PaperlayerView exact revision on arXiv

Source-linked AI summary

Resource Scheduling in Edge Computing: A Survey

Quyuan Luo, Shihong Hu, Changle Li, Guanghui Li, Weisong Shi

arXiv:2108.08059v1cs.NIeess.SP

TL;DR

IoT and wireless-network growth creates demanding requirements for edge services, motivating resource scheduling near users. This survey organizes edge-computing architecture, scheduling models, techniques, applications, and open challenges, finding trade-offs between centralized optimality and distributed flexibility.

  • Problem

    IoT and emerging services impose higher requirements for latency, energy efficiency, reliability, and service quality, motivating resource scheduling in edge computing.

  • Method

    The survey presents a three-tier edge-computing architecture, a unified offloading model, and a synthesis of scheduling issues, techniques, collaborative modes, applications, and challenges.

  • Results

    The surveyed techniques show trade-offs: centralized methods can obtain better optimal solutions with more overhead, while distributed methods are simpler, more flexible, easier to implement, and more adaptive.

  • Takeaways & Limitations

    Resource scheduling research must account for dynamic conditions such as unstable energy harvesting and user mobility, which affect offloading, caching, and service stability.

Abstract

from arXiv · show

With the proliferation of the Internet of Things (IoT) and the wide penetration of wireless networks, the surging demand for data communications and computing calls for the emerging edge computing paradigm. By moving the services and functions located in the cloud to the proximity of users, edge computing can provide powerful communication, storage, networking, and communication capacity. The resource scheduling in edge computing, which is the key to the success of edge computing systems, has attracted increasing research interests. In this paper, we survey the state-of-the-art research findings to know the research progress in this field. Specifically, we present the architecture of edge computing, under which different collaborative manners for resource scheduling are discussed. Particularly, we introduce a unified model before summarizing the current works on resource scheduling from three research issues, including computation offloading, resource allocation, and resource provisioning. Based on two modes of operation, i.e., centralized and distributed modes, different techniques for resource scheduling are discussed and compared. Also, we summarize the main performance indicators based on the surveyed literature. To shed light on the significance of resource scheduling in real-world scenarios, we discuss several typical application scenarios involved in the research of resource scheduling in edge computing. Finally, we highlight some open research challenges yet to be addressed and outline several open issues as the future research direction.

I. INTRODUCTION

IoT applications and emerging 5G scenarios demand lower latency, higher energy efficiency, reliability, and greater resource capacity than centralized cloud processing can consistently provide. This motivates edge computing and resource scheduling across heterogeneous devices, services, resources, objectives, and collaborative participants.

  • From Cloud Computing to Edge Computing: IoT applications raise requirements for bandwidth, latency, energy consumption, performance, and reliability that centralized cloud processing may struggle to meet.The surveyed applications include UAVs, connected autonomous vehicles, video services, smart cities, healthcare, manufacturing, and smart homes.
  • From Cloud Computing to Edge Computing: Edge computing moves computing, storage, caching, processing, and service delivery closer to users while retaining interaction with the cloud.Things can produce and consume data, request cloud services, or perform cloud computing tasks at the edge.
  • Resource Scheduling in Edge Computing: Resource scheduling assigns available communication, storage, and computing resources to tasks according to participant objectives.Tasks may originate from applications such as vehicle sensing, health monitoring, and surveillance, with objectives including low latency or reduced energy consumption.
  • Resource Scheduling in Edge Computing: Its main actions are computation offloading, resource allocation, and resource provisioning, covering processing location, resource quantities, and user-resource associations.These actions address how tasks are processed and how edge resources are assigned or provisioned.
  • Resource Scheduling in Edge Computing: Centralized methods collect global information through a control center, whereas distributed methods operate without collecting global information centrally.The two methodologies provide the organizing distinction for resource-scheduling techniques discussed in the survey.

C. Related Surveys

Existing surveys often addressed edge-computing resource scheduling from a single angle or at a high level. This survey responds with a comprehensive, systematic treatment spanning architecture, models, research issues, techniques, indicators, applications, and open challenges.

  • Related surveys: Existing surveys often focused on a single resource-scheduling angle or discussed edge computing only at a high level.Examples include computation offloading and resource provisioning, while some treatments lacked comprehensive depth.
  • Survey scope: The survey presents a comprehensive review of state-of-the-art research focused specifically on resource scheduling in edge computing.
  • Architecture: It introduces a three-tier architecture and examines things-edge, things-edge-cloud, edge-edge, and edge-cloud collaboration patterns.
  • Basic model and research issues: It organizes resource-scheduling research around computation offloading, resource allocation, and resource provisioning using basic models for QoS and QoE requirements.
  • Techniques, applications, and open issues: It compares centralized and distributed scheduling techniques, summarizes performance indicators, surveys applications, and identifies future challenges.Indicators include latency, energy consumption, cost, utility, profit, and resource utilization; applications include UAVs, CAVs, video service, smart cities, smart health, smart manufacturing, and smart homes.

