Consider a data center network using CEE (802.1Qau) with different tenants. If a tenant is running way larger number of flows, according to 802.1Qau, he can get most of the bandwidth because 802.1Qau only enforces per-flow fairness (even distribution) of bandwidth allocation. We cannot do (1) per-user even distribution and (2) weighted max-min fairness

This paper proposes the use of AFD (approximate fair dropping) algorithm in CEE to achieve “approximate fairness” with minimal of resources. It measures the arrival rate of flows \(r_i\) and find its fair-share rate

\[r_i^{\textrm{fair}}=\frac{W_i}{\sum_j Wj}\sum_j r_j.\]

Then there will be an “AF feedback” value by quantizing \([1-r_i^{\textrm{fair}}/r_i]^+\), which is zero when the fair-share rate is higher then current rate and more positive if the current rate exceeds more from the fair-share rate. The overall feedback value is then computed by \(Fb=(1-\alpha)Fb_{\textrm{QCN}}+\alpha Fb_{\textrm{AF}}\), which is higher than 802.1Qau’s feedback only when we have a large \(Fb_{\textrm{AF}}\) and smaller than that otherwise.

This per-flow (or per-user) feedback means to let a under-privileged flow increase its rate and requests a overshooting flow to decrease its rate. By assigning different weights, we can achieve weighed max-min fairness, in a time-averaged mannar.

Bibliographic data

@inproceedings{
   title = "AF-QCN: Approximate Fairness with Quantized Congestion Notification for Multi-tenanted Data Centers",
   author = "Abdul Kabbani and Mohammad Alizadeh and Masato Yasuda and Rong Pan and and Balaji Prabhakar",
   booktitle = "Proc. HotI",
   year = "2010",
}