AUTHOR=Mazied EmadElDin A. , Nikolopoulos Dimitrios S. , Hanafy Yasser , Midkiff Scott F. 

TITLE=Auto-scaling edge cloud for network slicing

JOURNAL=Frontiers in High Performance Computing

VOLUME=Volume 1 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/high-performance-computing/articles/10.3389/fhpcp.2023.1167162

DOI=10.3389/fhpcp.2023.1167162

ISSN=2813-7337

ABSTRACT=Next-generation wireless networks feature dynamic instantiation and termination of on-demand virtual radio access networks (RAN slices). Autoscaling edge computing resources is the essence of RAN slicing design. Autoscaling includes vertical scaling (VS) and horizontal scaling (HS). VS tailors resource allocation for running slices, within predefined limits, according to the variations in demands. If demand growth mandates allocating more resources than predefined limits, the system employs HS to instantiate a container. However, starting a container (taking 0.5 - 5 seconds) would violate the processing time required for RAN slices (0.5 - 50 milliseconds). Therefore, removing resource limits from slice configuration could contribute to meeting processing time requirements by avoiding the instantiation of a new slice. However, the lack of computational resources at the edge necessitates carefully considering resource usage per RAN slice. We address the challenge of removing resource limits to circumvent starting a new container by leveraging the capability of centralized logic in a RAN slicing architecture (i.e., slicing controller) to determine resource limits as outcomes of a decision-making process. This work introduces a resource control agent (RC) that determines resource limits as the number of computing resources packed into containers to meet the stochastic variations in demands. We aim to minimize the container deployment cost while the processing time is maintained below a threshold. RAN slicing workload (demand) is modeled using the most expensive radio function in wireless service architecture that supports surveillance and industry 4.0 use-cases, the Low-Density Parity Check (LDPC) decoding algorithm. Gaussian random LDPC iterations manifest variations in demands. Therefore, we formulate the problem as a variant of the stochastic bin packing problem (SBPP) where RAN slices (bins) need to be packed with computing resources (items) to satisfy the random variations in radio workload (stochastic demands). This paper embraces LDPC asymptotic analysis and Roofline’s definition of computing peak performance for problem formulation. Furthermore, we adopt chance-constrained programming to approach the SBPP resource control (S-RC) problem. The numerical evaluation demonstrates that S-RC maintains processing time requirement with higher probability than the case of configuring RAN slices with predefined limits. However, S-RC introduces 45% overall average cost overhead.