PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent Task Offloading for Heterogeneous V2X Communications

Kai Xiong, Supeng Leng, Chongwen Huang, Chau Yuen, Liang Guan

arXiv:2006.15855v1cs.NIcs.ITmath.IT

TL;DR

Resource-constrained onboard processors and limited vehicular communication resources make reliable, delay-aware task offloading a challenge. The paper models heterogeneous V2X offloading with stochastic network calculus and federated Q-learning; Sync-FQL has the lowest total P2 cost among the compared algorithms.

  • Problem

    VEC offloading must address limited vehicular communication resources, while prior work often overlooked joint interworking of DSRC, C-V2X, and mmWave and integrated offloading failure probability.

    Source: §II · p. 2

  • Method

    The paper derives offloading-delay upper bounds with stochastic network calculus and proposes federated Q-learning to optimize offloading under resource budgets and failure probability.

    Source: §I · p. 2, §V · p. 8

  • Results

    Sync-FQL has the lowest total P2 cost, while Greedy and Relaxation have much higher total cost than the learning-based algorithms.

    Source: A. Performance of Offloading Algorithms · p. 10

  • Takeaways & Limitations

    The delay upper bounds can guide offloading scheduling and communication-form selection, and can guide search pruning during training.

    Source: §I · p. 2, §III · p. 3

Abstract

from arXiv · show

With the rapid development of autonomous driving technologies, it becomes difficult to reconcile the conflict between ever-increasing demands for high process rate in the intelligent automotive tasks and resource-constrained on-board processors. Fortunately, vehicular edge computing (VEC) has been proposed to meet the pressing resource demands. Due to the delay-sensitive traits of automotive tasks, only a heterogeneous vehicular network with multiple access technologies may be able to handle these demanding challenges. In this paper, we propose an intelligent task offloading framework in heterogeneous vehicular networks with three Vehicle-to-Everything (V2X) communication technologies, namely Dedicated Short Range Communication (DSRC), cellular-based V2X (C-V2X) communication, and millimeter wave (mmWave) communication. Based on stochastic network calculus, this paper firstly derives the delay upper bound of different offloading technologies with a certain failure probability. Moreover, we propose a federated Q-learning method that optimally utilizes the available resources to minimize the communication/computing budgets and the offloading failure probabilities. Simulation results indicate that our proposed algorithm can significantly outperform the existing algorithms in terms of offloading failure probability and resource cost.

I. INTRODUCTION

The paper frames reliable, economical VEC offloading as a joint choice of communication technology and offloading target under resource and failure constraints. It introduces heterogeneous V2X modeling and federated learning, then describes the system and its stochastic delay analysis.

I. INTRODUCTION

  • VEC offloading addresses scarce onboard processing, but limited vehicular bandwidth and failures across communication or computing constrain reliable, economical service.The paper identifies prior VEC work as often assuming sufficient communication resources and rarely incorporating integrated communication and computing failure probability.Source: §I · pp. 1–2
  • Stochastic network calculus yields offloading-delay upper bounds for different V2X technologies, which the authors say can guide scheduling and communication-form selection.The contribution describes these as the first closed-form offloading-delay upper bounds for VEC.Source: §I · p. 2
  • The proposed Sync-FQL shares local Q-learning knowledge through a consensus Q-table, while the optimization accounts for communication and computing budgets and offloading failure probability.The authors also describe the federated scheme as parallel and designed for scalability and fast convergence.Source: §I · p. 2

II. RELATED WORK

  • The paper identifies a gap in jointly studying DSRC, C-V2X, and mmWave in heterogeneous VEC while accounting for offloading failure probability and resource cost.It notes that selecting access technologies and offloading targets is complicated by their dependency.Source: §II · p. 2

