In-Orbit Processing or Not? Sunlight-Aware Task Scheduling for Energy-Efficient Space Edge Computing Networks

Weisen Liu, Zeqi Lai, Qian Wu, Hewu Li, Qi Zhang, Zonglun Li, Yuanjie Li, Jun Liu

Abstract

As satellite constellations transition toward Space Edge Computing (SEC), managing the energy consumption of on-board AI tasks becomes critical. Satellites rely on solar power in sunlight and batteries in the Earth's shadow (eclipse zone). Frequent deep discharges significantly degrade battery life. This paper presents PHOENIX, a novel task-scheduling framework that treats sunlight as a primary resource. By intelligently offloading computation to illuminated satellites or delaying processing until sunlight is available, PHOENIX reduces battery stress by up to 54.8% and can extend operational lifespans by nearly 3x without compromising task deadlines.

Introduction

The deployment of "mega-constellations" like Starlink and OneWeb has ushered in the era of Space Edge Computing (SEC). Processing data—such as wildfire detection or ship tracking—directly on satellites reduces latency and saves downlink bandwidth. However, satellites are power-constrained. They operate in a cycle of "sunlight zones" (solar power) and "eclipse zones" (battery power).

Existing offloading techniques often ignore this cycle, leading to high Depth of Discharge (DoD) levels that prematurely age satellite batteries. The core challenge addressed by PHOENIX is: How can we schedule complex AI tasks across a constellation to minimize battery wear while meeting strict real-time deadlines?

The PHOENIX Framework

PHOENIX combines centralized planning with distributed execution to optimize energy usage across the constellation.

PHOENIX framework overview showing sunlight/eclipse zones and power management logic.

The framework operates through a three-tiered scheduling logic:

  1. Sunlight-Aware Orbit Assignment: A centralized mission center predicts sunlight windows and uses a Knapsack-based algorithm to pre-allocate "alternative orbit sets." This allows orbits with high task loads to offload to those with surplus solar energy.
  2. On-board Decision Making: Local managers choose between four execution paths for each task: immediate local processing, delayed local processing (waiting for sunlight), satellite-to-satellite offloading, or ground offloading.

Experimental Setup

The authors validated PHOENIX using a Hardware-in-the-loop (HiL) testbed. They utilized Nvidia Jetson AGX Orin modules to mimic satellite hardware, integrated with StarryNet for LEO network emulation. The study used real-world trajectory data for Starlink and OneWeb, alongside computer vision tasks like ship detection (YOLO on HRSID dataset) and wildfire segmentation (U-Net on Amazon Rainforest imagery).

Key Results and Insights

The primary insight of PHOENIX is shifting the optimization metric from raw energy consumption to Battery DoD. Energy consumed in the sun is "free," while energy from the battery is "expensive" due to hardware degradation.

Impact of Depth of Discharge (DoD) on battery cycle number and operational lifetime.

Key performance metrics include:

  • Energy Efficiency: PHOENIX reduced the maximum battery DoD by 54.8% compared to state-of-the-art baselines.
  • Lifespan Extension: By minimizing battery stress, the framework is estimated to prolong satellite battery lifetime by 2.9x to 5.3x.
  • Deadline Compliance: Unlike simple ground-offloading which suffers from downlink congestion, PHOENIX consistently met task deadlines by leveraging in-orbit processing during sunlit windows.

Conclusion

PHOENIX demonstrates that sunlight awareness is crucial for sustainable Space Edge Computing. By treating the sun/shadow cycle as a scheduling constraint, constellations can significantly extend their operational life. Future work will explore multi-hop ground offloading and adaptive adjustments for unexpected task bursts in heterogeneous hardware environments.

Summary updated

BibTeX
@inproceedings{liu2024orbit,
  author = {Weisen Liu and Zeqi Lai and Qian Wu and Hewu Li and Qi Zhang and Zonglun Li and Yuanjie Li and Jun Liu},
  title = {{In-Orbit Processing or Not? Sunlight-Aware Task Scheduling for Energy-Efficient Space Edge Computing Networks}},
  booktitle = {INFOCOM},
  year = {2024},
  doi = {10.1109/infocom52122.2024.10621268},
}
Figure