Alternative Best Effort (ABE) for Service Differentiation: Trading Loss Versus Delay

The idea of an Alternative Best Effort (ABE) per-hop behaviour (PHB) emerged about 20 years ago. It provides a low-delay traffic class in the Internet at the expense of more packet loss than Best Effort (BE). Therefore, ABE is better suited than BE for loss-tolerant but delay-sensitive applications. Furthermore, ABE traffic should not degrade the service for BE traffic in terms of packet loss and delay. Therefore, Internet service providers may leave the choice of using BE or ABE to their customers as they achieve service differentiation without compromising other traffic. In this work, we revisit ABE and pursue the fundamental question whether an ABE service is technically feasible, how its service would look like and interact with existing transport protocols? We present a novel scheduler called Deadlines, Saved Credits, and Decay (DSCD) for combined scheduling of BE and ABE traffic. It allows to control ABE's delay advantage over BE and copes with varying bandwidth. We provide an implementation of DSCD in the Linux network stack and demonstrate its efficiency. A side product of the implementation is an efficient approximation of the exponential function in the kernel and a bandwidth estimation method that even works at moderate link utilization. We study DSCD in a semi-virtualized testbed with real networking stacks to understand implications for transport protocols in a BE/ABE Internet. The study analyzes ABE's impact on loss and delay under various conditions and gives recommendations for configuration.