Towards approximate fair bandwidth sharing via dynamic priority queuing

We tackle the problem of a network switch enforcing fair bandwidth sharing of the same link among many TCP-like senders. Most of the mechanisms to solve this problem are based on complex scheduling algorithms, whose feasibility becomes very expensive with today's line rate requirements, i.e. 10-100 Gbit/s per port. We propose a new scheme called FDPA in which we do not modify the scheduler, but instead we use an array of rate estimators to dynamically assign traffic flows to an existing strict priority scheduler serving only few queues. FDPA is inspired by recent advances in programmable stateful data planes. We propose a design that uses primitives common in data plane abstractions such as P4 and OpenFlow. We conducted experiments on a physical 10 Gbit/s testbed, we present preliminary results showing that FDPA produces fairness comparable to approaches based on scheduling.