Network Burst Peak TCP Congestion Control

Abstract

In this paper, the common greedy congestion control protocols are studied, and their
shortcomings and improvement methods are analyzed. In data centers, network applies online
data-intensive (OLDI) applications which have the characteristics of low latency, high throughput,
and high burst, requiring query response latency control of less than 300 milliseconds in the data
center. The DCTCP congestion control protocol based on explicit congestion notification (ECN)
achieves better performance in data center network applications, but requires all Switches to support
ECN and are not compatible with all Switches. The greedy congestion control protocol is improved
to meet the needs of the data center. In this paper, there are two contributions. First, the experimental
research on greedy congestion control and DCTCP protocol is carried out; Secondly, a unified
management congestion control protocol based on the sending window is proposed. It can have better
versatility and performance similar to DCTCP, and the packet loss rate is 56 % lower than NewReno.