M21TCP: Overcoming TCP Incast Congestion in Data Centres

Modern data centres host a myriad of cloud services and applications with stringent delay and throughput requirements. The vast majority of these applications are of type partition/aggregate, where many servers simultaneously communicate with one client to produce a result. Unfortunately, the standard TCP/IP protocol, originally proposed for the Internet one-to-one transport, falls short in this environment. This is due to the TCP throughput collapse in such environment, known as TCP Incast congestion problem. This paper revisits the Incast congestion problem and identifies its root cause: severe packet drops that result from the switch buffers overflow. We propose a method of controlling congestion, named 'Many-To-one' (M21TCP). The intuition is that a switch can inform all parallel senders of the maximum rate at which they can send packets that will not cause buffers overflow. M21TCP has been tested and evaluated against known previous proposals such as DCTCP, RED and ECN. Results show that M21TCP outperforms previous solutions and completely eliminates Incast for the maximum number of servers.