Abstract
We propose a novel strategy for energy-efficient dynamic computation offloading in beyond 5G networks. The goal is to minimize the energy consumption of the overall system—comprising multiple User Equipment (UE), an access point (AP), and an edge server (ES)—under constraints on end-to-end service delay and packet error rate. To reduce energy consumption, we exploit low-power sleep operation modes, shifting the edge computing paradigm from an "always on" to an "always available" architecture. We introduce the Discontinuous Computation Offloading (DisCO) algorithm, which hinges on Lyapunov stochastic optimization to orchestrate radio and computational resources without requiring prior knowledge of traffic or channel statistics.
Introduction
As we transition toward 6G, Mobile Edge Computing (MEC) is essential for low-latency applications like autonomous driving and remote surgery. However, the "always-on" nature of current infrastructure leads to massive energy consumption. This paper introduces DisCO (Discontinuous Computation Offloading), a framework that minimizes total system energy by intelligently orchestrating sleep modes across the entire network chain.
The Challenge of Always-On Infrastructure
While computation offloading saves battery life for mobile devices, it often shifts the energy burden to Access Points (APs) and Edge Servers (ESs). In traditional deployments, these infrastructure elements remain active 24/7, consuming significant power even when idle. Existing research typically optimizes only one segment of the chain (e.g., the UE battery) or fails to account for the time-varying nature of wireless channels and task arrivals.
The objective of DisCO is to minimize total system energy while meeting two critical constraints:
- Average End-to-End (E2E) Delay: The total time from data generation at the UE to result reception.
- Reliability: Ensuring the probability of exceeding a maximum delay threshold remains below a strict limit.
The DisCO Framework
The DisCO algorithm treats the UE, AP, and ES as a holistic system. It utilizes Lyapunov Stochastic Optimization to transform long-term average energy and delay constraints into per-slot optimization problems. This allows for real-time decision-making based only on current observations.
As shown in Figure 1, the system is modeled as a series of queues: local queues at the devices ($Q_k^l$), transmission queues at the AP ($Q_k^a$), and computation queues at the ES ($Q_k^m$). The "Slot organization" detail highlights how time is partitioned to accommodate control signaling, computation/offloading, and transitions between sleep and active states.
In every time slot (approx. 10ms), DisCO simultaneously decides:
- Sleep Modes: Which entities (UE, AP, ES) should enter low-power states.
- Uplink/Downlink Parameters: Modulation and Coding Schemes (MCS) and transmit power.
- Computational Resources: CPU frequency allocation at the Edge Server.
Key Insights and Results
The core innovation of DisCO is the shift from "always on" to "always available." By synchronizing sleep cycles across the UE, AP, and ES, the system "wakes up" just in time to handle data bursts and then immediately returns to a low-power state.
Numerical results demonstrate several key advantages:
- Energy-Delay Trade-off: Operators can tune a single parameter ($V$) to prioritize either extreme energy savings or ultra-low latency.
- Superiority over Baselines: DisCO significantly outperforms strategies that lack sleep modes or only allow radio-level sleep. In high-traffic scenarios, it achieves up to a 10% energy gain through intelligent resource allocation.
- Reliability Guaranteed: The adaptive mechanism successfully maintains the out-of-service probability below target thresholds (e.g., $10^{-3}$), making it viable for mission-critical applications.
Conclusion and Future Work
DisCO provides a "globally green" solution by optimizing the entire computation offloading chain. While the current model focuses on a single-cell scenario, it establishes a foundation for energy-efficient MEC. Future work will extend this to multi-cell environments with user mobility and explore data-driven versions of the algorithm using Deep Reinforcement Learning (DRL) to handle even more complex environment patterns.
Summary updated