II. ARCHITECTURE

The survey presents a three-tier heterogeneous architecture comprising thing, edge, and cloud layers, and uses it to organize resource-scheduling collaboration. Things generate data that may be processed locally or offloaded according to service requirements.

  • Collaboration: The survey organizes scheduling through collaboration manners under the three-tier architecture, including interactions among things, edge, and cloud participants.The architecture section introduces these collaborative manners after describing the network composition.
  • Architecture: The architecture has three tiers: thing, edge, and cloud layers, reflecting complementary relationships among users, edge resources, and cloud resources.Cloud offers greater capacity, while edge computing can reduce response latency and energy consumption.
  • Thing Layer: The thing layer contains heterogeneous end-devices, including UAVs, connected vehicles, AR equipment, cameras, sensors, and smart devices.These devices perceive data and possess some storage and computing capability.
  • Thing Layer: Thing-generated data can be processed locally or offloaded to the edge or cloud according to QoE and QoS requirements.Static and dynamic end-devices make resource management more challenging.

2) Edge Layer:

The edge layer is the intermediate core between things and the cloud, supplying wireless access and stronger computing and storage than end-devices. Collaboration across layers supports scalable processing, while edge participation can reduce latency and satisfy QoE requirements.

  • Edge Layer: The edge layer forms the intermediate core between the thing and cloud layers and includes cellular towers, edge servers, roadside units, gateways, and controllers.It provides wireless access through LTE, WiFi, and DSRC.
  • Things-Edge Collaboration: Things-edge collaboration decides whether smart-device tasks are processed locally or offloaded to edge servers according to collaboration strategy and QoS/QoE requirements.One cited approach selects computation components to offload while minimizing mobile-device energy consumption.
  • Things-Edge-Cloud Collaboration: Things-edge-cloud collaboration combines edge and cloud resources to address the limited capacity of edge resources as smart devices and resource-intensive applications increase.Examples target offloading performance, scalability, energy consumption, and task-processing time.
  • Edge-Edge Collaboration: Edge-edge collaboration adds another task-processing option and commonly accompanies things-edge or things-edge-cloud collaboration.Studies include parked-vehicle resources, overload migration, and scheduling among mobile devices, edge servers, and the cloud.

E. Edge-Cloud Collaboration

The survey models resource scheduling as decisions over task processing and resource state, then reviews research across offloading, allocation, and provisioning. Its task model captures workload, parallelism, output size, and delay constraints, while the offloading variable permits local, partial, or full execution.

  • Resource Scheduling Model: The survey introduces a unified resource-scheduling model and reviews computation offloading, resource allocation, and resource provisioning.The model guides offloading decisions using communication and computing states together with QoE requirements.
  • Task Model: A task T is represented by data size, processing density, parallelizable fraction, result-to-input size ratio, and delay constraint.These five items characterize the workload and its processing requirements.
  • Offloading Decision: The offloading variable λ∈[0,1] denotes the fraction of task data sent away from the device for processing.λ=0 means local processing, λ=1 means full offloading, and intermediate values mean partial offloading.
  • Local Computing: Local energy consumption depends on local processing time and processor power, with per-core power modeled as proportional to the cube of CPU frequency.The local-power relation uses coefficient κ1 to connect processing capability and device-side power consumption.
  • Local Computing: Local computing time is derived using Amdahl’s law by separating serialized and parallelizable portions of the locally processed data.The serialized component is explicitly parameterized by processing density, nonparallelizable fraction, local data size, and CPU frequency.
  • Wireless Offloading: Wireless offloading uses a transmission rate r characterized by wireless models such as Shannon’s formula and path-loss-dependent links.The rate depends on assigned bandwidth and transmission conditions.

2) Task T offloaded to the edge:

Offloading task T to the edge adds wireless transmission and edge-computing stages, with delay and energy modeled from their respective communication and processing components.

  • Task T offloaded to the edge:: Offloading λD bits requires modeling transmission delay and end-device transmission energy.The transmission rate depends on P1, channel fading h, and noise power ω0.
  • Task T offloaded to the edge:: The edge computes offloaded data using assigned cores and per-core processing capability.The model distinguishes serialized and parallelizable computing parts.
  • Task T offloaded to the edge:: Edge-side energy consumption is modeled from the power used to compute the offloaded data.Per-core power is expressed as p_e = κ2(f_e)^3.

3) Result return:

After edge processing, the result is transmitted back to the end-device, adding return delay and edge transmission energy to task-processing costs.

