PaperlayerView exact revision on arXiv

Source-linked AI summary

Resource Management in Fog/Edge Computing: A Survey

Cheol-Ho Hong, Blesson Varghese

arXiv:1810.00305v1cs.DC

TL;DR

Fog/edge computing brings processing closer to devices, but its constrained, heterogeneous, and dynamic resources create a resource-management challenge. This survey reviews literature through 2018 and classifies architectures, infrastructure, and algorithms. More than 85% of reviewed articles were published from 2013, while the survey identifies lightweight-versus-performance needs and several open or impractical areas.

  • Problem

    Fog/edge resources are resource-constrained, heterogeneous, and dynamic, making resource management an important challenge compared with centralized cloud resources.

  • Method

    The survey reviews research on fog/edge resource management and classifies contributions across architectures, infrastructure, and algorithms.

  • Results

    More than 85% of reviewed articles were published from 2013, and 82 publications supported the survey’s architecture classification.

  • Takeaways & Limitations

    The survey identifies key data-flow architectures and highlights resource-management techniques spanning aggregation, sharing, and offloading.

  • Takeaways & Limitations

    Resource-limited edge devices cannot run heavyweight tools such as Apache Spark and deep learning libraries, while lightweight tools lack advanced analytics functions.

Abstract

from arXiv · show

Contrary to using distant and centralized cloud data center resources, employing decentralized resources at the edge of a network for processing data closer to user devices, such as smartphones and tablets, is an upcoming computing paradigm, referred to as fog/edge computing. Fog/edge resources are typically resource-constrained, heterogeneous, and dynamic compared to the cloud, thereby making resource management an important challenge that needs to be addressed. This article reviews publications as early as 1991, with 85% of the publications between 2013-2018, to identify and classify the architectures, infrastructure, and underlying algorithms for managing resources in fog/edge computing.

I. INTRODUCTION

Fog/edge computing moves processing resources closer to end devices and sensors to address cloud-induced latency, but its constrained, heterogeneous, and dynamic resources make management challenging. The survey classifies resource-management architectures, infrastructure, and algorithms while reviewing publications through 2018.

  • Motivation: Cloud-centered processing can increase communication latency as billions of connected devices transfer data to distant data centers.The resulting latency can degrade application Quality-of-Service and Quality-of-Experience.
  • Motivation: Fog/edge computing brings computing resources closer to users and sensors, reducing data sent to the cloud and consequently reducing communication latencies.The model decentralizes some data-center resources toward the network edge.
  • Motivation: Edge resources are resource-constrained, heterogeneous, and dynamic, making resource management a key fog/edge computing challenge.Their processors, power budgets, architectures, workloads, and competing applications vary across devices.
  • Scope and classification: The survey reviews architectures, infrastructure, and algorithms underpinning resource management in fog/edge computing.Its architecture coverage includes dataflow, control, and tenancy, while later sections address infrastructure and algorithms.
  • Survey scope: More than 85% of the reviewed articles were published from 2013, covering research published between 1991 and 2018.The review categorizes publications as books or chapters, reports, conference or workshop papers, and journal or magazine articles.
  • Architecture classification: The survey classifies data-flow architectures by workload movement, control architectures by resource control, and tenancy architectures by hosted entities.Examples include device-to-edge or cloud-to-edge transfer, centralized or distributed control, and single- or multi-application hosting.
  • Architecture classification: The architecture classification uses 82 research publications, with 86% published since 2013.The reviewed architecture literature is summarized in a histogram.
  • Data-flow architectures: Data-flow architectures include aggregation, sharing, and offloading.These categories organize how data or workloads are transferred within fog/edge environments.

1) Aggregation:

The survey organizes aggregation research by how aggregation is modeled and implemented and by techniques that improve its efficiency or quality. These techniques address bandwidth, latency, energy, quality, heterogeneity, and accuracy across several implementation families.

