Source-linked AI summary
Constella: A Novel Framework for Cost-Efficient Distributed AI Inference in LEO Space Data Centers
Andrija Stanisic, Milos Gravara, Juan Luis Herrera, Stefan Nastic
TL;DR
Heterogeneous LEO constellations make it difficult to choose a cost-effective processor-to-communicator composition while meeting distributed AI inference demands. Constella addresses this with offline split-layer and resource optimization plus online telemetry-driven routing, achieving lower cost and latency while retaining high inference success.
Problem
The central problem is selecting heterogeneous constellation size and role ratios for AI inference without over-provisioning expensive processors or limiting system efficiency.
Method
Constella uses OCRI to jointly optimize the DNN split layer and processor/communicator counts offline, then uses LIA to assign each inference result online using telemetry and ground-contact timing.
Results
Constella maintains at least 81.9% inference success while reducing deployment cost by one to two orders of magnitude, mean latency by up to 2.7×, and total energy by up to 74×.
Takeaways & Limitations
The framework provides a cost-efficient approach for distributed inference across heterogeneous LEO constellations under workload and operational constraints.
Takeaways & Limitations
The evaluation assumes homogeneous role-specific satellite parameters and excludes inference latency from optimization by requiring the task interval to suffice for full onboard DNN execution.
Abstract
from arXiv · showhide
Space data centers built from Low-Earth Orbit (LEO) satellite constellations are gaining increasing attention as a scalable computing infrastructure. With access to abundant solar energy and high-throughput optical inter-satellite links, such constellations can run AI workloads directly in orbit, enabling new in-space application types while optimizing existing ones such as Earth observation. However, managing satellite constellations that combine heterogeneous satellite roles introduces a cost optimization challenge. Determining the appropriate constellation size and satellite role ratio for a given workload is challenging, as over-provisioning processing satellites increases system cost, while under-provisioning limits system efficiency. To enable cost-efficient execution of AI inference workloads in such space data centers, we present Constella, a novel framework that leverages DNN splitting for distributed AI inference in LEO satellite constellations. Constella comprises an offline resource identifier that determines the optimal ratio of processor-to-communicator satellites and an online assignment algorithm. The algorithm utilizes constellation telemetry to adaptively route data within the constellation and to ground stations. We evaluate Constella on a real-world satellite dataset across scenarios of increasing complexity. Results demonstrate a reduction in system cost by up to two orders of magnitude and lower end-to-end inference latency by up to 2.7x compared to other approaches, while maintaining no less than 81.9% inference success rate.
1 Introduction
Constella addresses the cost and coordination challenges of distributed AI inference across heterogeneous LEO constellations by jointly planning resources offline and routing inference data online. Its evaluation reports lower cost, latency, and energy use while maintaining inference success.
- LEO constellations offer solar-powered compute and optical inter-satellite links for running AI workloads and supporting Earth observation.
- Processor satellites are more expensive than communicator satellites, making the processor-to-communicator ratio a key cost-design decision.
- Constella jointly determines the DNN split layer and minimum-cost constellation composition offline, then assigns processors to communicators online using ISL communication.
- Constella maintains at least 81.9% inference success while reducing deployment cost by one to two orders of magnitude, mean end-to-end latency by up to 2.7×, and total energy consumption by up to 74×.
- OCRI solves the offline configuration problem as a mixed integer linear program, while LIA performs latency-aware online assignment using communicator state and ground-contact timing.
2 Motivation & Related Work
The paper motivates Constella with disaster-detection workloads that require timely inference under heterogeneous satellite costs, dynamic communication windows, and limited resources. It positions the framework against prior centralized or pre-planned approaches by targeting decentralized per-task assignment under joint constraints.
- 2.1 Motivational Use Case: Disaster monitoring benefits from early detection and prediction because wildfires, earthquakes, and floods cause substantial human and economic losses.
- 2.1 Motivational Use Case: The use case separates processors, which capture imagery and compute inference, from lower-cost communicators dedicated to downlink transmission.
