Mobility-Included DNN Partition Offloading from Mobile Devices to Edge Clouds

Area of Science:

• Computer Science
• Artificial Intelligence
• Mobile Computing

Background:

• Deep Neural Networks (DNNs) offer advanced capabilities but face challenges with latency and energy consumption on mobile devices.
• Centralized cloud DNN processing incurs significant data transfer and latency issues.
• Edge cloud offloading is a potential solution but is hindered by mobile device mobility causing failures.

Purpose of the Study:

• To develop a novel algorithm, Mobility-Included DNN Partition Offloading (MDPO), that addresses the challenges of DNN computation offloading for mobile users.
• To minimize the total latency of completing DNN jobs while accounting for user mobility.
• To ensure the algorithm's effectiveness for various DNN topologies, including chain and graphic structures.

Main Methods:

• Developed a Mobility-Included DNN Partition Offloading (MDPO) algorithm.
• Designed MDPO to adapt DNN computation partitioning and offloading strategies based on user mobility patterns.
• Evaluated MDPO's performance against local-only and edge-only execution models.

Main Results:

• MDPO significantly reduces the total latency for DNN job completion on mobile devices.
• The algorithm demonstrates improved performance compared to traditional local-only and edge-only execution methods.
• MDPO exhibits adaptability to diverse network conditions and user mobility scenarios.

Conclusions:

• The proposed MDPO algorithm effectively mitigates latency and energy consumption issues associated with DNNs on mobile devices.
• MDPO provides a robust solution for computation offloading that accounts for user mobility, enhancing DNN application performance.
• The algorithm's flexibility makes it suitable for various DNN structures and dynamic network environments.