughput due to severe packet loss, a phenomenon known as Incast congestion. In this paper, we investigate the Incast problem and introduce a variant of TCP, namely M21TCP-A, which is developed with the idea that the switch can be used to inform each parallel sender (server) of how many other senders are communicating simultaneously. Each sender can then limit its congestion window based on this feedback from the switch. M21TCP-A is evaluated against RED, ECN and DCTCP, which are the most successful congestion control proposals for mitigating Incast. Performance results show that M21TCP-A mitigates Incast thoroughly and outperforms previous proposals.