TCP improvements for data center networks

Data centers are used to host computation-intensive distributed applications. Since all the applications are distributed in nature, their performance is significantly affected by the communication network. These applications generally create two kinds of flows: large data flows that require high throughput and small control flows that require low delay. These two contradictory requirements are not satisfied efficiently by conventional TCP protocols. The situation becomes worse due to the bursty nature of the data center traffic. DCTCP, a modified version of TCP that was specifically designed for data centers, was originally proposed by Alizadeh et al. in 2010. DCTCP uses an Active Queue Management (AQM) policy where the router marks a packet using Explicit Congestion Notification (ECN), when the number of packets in the queue is more than a marking threshold (K). DCTCP provides lower end-to-end delay in comparison with conventional TCP. However, it tends to provide lower throughput since it uses very small buffer space. In this paper, we have modified the existing DCTCP congestion control algorithm to increase the throughput without significantly increasing delay. We also present an algorithm for dynamic delayed ACK timeout calculation. The performance of TDCTCP (our modified DCTCP algorithm), DCTCP and the TCPNewReno have been studied using OMNeT++ simulator based models. It is shown that TDCTCP provides 15% more throughput and 15% more fairness when compared to DCTCP. Queue length and end-to-end delay increases slightly for TDCTCP. Additionally, TDCTCP provides more stable throughput than DCTCP and TCPNewReno. Though TDCTCP's delay is slightly higher than that of DCTCP, it provides much better throughput than DCTCP for the same marking threshold value.