SAMUS: Slice-Aware Machine Learning-based Ultra-Reliable Scheduling
Caner Bektas, Dennis Overbeck, Christian Wietfeld
Abstract
Multiple service types such as Ultra-Reliable Low Latency Communication (uRLLC) and Enhanced Mobile Broadband (eMBB) are envisioned to be incorporated into the next generation mobile communication standard 5G based on a single physical communication network. To unite these services with partly contradicting Quality of Service (QoS) requirements, Network Slicing is considered a key technology. uRLLC slices in particular are highly demanding, requiring extremely high reliability and low latency in the single-digit milliseconds range. Consequentially, the latency impact of radio resource management on the end-to-end latency is optimized in this work by using so-called Configured Grants (CGs), which aim to minimize latency-intensive scheduling requests by pre-allocating radio resources. As predicting future traffic demands and channel conditions are required to use CGs, a data-driven machine learning-based radio resource scheduler prototype is introduced and evaluated in this work based on a specifically developed 5G radio resource simulator. The results show promising latency optimizations and possible trade-offs in uRLLC and eMBB coexistence.