  • Aggregation overview: Aggregation partially computes data from multiple end devices at an edge node to reduce communication overhead and unnecessary traffic beyond the edge.The survey separates aggregation research into modeling or implementation techniques and techniques for improving aggregation.
  • Techniques for modeling and implementing aggregation: Modeling and implementation techniques include graph-based, cluster-based, Petri Net-based, and decoupled approaches.The survey discusses tree and directed-graph methods, cluster heads, High Level Petri Nets, and routing-aggregation decoupling.
  • Graph-based techniques: Tree-based aggregation includes Data Aggregation Trees and spatial index trees, with probabilistic network models used to represent lossy links for load balancing.EGF is an energy-efficient index tree for data collection and aggregation when sensors are unevenly distributed.
  • Graph-based techniques: Distributed Dataflow uses directed graphs, MQTT, multi-node flow deployment, and heterogeneous-device support for fog applications.It extends dataflow programming to fog computing and wireless sensor network applications.
  • Modeling and implementation techniques: Batch techniques group incoming data over time intervals for batch jobs, while hybrid techniques combine methods across network regions.Examples include Apache Flink, Spark D-Streams, and Tributary-Delta.
  • Techniques for improving aggregation: Improvement techniques optimize communication efficiency through bandwidth, latency, and energy objectives, or optimize aggregation and analytics quality.The survey presents efficiency-aware and quality-aware categories.
  • Efficiency-aware techniques: Latency-aware aggregation can service requests in near real-time with minimum latency, but the described approach trades off energy efficiency.Filtering policies are based on complex event processing.
  • Efficiency-aware techniques: Energy-aware research minimizes transfer power consumption, but edge-side data-processing overhead remains, leaving energy awareness an open research area.The literature explores trade-offs between aggregation latency and energy consumption.

2) Sharing:

The survey classifies workload sharing among edge peers by control, optimization objectives, and cooperation setting. It also distinguishes single-hop, multi-hop, and opportunistic sharing, while highlighting energy, heterogeneity, security, and fairness concerns.

  • Sharing: Sharing distributes workloads among peer devices instead of offloading them to the cloud, creating dynamic networks that require energy-aware cooperation.Devices may join or leave without notice, and participating devices are likely to be battery-powered.
  • Classification: Sharing techniques are classified by control, optimization, and cooperation, covering centralized or distributed control, objective-specific methods, and ad hoc or infrastructure-based settings.The survey presents these categories as the main dimensions of its sharing-technique classification.
  • Optimization: Connectivity-aware sharing includes single-hop, multi-hop, and opportunistic techniques for selecting peers as devices move into or out of communication range.Multi-hop methods access a larger resource pool, while opportunistic methods use contextual or historical connectivity information.
  • Optimization: Heterogeneity-aware methods address differences in processor architectures, operating systems, and workload deployment across mobile edge devices.Honeybee uses cycles from heterogeneous mobile devices to serve a workload from a given device.
  • Optimization: HoneyBot identified a malicious device in 20 minutes and isolated it from the network to maintain network safety.The technique selects some devices as HoneyBots to monitor malicious behavior.
  • Optimization: Fairness is formulated as multi-objective optimization balancing mobile-device battery drain against performance gains from workload sharing.The referenced approach models application processing as a DAG of embarrassingly parallel tasks and solves each task using a heuristic technique.

3) Offloading:

The survey organizes offloading by data-flow direction and by how applications or cached data are moved across devices, edge nodes, and clouds. It covers partitioning, caching, server migration, and several planning heuristics, while noting resource constraints at the edge.

  • Offloading: Offloading moves a server, application, and associated data to the edge to augment user-device computing or bring cloud services closer to request sources.The survey distinguishes offloading from user devices to the edge and from the cloud to the edge.
  • Offloading from User Device to Edge: Device-to-edge offloading uses application partitioning and caching mechanisms to augment computing on user devices.Examples include Cloudlet VMs processing videos from multiple mobile devices and edge caching of shared content.
  • Planning Approaches: 72 exhaustive brute-force offloading plans are impractical for deriving plans quickly, although they can provide insight into the search space.The plans cover combinations involving the cloud, edge nodes, and user devices, with the minimum-execution-time plan selected.
  • Application Partitioning: Application partitioning is modeled through graph-based, component-based, or neural-network-based approaches.These approaches partition application tasks, browser components, or neural-network layers across devices, edge nodes, and clouds.
  • Caching: Caching mechanisms include chunking and aggregation, as well as reverse-auction games that incentivize cooperative caching among edge nodes.Multi-RAT uploads file chunks to edge VMs in parallel before an aggregation VM combines them for cloud transfer.
  • Offloading from Cloud to Edge: Cloud-to-edge offloading includes server offloading, caching mechanisms, and web programming, with server migration implemented through replication or partitioning.Partitioning may be functionality-aware, geography-aware, or latency-aware.

B. Control

The survey classifies resource-management control architectures as centralized or distributed according to where decisions are made. Centralized control uses one controller, whereas distributed control spreads decision-making across edge nodes.

  • B. Control: Control architectures are classified as centralized or distributed based on how resource decisions are made.The classification covers decisions about computations, networks, and communication of edge resources.
  • B. Control: Centralized control uses a single controller to make decisions for edge-resource computations, networks, or communication.Distributed control instead places decision-making across edge nodes.

