Adaptive Approximate Fair Queueing for Shared-Memory Programmable Switches

Fair Queueing (FQ) is an ideal fair bandwidth allocation scheme but is rarely deployed in production networks due to its high complexity. Driven by the prevalence of commercial programmable switching ASICs (e.g., Broadcom Trident 4, Cisco Silicon One, and Intel Tofino), several recent approaches have shown that FQ can be approximated with limited FIFO queues. However, these approaches (implicitly) assume that each queue has a dedicated buffer, while commodity switching chips usually employ a globally shared memory and dynamically allocate buffer to each queue. When directly applied to shared-memory switches, these approaches are inadaptive to the traffic dynamics. In this paper, we reveal this problem with simulations and explore the intrinsic trade-off between buffer efficiency and fairness. Based on the observations, we design Adaptive Approximate Fair Queueing (A <tex-math notation="LaTeX">²</tex-math> FQ) , a practical approximate FQ algorithm that is adaptive to traffic dynamics. At its heart, A <tex-math notation="LaTeX">²</tex-math> FQ dynamically changes the number of effective queues according to traffic characteristics. Extensive experiments and simulations show that A <tex-math notation="LaTeX">²</tex-math> FQ can improve fairness by up to 19.7×.