- 2.1 Motivational Use Case: Thousands of satellites and increasingly complex DNNs create combinatorial growth in constellation-composition decisions, motivating automated cost-efficient provisioning.
- 2.1 Motivational Use Case: Runtime assignment must balance transfer delay, buffering, communicator overload, energy feasibility, and changing ground-contact windows.
- 2.2 Related Work: Prior routing approaches compute decisions centrally over pre-planned schedules, whereas the stated gap is decentralized per-task processor-to-communicator assignment under buffer, energy, and communication-window constraints.
3 Constella Framework Overview
Constella combines offline constellation and split-layer optimization with online, telemetry-driven routing. OCRI selects a minimum-cost feasible configuration, while LIA routes each inference result to an eligible communicator or falls back to direct downlink.
- Framework Overview: The framework runs OCRI before deployment to determine the minimum-cost configuration, then uses LIA during execution to assign processors to communicators.
- System Model: Processors provide sensing, computation, ISL, and downlink capabilities, while communicators provide ISL and downlink capabilities at lower assumed usage cost.
- System Model: The model assumes known hardware efficiencies, energy budgets, throughput, homogeneous role-specific hardware, uniformly interleaved satellites, and negligible ISL transmission cost relative to ground transmission.
- OCRI: The split layer determines which DNN layers execute onboard and induces cumulative computation W(l) and intermediate-data size D(l).
- OCRI: OCRI minimizes weighted processor and communicator deployment cost while enforcing downlink-capacity, processor-energy, aggregate-energy, and integer split-layer constraints.
- LIA: LIA independently evaluates each inference task, filters communicators by contact timing, buffer capacity, and energy feasibility, and selects the eligible communicator with earliest ground contact.
- LIA: If no communicator is eligible, LIA uses direct downlink when processor energy is sufficient; otherwise, transmission is infeasible.
4 Evaluation
The evaluation compares Constella with naive and traditional baselines across cost, success rate, latency, energy consumption, and execution overhead. Across scenarios, Constella identifies lower-cost configurations while maintaining inference completion and reducing latency and energy use.
- Experimental setup: The evaluation uses naive and traditional baselines spanning over-provisioned split inference and communication-only processing.The naive baseline uses all satellites with a midpoint split and static assignments, while the traditional baseline transmits raw imagery directly to the ground station.
- Evaluation objectives: The experiments assess cost-success trade-offs, end-to-end latency, per-orbit energy consumption, and offline and online computational overhead.These dimensions evaluate minimal-cost task completion, adaptive routing, constellation energy balance, and practical feasibility.
- Cost and success rate: 81.9% minimum success rate accompanies one-to-two-orders-of-magnitude lower deployment cost than TB and NB across scenarios.Constella achieves 100% success at minimal cost in extra-small and small scenarios; fixed baseline configurations degrade under energy and buffer limits in the extra-large scenario.
- Inference latency: 1.1–2.7× lower mean latency is reported for Constella across scenarios compared with TB and NB.In the medium scenario, mean latency is 1541 s for Constella, versus 4129 s for NB and 3518 s for TB.
- Energy consumption: 3.5–45.5× lower energy than TB and up to 74× lower than NB is achieved in all scenarios except medium.The medium exception reflects TB using less energy because it activates no communicators, while Constella achieves lower cost and latency.
5 Conclusion
Constella addresses the trade-off among deployment cost, energy consumption, and execution performance in distributed space inference. Across scenarios, it combines offline constellation and split-layer optimization with online adaptive assignment while maintaining inference success and low execution overhead.
- Constella maintains at least 81.9% inference success while achieving one to two orders of magnitude lower deployment cost, up to 2.7× lower mean latency, and up to 74× lower total energy consumption than TB and NB.
- The framework jointly optimizes constellation configuration and split-layer selection offline, then performs adaptive processor-to-communicator assignment online.
- OCRI and LIA remain lightweight, with mean runtimes below 24 ms per configuration and 5 ms per orbit, respectively.
- Future work will model inference time for hardware heterogeneity and optimize satellite role placement within constellation topology, including hybrid roles.