Avoiding TCP Incast through Scheduling Data Requests

摘要

In high bandwidth, low latency data center networks, when multiple data senders simultaneously communicate with a single receiver, namely in many-to-one communication pattern, the burst data overload the receiver's switch buffers, which leads to TCP throughput collapse. That is so-called TCP In cast problem. It has become a new hot research topic in recent years. And, many proposals have been put forward from the aspects of multiple layers including link layer, transport layer, and application layer, etc. This paper gives an application-level solution to avoid TCP In cast by scheduling the clients' requests to servers. Our solution does not modify TCP protocol or the network switch and can thus be applied to any TCP variant, and works regardless of the type of data center network topology. The simulation results confirm the validity of our solution.