On the Fairness of DCTCP and CUBIC in Cloud Data Center Networks

Data center networks (DCNs) hold a key role in information industry as the infrastructure for hosting cloud data services and related applications. The increasing number of customers using cloud services sets, however, new challenges for the efficient management of the DCN traffic handled by the transport protocols. TCP is adapted as the main transport control protocol in DCNs, but the standard congestion control algorithm it employs has been found to be very inefficient in providing a fair sharing of the available channel capacity among the competing DCN traffic flows. DCTCP and CUBIC are two promising TCP variants designed to resolve the fairness issues identified in these environments. In this paper, we evaluate the fairness properties of these two TCP variants in comparison to the conventional TCP Reno with the use of network simulator 3 (ns-3) in realistic DCN scenarios. Our results show that DTCP improves substantially the fairness properties of TCP, but it requires more buffer space for large number of flows; it also requires the intermediate switches to be Explicit Congestion Notification (ECN)-aware. CUBIC achieves very low queue occupancies, but it demonstrates low fairness when a large and a small set of flows compete at the output port of a bottleneck switch (TCP Outcast issue).