This is from the mini-conference of INFOCOM. It proposed FQCN as an extension of QCN in IEEE 802.1Qau.

It is found that QCN does not provide fairness to flows. There are great rate disparity between flows according to the experiment reported in this paper. The author therefore proposed “fair QCN”, which sends feedback of \(Fb<0\) to not only the sampled flow, but all the overrated flows, i.e. those who sent at rate higher than the equal share. The congestion parameter to send to each overrated sender is given by \(Fb(i)=\frac{A_i}{\sum_{k=1}^N A_k}Fb\), which \(N\) is the total number of overrated senders, and \(A_k\) is the number of packets from flow \(k\).

According to the simulation, it is found that with QCN, an incast scenario of 16 servers can give a throughput of 400Mbps on a 1Gbps link, but FQCN can maintain 900Mbps utilization even with large number of servers. The poor performance of QCN is due to rate unfairness, which no new traffic is sent until the slowest flow of previous job is completed. By being fair, we can therefore reduce such “wait time”

Bibliographic data

@inproceedings{
   title = "On Mitigating TCP Incast in Data Center Networks",
   author = "Yan Zhang and Nirwan Ansari",
   booktitle = "Proc. INFOCOM",
   pages = "51--55",
   month = "Apr",
   year = "2011",
   address = "Shanghai, China",
}