1) Centralized:

The survey classifies fog/edge resource-management architectures by centralized and distributed control, and by tenancy across applications and users.

  • 1) Centralized:: Centralized architectures include solver-based and graph matching-based approaches for resource-management decisions.Solver-based methods generate assignment plans, while graph matching maps offloading decisions to a minimum weight-matching problem.
  • 2) Distributed:: Distributed architectures are classified as blockchain-based, game theoretic-based, genetic algorithm-based, and sensor function virtualization-based.These approaches distribute decision-making across edge nodes or modular functions.
  • C. Tenancy: Tenancy is classified across two dimensions—applications and users—yielding four combinations: SASU, SAMU, MASU, and MAMU.The taxonomy ranges from one application and one user to multiple applications and multiple users.
  • C. Tenancy: MAMU represents multiple applications serving multiple users, while SASU provides exclusive hardware use for one application and one user.The survey describes MAMU as representative of publicly accessible infrastructure.
  • C. Tenancy: System virtualization and network slicing are the two techniques identified for supporting multi-tenancy.Virtualization isolates resources, while network slicing runs multiple logical networks over shared physical infrastructure.

III. INFRASTRUCTURE

The survey classifies fog/edge resource-management infrastructure into hardware, system software, and middleware, reviewing 63 publications in this section.

  • III. INFRASTRUCTURE: Infrastructure comprises hardware and software facilities for managing computation, network, and storage resources for fog/edge applications.The survey uses three categories: hardware, system software, and middleware.
  • Hardware: Hardware includes small-form-factor devices such as gateways, WiFi APs, set-top boxes, cars, drones, and commodity computers.Single-board computers are increasingly used to provide computing capabilities on these devices.
  • System software: System software runs directly on fog/edge hardware and manages device computation, network, and storage resources.Examples include operating systems and virtualization software.
  • Middleware: Middleware runs on an operating system to provide complementary services, coordinate distributed nodes, and deploy VMs or containers.Its services include coordination and orchestration, performance monitoring, and communication facilities.
  • III. INFRASTRUCTURE: 83% of the 63 publications reviewed for infrastructure were published since 2013.Figure 10 presents the publication histogram across books or chapters, reports, conference or workshop papers, and journal or magazine articles.

A. Hardware

Fog/edge hardware spans computation and network devices, while system software provides virtualization and isolation for resource-constrained edge nodes.

  • A. Hardware: Fog/edge computation devices include single-board computers and commodity products used to process fog/edge data.Examples include Raspberry Pi boards, desktops, laptops, and smartphones.
  • A. Hardware: Single-board computers integrate CPU, memory, networking, and storage on one circuit board and are commonly used as fog/edge nodes.Examples include Raspberry Pi boards used in connected vehicles, drones, and IoT gateways.
  • A. Hardware: Commodity products such as desktops, laptops, and smartphones can provide idle computational resources as fog/edge nodes.One study used such devices in classrooms, movie theaters, and cafes.
  • A. Hardware: Network devices include gateways, routers, WiFi APs, and edge racks located at the network edge to process network traffic.Examples include smart gateways, ubiquitous wireless gateways, and GENI racks combining network, computing, and storage resources.
  • System software: System virtualization supports multi-tenancy by isolating tenant resources and preventing one tenant from monopolizing available resources.VMs virtualize physical resources through a hypervisor, whereas containers provide lightweight process-level virtualization.
  • System software: Containers offer low overhead, rapid startup and shutdown, and high application density on performance-limited fog/edge hardware.They package applications and dependent libraries for deployment across devices.

2) Network Virtualization:

Network virtualization combines software and hardware mechanisms to connect fog/edge resources, support distributed control, and adapt capacity to demand.

  • 2) Network Virtualization:: Network virtualization combines hardware and software network resources into a software-based virtual network for a tenant.It creates a logical network distinct from the underlying physical network.
  • 2) Network Virtualization:: SDN separates control and data planes, while NFV decouples networking functions such as routing and firewalls from dedicated hardware.Together they provide software-based network management for fog/edge clouds.
  • 2) Network Virtualization:: SDN/NFV can interconnect geographically dispersed fog/edge clouds through virtual networks supporting L2, L3, IPv4, IPv6, and varied addressing modes.The HFC framework supports automated provision of virtual networks to distributed fog/edge clouds.
  • 2) Network Virtualization:: Virtualized networks can place VMs for the same tenant in one VLAN across different locations and route data between VMs using virtual switches.SDN routing strategies can allocate more VMs to higher-priority video-analysis tasks.
  • 2) Network Virtualization:: NFV-enabled mobile fog/edge clouds support dynamic resource allocation and infrastructure scaling based on demand.The survey describes algorithmic resource allocation for such virtualized environments.
  • 2) Network Virtualization:: A hybrid SDN control plane adds runtime backup controllers because a centralized controller can be a single point of failure and reduce reliability.The hybrid design retains a centralized controller while attaching additional controllers as backups.

