Abstract
The Multi-access Edge Computing (MEC) and Fog Computing paradigms are enabling the opportunity to have middleboxes either statically or dynamically deployed at network edges acting as local proxies with virtualized resources for supporting and enhancing service provisioning in edge localities. However, migration of edge-enabled services poses significant challenges in the edge computing environment. In this paper, we propose an edge computing platform architecture that supports service migration with different options of granularity (either entire service/data migration, or proactive application-aware data migration) across heterogeneous edge devices (either MEC-based servers or resource-poor Fog devices) that host virtualized resources (Docker Containers). The most innovative elements of the technical contribution of our work include i) the possibility to select either an application-agnostic or an application-aware approach, ii) the possibility to choose the appropriate application-aware approach (e.g., based on data access frequencies), iii) an automatic edge services placement support with the aim of finding a more effective placement with low energy consumption, and iv) the in-lab experimentation of the performance achieved over rapidly deployable environments with resource-limited edges such as Raspberry Pi devices.
Introduction
As we move toward 5G, Industry 4.0, and real-time AR/VR, low latency is non-negotiable. Edge computing (MEC and Fog) solves this by placing resources near the user. However, users are mobile. When a user moves out of range of one edge node, their active services and associated data must "follow" them to a new node to maintain a seamless experience.
Traditional Virtual Machine (VM) migration is too heavy for the edge, consuming excessive bandwidth and time. While containers like Docker are more lightweight, migrating a multi-gigabyte container still causes significant service interruption. Furthermore, the edge is heterogeneous; a service might need to move from a high-performance server to a power-limited gateway. Existing solutions often ignore these hardware differences and the specific characteristics of the data being moved. This paper introduces a flexible framework for migrating Docker-based services across heterogeneous edge nodes, ranging from powerful MEC servers to resource-constrained Raspberry Pis.
Proposed Migration Framework
The proposed multi-layer architecture treats the Service Layer (business logic) and Data Layer (state) as distinct entities. This separation allows for more granular control over what is moved and when. The core of the system resides within the Edge Node, which manages the migration logic through several specialized modules.
Differentiated Migration Modes
The framework supports two primary modes of operation:
- Application-Agnostic: A baseline mode that moves the entire container without knowing its internal logic. This is useful for simple services where the overhead of analysis is not justified.
- Application-Aware: A smarter mode that uses a Decision Module to analyze data access frequencies. By understanding the data access pattern, the system can prioritize which data chunks are critical for immediate service resumption.
Proactive Handoff and Energy Optimization
Instead of waiting for a disconnection, the system uses mobility prediction—based on historical paths or signal strength (RSSI)—to start moving "cold data" (rarely accessed records) to the target node in advance. For short-term handoff windows, the system prioritizes moving data chunks with the highest migration probability first.
To address the heterogeneity of edge hardware, the authors introduced a novel Energy Affinity Placement algorithm. This matches a service's resource needs to a node's capacity, aiming for an "Energy Affinity" near 1.0. This ensures that resource-poor Fog nodes are not overloaded while high-performance MEC nodes are utilized efficiently.
Experimental Evaluation
The researchers evaluated their framework using a mix of real-world experimentation and simulation. The in-lab testbed consisted of three Linux boxes (MEC nodes) and a Raspberry Pi 3 (Fog node) running Docker 18.09. They tested the system using a tiny Busybox container and a large (>1GB) OpenCV-based face recognition application to simulate heavy data loads.
Key Insights
- Data "Temperature" Matters: By identifying "cold" data that hasn't been updated recently, the system can safely move the bulk of the state before the user even switches nodes, significantly reducing the final handoff time.
- The Variability Trade-off: Proactive migration is most effective when data variability is low. If data changes too rapidly, the cost of reconciling changed data (using hash values to ensure consistency) can outweigh the benefits of moving it early.
Results and Discussion
The proactive migration strategies achieved total migration time reductions of 30% to 50% compared to reactive baselines. In a 10Mb/s Wi-Fi environment, migrating a 1GB face recognition container reactively took over 1500 seconds—a massive service gap. The proactive framework significantly mitigated this by pre-loading the service and cold data.
Furthermore, the placement logic successfully identified target nodes that provided the necessary Quality of Service (QoS) while minimizing the energy footprint on battery-powered edge devices.
Limitations and Future Work
While promising, the system's efficiency is heavily reliant on the accuracy of mobility prediction. Unexpected user movements can lead to wasted migration effort. Additionally, the current implementation is optimized for specific database structures like MongoDB; expanding this to general, unstructured filesystems remains a challenge. Future work will focus on handling multiple, simultaneous migration sessions and testing the solution in geographically distributed federated cloud environments.
Summary updated