PaperlayerView exact revision on arXiv

Source-linked AI summary

Joint DNN Partitioning and Resource Allocation for Satellite-Terrestrial Collaborative Inference Systems

Wenyu Liang, Zesong Fei, Peng Liu, Xinyi Wang, Ming Zeng

arXiv:2609.34543v1eess.SP

TL;DR

The paper addresses joint DNN partitioning and resource allocation for satellite–terrestrial inference, where satellite constraints and shared GS resources make the decisions coupled. It develops a two-layer optimization algorithm and reports lower latency than binary offloading, with gains over fixed resource-allocation schemes as well.

  • Problem

    Satellite–terrestrial DNN partitioning remains underexplored despite limited onboard resources and competition for shared bandwidth and GS computing capacity.

    Source: §I · p. 1

  • Method

    A two-layer algorithm combines closed-form updates and nested bisection for continuous resources with random-restart coordinate descent for partitioning.

    Source: §I · p. 2, 1) Inner-Layer Continuous Resource… · p. 3, 2) Outer-Layer DNN Partitioning Problem · p. 4

  • Results

    The proposed scheme reduces task completion latency compared with binary offloading, and reported gains over fixed resource-allocation schemes support jointly optimizing partitioning and communication-computing resources.

    Source: §I · p. 2, §V · p. 5

  • Takeaways & Limitations

    Jointly adapting partitioning and communication-computing resources reduces latency while satisfying satellite energy constraints in the studied multi-satellite system.

    Source: §I · p. 1, §V · p. 5

Abstract

from arXiv · show

This letter investigates a satellite-terrestrial collaborative inference system for remote sensing information processing, where a ground station (GS) serves multiple low Earth orbit (LEO) satellites running heterogeneous deep neural network (DNN) inference tasks. By partitioning the DNN between the satellite and the GS, each satellite executes the front-end layers locally, transmits the resulting intermediate features to the GS, thereby offloading the remaining layers to the GS for inference completion. We jointly optimize DNN partitioning, satellite and GS computing resources, satellite-ground bandwidth allocation, and satellite transmit power to minimize the average task completion latency under per-satellite energy constraints. To address the coupling between discrete partitioning decisions and continuous communication-computing variables, we propose a two-layer optimization algorithm. The inner layer performs closed-form updates and efficient nested bisection searching to update the continuous resource variables, while the outer layer employs random-restart coordinate descent to refine the partitioning decisions. Simulation results show that the proposed algorithm effectively reduces task completion latency compared with conventional binary offloading.

I. INTRODUCTION

Binary offloading cannot exploit layer-wise differences in DNN workload and feature size, while satellite–terrestrial partitioning remains underexplored despite shared resource constraints. The paper jointly optimizes partitioning and communication-computing resources to reduce average task completion latency under resource and energy constraints.

I. INTRODUCTION

  • DNN partitioning executes front-end layers onboard and sends intermediate features to the GS, offering a finer computation–communication tradeoff than binary offloading.Suitable bottleneck layers can produce compact features after moderate preceding computation.Source: §I · p. 1
  • Satellite–terrestrial partitioning is underexplored, and satellites contend for bandwidth and GS computing capacity while facing limited onboard computing and energy resources.These constraints couple partition decisions with communication and computing allocation.Source: §I · p. 1
  • The paper jointly optimizes DNN partitioning, computing resources, satellite–ground bandwidth, and transmit power to minimize average task completion latency subject to resource and energy constraints.Source: §I · p. 1

II. SYSTEM MODEL

  • The system comprises a GS serving visible LEO satellites with varying task arrival rates and model architectures over frequency-division multiple access.Satellite–ground topology is modeled as quasi-static, and visibility depends on an elevation-angle threshold.Source: §II · p. 2
  • The satellite–ground channel model includes antenna gains, free-space path loss, elevation-dependent atmospheric and rain attenuation, and shadowed-Rician fading.Source: §II · p. 2

F DNN

The DNN model characterizes layer workloads and feature sizes, then uses candidate partition points to determine what runs onboard and what is transmitted for GS processing.

F DNN

  • Under q-bit quantization, output feature size after layer l is Datak(l) = qNk,l, and non-computational operators have negligible inference workload.These operators include the virtual input layer, Flatten, and Dropout.Source: F DNN · p. 2
  • Each satellite's candidate partition point determines transmitted feature size and splits the DNN workload between satellite-side and GS-side execution.The candidate endpoints represent full GS execution and full onboard execution, respectively.Source: F DNN · p. 2
  • Task completion latency combines satellite–GS transmission and propagation with satellite-side and GS-side computation latency.Satellite energy accounting includes onboard computation and data transmission, while static and GS-side energy are omitted.Source: F DNN · p. 2

A. Problem Formulation

The formulation minimizes average task completion latency across connected satellites by jointly choosing partition points and continuous communication-computing resources under system constraints.

A. Problem Formulation

  • The objective jointly selects DNN partitioning, satellite and GS computing allocations, satellite–ground bandwidth, and transmit power to minimize average task completion latency.Source: A. Problem Formulation · p. 3
  • Partition points are restricted to predefined candidates, while per-satellite and GS computing resources, total bandwidth, and transmit power are bounded.Each satellite's energy consumption must also remain within its prescribed energy budget.Source: A. Problem Formulation · p. 3

B. Algorithm Design

The optimization is difficult because discrete partition choices interact with continuous resource allocation and coupled latency and energy terms.