  • Result return:: Returning γD bits produces a transmission delay determined by the result-return data rate.The return-delay model parallels the offloading-delay model.
  • Result return:: The edge incurs transmission energy when sending the γD-bit processing result to the end-device.This energy is modeled separately from the earlier offloading and edge-computing energy terms.
  • Result return:: Total task cost combines energy consumption, bandwidth use, and computing-resource use.Bandwidth and computing costs are parameterized by per-unit prices p1 and p2.

6) Computing acceleration:

Computing acceleration compares edge-assisted processing with local computation while accounting for both computation and transmission delays.

  • Computing acceleration:: Local and edge processing speedups are formulated with Amdahl’s law for their respective task portions.The formulas distinguish the locally computed (1−λ)D bits from the edge-computed λD bits.
  • Computing acceleration:: Actual edge-assisted acceleration includes computing delay and transmission delay rather than computation time alone.The relevant latency includes offloading, edge computation, and result-return delays.
  • Computing acceleration:: Resource scheduling must jointly address energy consumption, latency, cost, and computing acceleration.The survey maps these requirements to offloading, resource allocation, and resource provisioning decisions.

B.1. Direction

Computation offloading is classified by direction and granularity, and is commonly studied jointly with allocation and provisioning across heterogeneous edge systems.

  • B.1. Direction: Device-to-edge offloading compensates for resource-constrained devices while supporting QoS and QoE requirements.It can also alleviate cloud-center overload and reduce wireless-transmission delay.
  • B.1. Direction: Offloading directions include device-to-edge, edge-to-cloud, cloud-to-edge, device-to-device, and edge-to-edge.The survey groups upward and downward directions as vertical offloading and peer transfers as horizontal offloading.
  • B.1. Direction: Edge-to-cloud offloading can balance overload when edge computing nodes cannot process tasks in time.The surveyed edge nodes include cloudlets, ENs, BSs, and mini data centers.
  • B.1. Direction: Cloud-to-edge offloading brings tasks closer to users to reduce transmission latency and shorten application response time.Examples include video transcoding, application cloning, and data replication at the edge.
  • B.1. Direction: Edge-to-edge offloading distributes workloads through task scheduling, service migration, and offload forwarding among peer ENs.These mechanisms address overloaded edge nodes and heterogeneous edge environments.
  • B.1. Direction: Device-to-device offloading uses idle resources of peer devices and may operate within things-edge or things-edge-cloud collaboration.Vehicle edge computing is given as an example application context.
  • B.1. Direction: Offloading granularity is binary or partial, depending on whether a task is indivisible or partitionable.Partial offloading uses an offloading ratio to determine the proportion processed at the destination.
  • B.1. Direction: Existing studies often combine offloading decisions with resource allocation and resource provisioning.Allocation covers computing, communication, and storage resources.

2) Computing and communication (CC):

Resource scheduling research jointly considers communication, computing, and storage resources across allocation, provisioning, placement, and task scheduling problems. These problems are often non-convex or NP-hard, motivating convex optimization and related solution techniques.

  • Resource allocation: Communication, computing, and storage resources are commonly modeled and optimized jointly rather than allocated individually.
  • Resource provisioning: Dynamic workload fluctuations create over-provisioning and under-provisioning risks, respectively increasing system cost or degrading QoS and task completion.
  • Resource provisioning: Resource provisioning research covers task allocation, cloudlet and edge-node placement, application placement, service placement, and virtualization-based resource management.
  • Solution techniques: The surveyed optimization models are typically non-convex or NP-hard, so many studies transform them into near-convex or convex problems.
  • Solution techniques: Convex-optimization techniques include Lyapunov, decomposition, SCA, and ADMM methods, which can yield sub-optimal solutions but often require complex calculations.

1) Convex optimization:

The survey organizes scheduling techniques around approximate, heuristic, machine-learning, game-theoretic, matching, and auction-based approaches for difficult distributed or optimization problems. Each approach offers practical benefits but also has stated limitations.

  • Approximate algorithm: Approximate algorithms apply relaxation, bounding, local search, or dynamic programming to established NP-hard problems, offering simple and flexible implementation.
  • Heuristic algorithm: Heuristic scheduling research commonly uses greedy and genetic algorithms, while simple heuristics risk local optima and meta-heuristics require difficult-to-reuse parameter settings.
  • Machine learning: Machine learning is applied because traditional static methods cannot make optimal decisions in dynamic environments modeled as MDPs.
  • Game theory: Game-theoretic distributed methods model users as players who choose responses to pursue their interests, generally requiring proof of a Nash Equilibrium.
  • Matching theory: Matching theory provides decentralized, low-complexity scheduling and is effective in dynamic networks, but is generally unsuitable for partial offloading because it targets binary offloading.
  • Auction: Auction mechanisms represent task-processing entities as bidders, service providers as sellers, and a trusted entity as auctioneer.

