Meeting Deadlines in Datacenter Networks: An Analysis on Deadline-Aware Transport Layer Protocols

Data centers had come to the scene as an effective framework for running a huge range of real-time and cloud applications. When TCP is deployed in a broad data center environment, it fails to maintain high throughput and deadline guarantees, mainly due to the unfair sharing and flow quenching. Several transport layer designs have been proposed over the last few years: some of them were claimed to outperform the currently used ones. This paper provides an analysis of some of the transport-layer protocols, namely D3 Deadline Driven Delivery), PDQ (Preemptive Distributed Quick), D2 TCP (Deadline-aware Data Center TCP), DCTCP (Data Center TCP) and MCP (Multiflow Conversation Protocol). The analysis includes the evaluation of their design implementation, performance and problems. The FCFS mechanisms at the heart of the D3 forces the protocol to be tied to the arrival order of the flows, which leads to the priority inversion issue resulting in flows unjustifiably missing their deadlines. The PDQ protocol has been described as a mechanism for a flow urgency regulation which emulates the EDF behavior. While PDQ has improved upon D3, it was pointed out that the management of priority lists and per-flow states in this line of protocols puts a great burden on the network core, thus restricting the efficient router’s bandwidth. Motivated by DTCP, D2 TCP and MCP utilize congestion window modeling at the end hosts to handle the deadline flows ensuring network stability and deadline meeting guarantees. Also, deployability problem is likely to be solved by recent advances of SDN introducing a centralized control over networks.