Scheduling URLLC users with reliable latency guarantees

This paper studies Ultra-Reliable Low-Latency Communications (URLLC), an important service class of emerging 5G networks. In this class, multiple unreliable transmissions must be combined to achieve reliable latency: a user experiences a frame success when the entire L bits are received correctly within a deadline, and its latency performance is reliable when the frame success rate is above a threshold. When jointly serving multiple users, a natural URLLC scheduling question arises: given the uncertainty of the wireless channel, can we find a scheduling policy that allows all users to meet a target reliable latency objective? This is called the URLLC SLA Satisfaction (USS) problem. The USS problem is an infinite horizon constrained Markov Decision Process, for which, after establishing a convenient property, we are able to derive an optimal policy based on dynamic programming. Our policy suffers from the curse of dimensionality, hence for large instances we propose a class of knapsack-inspired computationally efficient — but not necessarily optimal — policies. We prove that every policy in that class becomes optimal in a fluid regime, where both the deadline and L scale to infinity, while our simulations show that the policies perform well even in small practical instances of the USS problem.