B. Algorithm Design

  • The problem combines discrete DNN partitioning choices with continuous resource-allocation variables.Source: B. Algorithm Design · p. 3
  • Computation latency, transmission latency, and energy consumption are coupled in the optimization.Source: B. Algorithm Design · p. 3
  • Bandwidth and transmit power are nonlinearly coupled, motivating the proposed two-layer optimization algorithm.Source: B. Algorithm Design · p. 3

1) Inner-Layer Continuous Resource Allocation Problem:

For a fixed DNN partition, the inner layer allocates computing, bandwidth, and transmit power under coupled resource and energy constraints. KKT updates and bisection searches provide the continuous resource solution.

1) Inner-Layer Continuous Resource Allocation Problem:

  • For a given partition, the inner layer fixes the workload and feature size before optimizing continuous resource variables.Source: 1) Inner-Layer Continuous Resource… · p. 3
  • For fixed partitioning, the continuous problem couples computation, communication, and energy variables, while fixed dual variables let resource updates decompose across satellites.The Lagrangian includes multipliers for GS computing, total bandwidth, and per-satellite energy constraints.Source: 1) Inner-Layer Continuous Resource… · p. 3
  • GS computing capacity is allocated in proportion to the square root of each satellite’s ground-side computational workload.Source: 1) Inner-Layer Continuous Resource… · p. 3
  • When a satellite’s energy constraint is nonbinding, it operates at maximum computing frequency; when binding, its frequency is reduced to meet the energy budget.Source: 1) Inner-Layer Continuous Resource… · p. 3
  • For positive transmitted feature size, the total-bandwidth constraint is active, and the transmit-power scalar optimization is strictly unimodal for μ > 0.The optimal scalar is found by bisection; nested bisection updates energy multipliers and the bandwidth multiplier to satisfy their constraints.Source: 1) Inner-Layer Continuous Resource… · pp. 3–4

2) Outer-Layer DNN Partitioning Problem:

The outer layer chooses discrete DNN partition points and evaluates each using the inner layer’s minimum latency. Random-restart coordinate descent searches this space without exhaustive enumeration.

2) Outer-Layer DNN Partitioning Problem:

  • The outer layer selects the DNN partitioning vector, which specifies workloads and transmitted feature sizes for the inner-layer optimization.Source: 2) Outer-Layer DNN Partitioning Problem · p. 4
  • Because exhaustive search over the partition choices is intractable, the method uses random-restart coordinate descent to approximate the global optimum.Source: 2) Outer-Layer DNN Partitioning Problem · p. 4
  • Coordinate descent tests candidate values for one satellite at a time and stops at the iteration limit or when a full sweep yields no strict objective reduction.Multiple restarts from predefined representative configurations and randomly sampled feasible partitions reduce sensitivity to local minima.Source: 2) Outer-Layer DNN Partitioning Problem · p. 4

3) Overall Algorithm and Complexity:

The framework combines inner resource allocation with outer coordinate descent; its stated complexity depends on the bisection iterations, candidate counts, and restart and iteration limits.

3) Overall Algorithm and Complexity:

IV. SIMULATION RESULTS

Monte Carlo simulations evaluate one GS serving eight LEO satellites with heterogeneous inference tasks against five resource-allocation and offloading benchmarks. The proposed algorithm adapts partitioning and resource choices across computing, bandwidth, and energy settings, achieving lower latency in the reported comparisons.

IV. SIMULATION RESULTS

  • The simulated system has one GS and eight LEO satellites; half run AlexNet-based tasks and half run ResNet18-based tasks.The benchmarks include AO-GS, random partitioning, equal bandwidth, fixed transmit power, and binary offloading.Source: §IV · p. 4
  • The proposed algorithm consistently achieves the lowest latency as satellite computing capacity varies, although performance gaps narrow as onboard resources increase.Fig. 2(b) reports that partitioning becomes less critical when onboard resources are abundant.Source: Figure 2 · p. 4, §IV · p. 5
  • As GS computing capacity increases, latency decreases for all schemes and the proposed algorithm’s advantage over partitioning baselines becomes more pronounced.The authors attribute this to adapting partitioning to use added GS resources without excessive feature transmission.Source: Figure 2 · p. 4, §IV · pp. 4–5
  • Joint bandwidth allocation and partitioning outperform equal bandwidth, while the proposed algorithm’s advantage over fixed transmit power is more evident under stringent energy constraints.Increasing bandwidth and energy budgets reduces latency; the energy-budget comparison links the advantage to a better computation–transmission energy tradeoff.Source: Figure 3 · p. 5, §IV · p. 5

V. CONCLUSION

The letter formulates satellite–terrestrial collaborative inference as a joint latency-minimization problem and addresses its coupled discrete and continuous decisions with a two-layer optimization algorithm. Simulations report gains over alternative partitioning and resource-allocation schemes.

V. CONCLUSION

  • The problem minimizes average task completion latency by jointly optimizing DNN partitioning, satellite- and GS-side computing, bandwidth allocation, and satellite transmit power.Source: §V · p. 5
  • The two-layer algorithm uses closed-form updates, nested bisection searching, and random-restart coordinate descent for coupled discrete and continuous variables.Source: §V · p. 5
  • Gains over binary offloading and fixed resource-allocation schemes demonstrate the benefits of jointly optimizing DNN partitioning and communication-computing resources.Source: §V · p. 5
  • Gains over AO-GS and random partitioning validate the inner- and outer-layer optimization methods, respectively.Source: §V · p. 5

Related papers

Loading 2609.34543v1…