Towards a Low Latency Network-Slice Resistant to Unresponsive Traffic

This paper studies what mechanisms a network must have to offer a very low-latency service to applications (featuring a maximum end-to-end packet delay). We assume very concrete requirements, not seen in the literature, that raise the challenge level: i) applications might be unresponsive to traffic warnings from the network; and ii) applications do not inform or require any network resources, exactly as the Internet works today (i.e., there is no admission control procedures). We present an architecture/algorithm with a minimum of state information and good scalability properties. Obviously, it is not applicable to the wide Internet. Even more, the architecture is not TCP-friendly (because control loops must be shorter than the Round Trip Time (RTT) magnitudes and oscillations, and packet losses are higher). Instead, it is appropriate to an end-to-end slice network based on a virtualization of the physical network with independent queues and line bandwidths. It is designed for interactive applications and for certain real-time ones. We use plain backpressure control supported by cooperation amongst the routers to isolate offending traffic. We are particularly concerned in situations of very high load, as they will be very common in the future. One objective is to reach a predictable network behaviour that in the limit (heavy network overload) is maintained, contrary to the current Internet. In the future, new pace-based congestion control algorithms for applications can be designed to take the most out of this type of network.