4) Federated learning:

Federated learning extends distributed scheduling by training models across edge devices without exchanging local data samples, while blockchain supports decentralized and traceable resource transactions. The survey contrasts their benefits with latency, device-coordination, and security limitations.

  • Federated learning: Federated learning trains resource-scheduling algorithms across distributed edge devices or servers without exchanging local data samples.
  • Federated learning: Federated learning can protect privacy, reduce wireless transmission burden, and reduce individual training time by exchanging model parameters instead of local data.
  • Federated learning: Federated learning involves multiple devices and remains vulnerable to malicious attacks.
  • Blockchain: Blockchain-based scheduling can maintain resource-transaction data security and integrity through decentralization, non-tampering, irreversibility, and traceability.
  • Blockchain: Blockchain-based methods have relatively high latency and involve multiple devices, while federated learning is described as a newer direction requiring further work.
  • Centralized and distributed methods: Centralized methods can obtain better optimal solutions but incur more overhead, whereas distributed methods are simpler, flexible, easier to implement, and adaptive to dynamic environments.

V. RESOURCE SCHEDULING IN APPLICATIONS CONTEXT

Resource scheduling is studied across UAVs, connected autonomous vehicles, video services, smart cities, and related IoT applications. These scenarios use edge resources to address limited device capacity, large data volumes, latency demands, and heterogeneous services.

  • UAVs: UAV scheduling research offloads computing-intensive, latency-sensitive tasks to ground base stations or uses UAVs as mobile edge resources.
  • Connected autonomous vehicles: Connected autonomous vehicles generate up to 4TB of raw data daily, motivating offloading to roadside or other edge resources for low-latency applications.
  • Connected autonomous vehicles: Parked vehicles can supplement edge infrastructure by providing idle computing resources for partitioned mobile-user tasks and collaborative service.
  • Video services: Video processing is scheduled to the edge because smart devices have limited storage and computing capabilities while video applications are computation-intensive and bandwidth-hungry.
  • Video services: VideoEdge optimizes computer-vision component placement, while LAVEA performs task selection and prioritized offloading using a serverless architecture.
  • Smart cities: Smart-city studies address diverse, heterogeneous, privacy-sensitive data through edge-collaborative processing and resource or service placement.

E. Smart Health

The survey presents smart health as an application area where growing medical data and operating costs motivate processing medical data at the edge. It also identifies several unresolved scheduling challenges, including task modeling, migration, energy harvesting, and heterogeneous architectures.

  • E. Smart Health: Edge resources are deployed to process medical data because medical-data growth increases medical-system operating and maintenance costs.
  • E. Smart Health: Existing task models commonly estimate computing capacity as task data size multiplied by a CPU-cycles-per-bit processing density.The survey notes that this is a linear, one-size-fits-all representation that may not suit different application tasks.
  • E. Smart Health: Computation migration involves environment sensing, task division, migration decisions, uploading, execution, and result return.Task division and migration decisions are identified as the two most critical steps.
  • E. Smart Health: Most energy-harvesting scheduling studies assume extra energy can be harvested continuously, although practical harvesting may be unstable.The survey calls for greater attention to the energy-harvesting process when designing scheduling strategies.
  • E. Smart Health: Existing scheduling research is concentrated mainly on the thing-edge-cloud architecture, while space-air-ground integrated networks are identified as a future heterogeneous architecture.

B. Feasibility

The survey identifies feasibility barriers beyond algorithm design, including limited deployment studies, weak treatment of security and privacy, user mobility, queueing, workload realism, and scarce real-system evaluation. It also organizes a review of architectures, scheduling issues, techniques, indicators, applications, and open challenges.

  • B. Feasibility: Few studies address deployment of edge nodes, despite geographical location affecting resource scheduling and service-range utilization.User mobility further complicates edge-node deployment.
  • B. Feasibility: Infrastructure scheduling remains mostly theoretical, while implementation, scalability, flexible deployment, and rapid configuration require further attention.
  • B. Feasibility: Security and privacy are insufficiently explored, although attacks or edge-node failures can threaten the reliability and robustness of the edge system.The survey specifically identifies authentication and privacy protection as needed for edge services and data.
  • B. Feasibility: User mobility can invalidate current offloading and caching decisions when users move between or outside edge-node service ranges.The survey suggests trajectory prediction and mobility-management policies to support QoS and seamless access.
  • B. Feasibility: Most evaluations use simulators or Matlab, and few studies assess scheduling algorithms in real edge systems.The survey calls for testbeds or prototypes to evaluate practical feasibility.
  • B. Feasibility: The survey systematically reviews architectures, collaborative manners, three scheduling issues, centralized and distributed techniques, six performance indicators, application scenarios, and open challenges.
Loading 2108.08059v1…