III. SYSTEM MODEL

  • Roadside VEC servers sit with cellular base stations and connect through X2 interfaces into a VEC pool, while a platoon header manages resources within a platoon.The VEC pool centrally manages resources for C-V2I offloading; the platoon header is a resource-rich vehicle elected from the platoon.Source: §III · p. 2
  • Sync-FQL runs at the VEC pool or platoon header and takes task delay and traffic attributes plus server service attributes as inputs.The listed inputs include Tmax, arrival rate λ, burstiness o, envelope service rate ξ, and peak service capacity η.Source: §III · pp. 2–3
  • The stochastic network-calculus analysis models traffic and service statistically to derive offloading performance metrics that can guide training search pruning.Traffic uses arrival rate and burstiness, while service is characterized by envelope rate and peak capacity; processing delay is bounded for a failure probability.Source: §III · p. 3

IV. PERFORMANCE ANALYSIS ON VEC

This section derives end-to-end delay upper bounds for offloading in heterogeneous VEC architectures. For platoon-based computing, it details how communication technology and its service assumptions shape those bounds.

A. Platoon-based Edge Computing

  • DSRC uses competition-based access with exponential backoff, and its offloaded tasks also share the on-board processor.The analysis models DSRC transmission as a latency-rate service with access delay before concatenation with processing.Source: 1) Delay Upper Bound of DSRC · pp. 3–4
  • The delay bound for each offloading technology is expressed for a specified failure probability, including transmission and processing delay.For DSRC, the bound is derived by treating transmission and on-board processing as concatenated services.Source: 1) Delay Upper Bound of DSRC · pp. 4–5
  • C-V2V mode 3 preassigns each communication’s resources through a nearby base station, eliminating transmission access competition but not on-board CPU competition.The delay bound combines the C-V2V communication service with on-board processing.Source: 2) Delay Upper Bound of C-V2V · p. 5
  • CmmW beam alignment uses RTS-like and CTS-like beacons, and DSRC traffic on the competition-based control channel worsens CmmW offloading performance.The mmWave service curve combines alignment and data transmission.Source: 3) Delay Upper Bound of mmWave · pp. 5–6
  • RmmW uses a reservation-based control channel; with negligible control-traffic burstiness, its delay bound is similar to C-V2V’s, so it is not explored further.The reserved control channel may be purchased from the operator or licensed in the future.Source: 3) Delay Upper Bound of mmWave · pp. 6–7

B. Infrastructure-based Edge Computing

C-V2I offloading combines uplink transmission with processing in the VEC pool.

B. Infrastructure-based Edge Computing

C. Local Processing

Local processing is bounded while accounting for competition among tasks on the vehicle’s on-board processor.

C. Local Processing

V. MODEL OPTIMIZATION

The optimization model accounts for communication and computing costs alongside offloading failure probability. Because its objective is non-convex, the paper proposes federated Q-learning and a relaxation-based solution.

V. MODEL OPTIMIZATION

  • Computing cost is assigned to VEC-pool processing and to on-board processing used by DSRC, C-V2V, and mmWave offloading.Local processing is stated to incur no communication or computing cost.Source: §V · pp. 7–8
  • The model treats offloading failure probability as an optimization objective alongside communication and computing costs.For C-V2X transmission, the cellular operator may charge a fee per Mbps.Source: §V · pp. 7–8
  • The failure probability is inversely proportional to the delay bound, and minimizing it makes the delay constraint tight.Source: §V · p. 8
  • The non-convex optimization problem is addressed with federated Q-learning and a relaxation that converts it to a convex problem with low complexity.Federated Q-learning uses parallel local learning and a consensus Q-table for aggregation.Source: §V · p. 8

A. Federated Q-Learning

The paper formulates offloading as an MDP and uses local Q-learning for each technology, followed by global aggregation into a shared consensus Q-table. It focuses on synchronous aggregation, which outperforms asynchronous aggregation in the simulations.

