Abstract
Fog Computing is a compelling paradigm that shifts information processing to the network edge. However, existing models often rely on Cloud-centric programming that struggles with the resource-scarce and dynamic nature of IoT environments. This paper introduces HORNET (HOlistic pRocessing and NETworking), a Software Defined Networking (SDN) solution that jointly optimizes data processing and networking. By utilizing a Value-of-Information (VoI) metric and a Multi-Layer Routing (MLR) approach, HORNET enables "acceptable lossiness"—prioritizing high-value data while dropping low-utility packets to maintain system performance under stress.
Introduction
The proliferation of the Internet of Things (IoT) has driven the need for Fog Computing, which brings computational power closer to sensors and users. Despite its potential, current Fog implementations often treat the edge as "Cloud-lite," failing to address the unique constraints of edge nodes. When congestion occurs, traditional systems lack the context to distinguish between critical alerts and routine background data.
HORNET addresses this gap by integrating the management of service deployment and network routing. It moves away from standard Quality of Service (QoS) metrics toward a utility-based model, ensuring that the most relevant information reaches its destination even when resources are pushed to their limits.
The HORNET Architecture
The core of HORNET is an SDN-based framework that operates across two distinct but interconnected layers: the Service Layer and the Communication Layer.
1. Value-of-Information (VoI)
VoI serves as the unifying metric for the entire system. Unlike technical metrics like latency or jitter, VoI quantifies the actual utility of a data object to the end-user. This allows HORNET to treat resource allocation as a single-objective optimization problem, simplifying the decision-making process for task prioritization and packet shedding.
2. Multi-Layer Routing (MLR)
HORNET employs a flexible routing strategy that can dynamically switch between three modes:
- Pure IP Routing: Maintains compatibility with legacy applications.
- Overlay-based Dispatching: Uses flow identifiers to manage collaborative processing across nodes.
- VoI-aware Dissemination: A content-aware approach that routes and prioritizes packets based on their calculated value.
3. Holistic Management
A centralized Fog SDN Controller oversees both layers. It manages the lifecycle of software components (the Service Layer) and configures network flows (the Communication Layer) simultaneously, using the SPF (Sieve, Process, and Forward) middleware to track information objects as they traverse the network.
Experimental Evaluation
The researchers evaluated HORNET using a dual-pronged approach:
- Physical Testbed: A network of five Raspberry Pi 3 Model B+ nodes arranged in a "kite" topology. This setup was used to measure the real-world overhead of rule enforcement and the speed of network reconfiguration.
- Smart City Simulation: A large-scale simulation of downtown Washington D.C. using the Phileas discrete event simulator. The scenario included environmental sensors and traffic cameras, testing the system's ability to handle heterogeneous IoT data.
Key Results and Insights
The experimental results highlight HORNET's efficiency and agility:
- Resource Efficiency: In the Smart City simulation, the system maintained a high Quality of Experience (QoE) while processing only 23.6% of video frames and 14.4% of pollution data. By ignoring "boring" or redundant data, HORNET significantly reduced the load on edge resources.
- Agility: Deploying new rules in the VoI-aware overlay layer took just 59ms, compared to 483ms for native IP routing. This 8x improvement is attributed to avoiding the expensive context switches required to modify OS-level routing tables.
- Resilience: HORNET demonstrated the ability to deliver high-priority services even when data links were 100% saturated by temporarily suppressing lower-priority traffic.
Conclusion and Future Work
HORNET represents a significant shift in Fog Computing, moving from rigid Cloud-based models to a dynamic, value-driven approach. While the current implementation uses Genetic Algorithms for optimization, future iterations may explore Quantum-based Particle Swarm Optimization for faster convergence. Additionally, the team is working on hierarchical federation to scale the centralized controller model to massive, city-wide deployments.
Summary updated