Adaptive Task Partitioning at Local Device or Remote Edge Server for Offloading in MEC

Jianhui Liu, Qi Zhang

Introduction

Mobile Edge Computing (MEC) is a pivotal technology for enabling ultra-low latency in resource-constrained IoT devices. By offloading computationally intensive tasks to nearby edge servers (ES), applications like Augmented Reality (AR), autonomous driving, and VR can function smoothly. However, as task complexity grows, a single edge server may become a bottleneck. This necessitates partitioning tasks across multiple servers to leverage parallel processing.

The Challenge of Static Offloading

Current MEC research typically relies on one of two static frameworks for task partitioning:

  1. Decision on UE (DoU): The mobile device (User Equipment, UE) partitions the task locally and transmits segments to different edge servers over multiple wireless links.
  2. Decision on ES (DoE): The UE sends the entire task to a primary edge server. This primary server then partitions the task and forwards portions to secondary servers via a high-speed backhaul link.
Static offloading frameworks and their timing diagrams.

Neither approach is universally optimal. DoU is sensitive to the quality of wireless links to secondary servers, while DoE depends heavily on the primary server's load and the backhaul's capacity. This paper addresses the need for a dynamic selection mechanism that adapts to real-time network conditions.

Proposed Adaptive Framework

The authors propose a framework that dynamically chooses between DoU and DoE to minimize average latency. This involves jointly optimizing task partitioning ratios ($\lambda$), resource block (RB) allocations, and CPU frequency assignments.

The optimization problem is formulated as a non-convex program. To solve it efficiently, the researchers developed an Interior Point Algorithm (IPA) that relaxes integer constraints, allowing for rapid computation of near-optimal resource allocations.

The Decision Rule

A key contribution of this work is the derivation of a closed-form decision rule (Theorem 1). This rule identifies the "inflection point" where the optimal strategy switches from DoE to DoU based on the current state of the network.

Decision regions for DoE and DoU schemes based on link quality and resource ratios.

As shown in the figure above, the choice primarily depends on:

  • Wireless Link Quality Ratio ($r_i^s/r_i^p$): The relative strength of the channel to the secondary server versus the primary server.
  • Computation Resource Ratio: The relative processing power available at each server.

Performance Evaluation

The framework was tested using numerical simulations modeled after 5G MEC environments (20 MHz bandwidth, 100 RBs, and 4K VR-sized tasks). Key findings include:

  • Latency Reduction: The adaptive algorithm consistently outperformed static DoU or DoE, showing up to a 14% improvement in scenarios with limited primary server capacity.
  • Backhaul Impact: High-speed forwarding links (e.g., 1 Gbps) favor DoE, as the overhead of local partitioning on the UE outweighs the speed of the backhaul.
  • Robustness: The scheme correctly identifies when wireless transmission is the bottleneck, opting for local partitioning (DoU) when the primary server's wireless link cannot handle the full task load efficiently.

Conclusion and Future Directions

This research demonstrates that adaptively choosing the partitioning point is crucial for meeting the strict latency requirements of next-generation IoT. While the current model assumes tasks are independently partitionable, future work could extend this to complex tasks with inter-dependencies (DAGs) and incorporate energy consumption constraints to further optimize for mobile device battery life.

Summary updated

BibTeX
@inproceedings{liu2020adaptive,
  author = {Jianhui Liu and Qi Zhang},
  title = {{Adaptive Task Partitioning at Local Device or Remote Edge Server for Offloading in MEC}},
  booktitle = {WCNC},
  year = {2020},
  doi = {10.1109/wcnc45663.2020.9120484},
}
Figure