A. Federated Q-Learning

  • The offloading controller is modeled as an MDP whose state includes technology assignments and reserved bandwidth, with transitions defined for state-action choices.Source: A. Federated Q-Learning · p. 8, 1) State · p. 8
  • Actions adjust the traffic share assigned to offloading technologies, while the agent selects the highest-valued action or explores randomly.Share changes use increments of 0.01 or 0.1.Source: 2) Action · p. 8
  • The reward is the negative total cost, and Q-table entries index learned action-state performance by technology, application category, and offloading percentage.Source: 4) Reward · p. 8, 5) Q-Table · p. 9
  • Each technology updates a local Q-table, and an aggregator combines local tables into a consensus table that is returned to local learners.Source: 5) Q-Table · p. 9
  • Synchronous aggregation waits for all local updates and outperforms asynchronous aggregation, whose updates may use outdated models.Source: 5) Q-Table · p. 9

B. Relaxation Optimization

The Relaxation algorithm approximates the non-convex problem P2 with a convex problem P3 to obtain a suboptimal result more quickly. This simplifies computation but yields a higher total cost than P2.

B. Relaxation Optimization

VI. PERFORMANCE EVALUATION

The evaluation compares synchronous federated Q-learning with asynchronous federated Q-learning, original Q-learning, Greedy, and Relaxation.

VI. PERFORMANCE EVALUATION

  • Sync-FQL is evaluated against Asyn-FQL, original Q-learning, Greedy, and Relaxation.Source: §VI · p. 9
  • Asyn-FQL aggregates after each local learner updates, whereas original Q-learning has no global aggregation process.Source: §VI · p. 9
  • Greedy selects the maximum-reward action, while Relaxation results come from solving P3 directly with CVX.Source: §VI · p. 9

A. Performance of Offloading Algorithms

Sync-FQL has the lowest total cost among the evaluated algorithms, while the best heterogeneous framework depends on traffic load. The results also motivate different deployment choices for delay-tolerant and delay-sensitive applications.

A. Performance of Offloading Algorithms

  • The algorithm comparison uses 500 simulations, 200 training iterations for learning-based methods, five tasks per vehicle, and a default of five vehicles.Source: A. Performance of Offloading Algorithms · p. 9
  • Greedy and Relaxation have low execution time; the authors propose edge-server-trained Sync-FQL for delay-tolerant applications and Relaxation for delay-sensitive ones.The authors state that delay-sensitive applications occupy a small share of wireless bandwidth, so the overall offloading efficiency would not decline significantly.Source: A. Performance of Offloading Algorithms · p. 10
  • Sync-FQL has the lowest P2 total cost, while Greedy and Relaxation cost much more than the learning-based algorithms.Source: Figure 4 · p. 10, A. Performance of Offloading Algorithms · p. 10
  • Under light traffic CV2X-RmmW has the lowest total cost, but under heavy traffic CV2X-DSRC performs slightly better and DSRC-CmmW has the largest cost.CmmW competes with DSRC for bandwidth, and RmmW requires RTS-like and CTS-like transmissions whereas DSRC uses one data transmission.Source: Figure 5 · p. 10, A. Performance of Offloading Algorithms · p. 10

B. Performance of Offloading Technologies

Failure probabilities and resource costs vary across offloading technologies and algorithms. Relaxation has the lowest failure probability, while learning-based algorithms have lower resource cost than Greedy and Relaxation.

B. Performance of Offloading Technologies

C. Performance with Traffic Loads

Offloading performance depends on traffic arrival rate, burstiness, and delay tolerance. The reported results favor offloading high-arrival workloads, C-V2I for highly bursty tasks, and C-V2V for delay-tolerant tasks.

C. Performance with Traffic Loads

VII. CONCLUSION

The conclusion presents stochastic-network-calculus delay bounds and an intelligent offloading scheme aimed at reliability and lower resource cost. It reports different preferred technology combinations by traffic load and task characteristics.

VII. CONCLUSION

  • C-V2X combined with RmmW performs best under light traffic, while C-V2X combined with DSRC performs best under heavy traffic.Source: §VII · p. 11
  • The paper derives offloading-delay upper bounds using stochastic network calculus as indicators for automotive-task quality of service.Source: §VII · p. 11
  • For high-arrival-rate traffic, vehicles should offload to others; for bursty tasks, they should avoid CmmW, while delay-tolerant tasks favor C-V2V.Source: §VII · p. 11
Loading 2006.15855v1…