Abstract
The Internet of Things has brought digitalization to intensive domains through the automation of their real-world processes. However, the criticality of these processes is reflected in high Quality of Service (QoS) requirements for the application to work properly. Moreover, business-level QoS, such as the operational cost, are also key to the feasibility of these applications. This QoS depends on three, closely-related dimensions: the application software, the computing devices and the communication network, which provide high flexibility to obtain different performances at different costs. Thus, to achieve optimal QoS in these scenarios, the application, computing and networking dimensions must be optimized, considering their crucial interplay in a joint effort. Furthermore, this solution must allow multi-objective optimization, finding the optimal trade-off between operational cost and application performance. In this paper, we present Multi-Objective SDN Fog Optimization (MO-SFO), a holistic framework that allows for the optimization of both the response time and the deployment cost. MO-SFO is evaluated over an emulated smart city case study, showing the cost and performance trade-off achieved in different topologies.
Introduction
Deploying modern Internet of Things (IoT) applications, such as smart city surveillance systems, requires a complex balancing act between low-latency performance and operational feasibility. To meet these demands, developers typically leverage Fog Computing and Software-Defined Networking (SDN). However, these infrastructure layers are often optimized in isolation: network engineers place SDN controllers, systems engineers deploy Fog nodes, and software developers distribute microservices independently. Because these layers are deeply interdependent, siloed optimization leads to sub-optimal configurations that are either excessively slow or unnecessarily expensive. This paper addresses this "coupling" problem by proposing a unified, multi-objective optimization approach.
Methodology
The authors introduce MO-SFO (Multi-Objective SDN Fog Optimization), a holistic framework designed to simultaneously optimize the placement of network controllers, fog computing nodes, and application microservices.
The core of MO-SFO is a Mixed Integer Linear Programming (MILP) model. This approach identifies the optimal configuration based on physical constraints such as RAM limits on fog nodes and network link capacities. A weighting parameter, $\alpha$, allows for flexible optimization:
- $\alpha = 0$: Minimizes total cost (CAPEX and OPEX).
- $\alpha = 1$: Minimizes application response time.
- $\alpha = 0.5$: Seeks a balanced trade-off between cost and performance.
The framework was evaluated using a realistic emulated environment built with Kathará and Docker. The case study involved five real-world analytics microservices from Intel’s OpenVINO (e.g., vehicle classification, pedestrian detection) deployed across "Small" (7 switches, 6 devices) and "Large" (20 switches, 18 devices) network topologies.
Results and Discussion
The evaluation revealed several critical insights into the interplay between infrastructure and application performance.
The Cost of Performance
A primary finding was the dominance of operational expenses. In a 5-year cost analysis, the energy consumption required for microservice execution (OPEX) accounted for 65% to 79% of the total deployment cost. Physical SDN switches also represented a significant portion of the budget (13-35%), highlighting that network topology is as critical to cost as compute power.
Optimization Trade-offs
The "Response Time" objective resulted in deployments that were 20-23% more expensive than the "Cost" objective but significantly faster. However, the "Trade-off" mode ($\alpha=0.5$) proved highly efficient. In the large topology, while the cost-only objective was 84% slower than the performance-focused setup, the trade-off version was only 11% slower while remaining significantly cheaper—costing only 3-6% more than the bare-minimum cost configuration.
Conclusion
MO-SFO demonstrates that treating networking, computing, and application layers as a single optimization problem yields superior results compared to siloed approaches. While the MILP model provides optimal solutions, its computational complexity poses challenges for scaling to massive networks. Future work will explore heuristic and meta-heuristic methods to reduce optimization time and investigate integration with orchestration tools like Kubernetes for automated, self-healing deployments.
Summary updated