On the Optimization of Edge Server Scaling and Placement for Open Radio Access Network (O-RAN) Slicing

Abstract

Managing edge server resources for O-RAN slicing can be viewed from two complementary perspectives: service providers, who aim to satisfy use-case requirements with minimal leased computing resources, and infrastructure providers, who seek to deploy the minimum number of edge servers across geographically distributed regions. From the service-provider perspective, this dissertation investigates static scaling of edge computing resources for O-RAN slicing workloads with stringent delay requirements. While static scaling enables proactive resource reservation, it inherently introduces over-provisioning that must be carefully minimized. To address this challenge, we develop a chance-constrained static scaling framework that accounts for workload uncertainty while guaranteeing processing time requirements. Workload variability is modeled using LDPC decoding, where iteration counts capture stochastic fluctuations induced by wireless channel conditions. Containers are modeled as bins that allocate computing resources under uncertainty, leading to a Stochastic Bin Packing Problem (SBPP) that integrates LDPC asymptotic analysis and the Roofline performance model. Results show that the stochastic formulation achieves higher reliability than deterministic designs, at the cost of moderate additional resource reservation. From the infrastructure-provider perspective, this dissertation determines the minimum number of edge servers required to support O-RAN slicing in both urban and rural regions under uncertainty. In urban environments, uncertainty arises from dynamic slicing traffic and user activity, modeled using point Poisson processes. In rural, resource-constrained regions, uncertainty is driven by wireless fronthaul behavior, modeled through a random number of retransmissions caused by power-limited communication links. We formulate the edge server placement problem as a chance-constrained mixed-integer linear program and develop two solution approaches: one based on sample-average approximation and another based on inverse cumulative distribution function (ICDF) transformations. Using realistic wireless testbed measurements, we show that under stable rural channel conditions with low retransmission variability, deterministic designs based on empirical averages closely match stochastic solutions. However, the framework explicitly characterizes the regimes in which uncertainty becomes impactful, providing a principled foundation for data-driven, reliability-aware edge server deployment.