C. Middleware

Middleware supplies complementary services for fog/edge resource management and supports hierarchical processing across edge and cloud tiers. The surveyed platforms provide coordination, monitoring, communication, and deployment capabilities.

  • Middleware provides performance monitoring, coordination, orchestration, communication facilities, and protocols for fog/edge computing.
  • Nebula enables a decentralized edge cloud in which volunteer nodes contribute resources for applications and data processing.Its ComputePool coordinates with DataStore to place computation near input data.
  • Hierarchical platforms process prompt-reaction tasks at fog/edge nodes while assigning complex or long-term analysis to cloud nodes.
  • Mobile Fog maps application processes across smartphone or vehicle leaf nodes, intermediate fog nodes, and data-center root nodes.It also supports communication between hierarchy levels and runtime scaling.
  • A smart-city platform uses sensor, low-power fog/edge, and more powerful upper layers to provide timely analysis near data sources.
  • A unified edge-cloud platform performs simple healthcare analytics at the edge and comprehensive long-term analytics in the cloud.

3) Mobile Fog/Edge Computing:

Mobile fog/edge computing extends mobile cloud computing by using nearby devices and hierarchical infrastructure to address latency, bandwidth, mobility, and resource constraints. The surveyed systems use diverse coordination, virtualization, and orchestration strategies.

  • Cloud-only offloading cannot provide low network latency and high bandwidth, motivating task offloading to nearby fog/edge resources.
  • FemtoClouds form compute clusters from powerful nearby smartphones and laptops, scheduling requests on idle devices with sufficient capability.
  • SOUL virtualizes device-specific sensors through APIs and externalizes them to an edge-cloud for shared computational and storage services.
  • WiFi-Direct groups let nearby mobile devices cooperate as computing-resource pools for data caching and video streaming.
  • Human-driven edge computing combines mobile edge computing with mobile crowdsensing, using smartphones and tablets as cooperative edge nodes without a controller.
  • Fog/edge devices have limited capabilities, may be numerous, and can move with intermittent network connectivity, complicating coordination.
  • Foggy orchestrates containerized application components across cloud, cloudlet, gateway, and sensor-adjacent Swarm of Things tiers.
  • IoT-cloud middleware provisions software-defined gateways by selecting compatible images, deploying them, and configuring their virtual environments.

IV. ALGORITHMS

The survey organizes fog/edge resource-management algorithms around discovery, benchmarking, load balancing, and placement. It describes their purposes, implementation techniques, and practical challenges in dynamic, constrained environments.

  • The surveyed algorithms cover discovery, benchmarking, load balancing, and placement of workloads across fog/edge resources.
  • Discovery identifies usable edge resources through programming infrastructure, handshaking protocols, or message passing.
  • Discovery: EaaS provides lightweight discovery for homogeneous edge resources, but its demonstrated deployment involved only a single collection of edge nodes.Further research is required for federated edge environments.
  • Benchmarking: Benchmarking captures resource performance across entities such as memory, CPU, storage, and network for deployment decisions.
  • Benchmarking: Near-real-time benchmarking is difficult because dynamic conditions change, while traditional cloud techniques can require hours and incur substantial costs.
  • Benchmarking: Edge benchmarking lacks diverse edge-specific application benchmarks, and resource-constrained nodes require lightweight tools.
  • Benchmarking: Integrated benchmarking is needed to evaluate cloud-edge deployment combinations for overall application performance.
  • Load balancing: Edge load balancing uses particle swarm optimization, cooperative models, graph-based balancing, and breadth-first search.

D. Placement

Placement algorithms assign computation tasks to suitable fog/edge resources while accounting for resource availability, changing conditions, and service constraints. The survey concludes that many resource-management challenges are addressed, but lightweight processing and containerized GPU management remain open issues.

  • Placement: Placement must account for fog/edge resource availability and environmental changes, including user mobility and changing resource availability.
  • Placement: A dynamic condition-aware method predicts placement parameters such as user location, preference, database location, and system load to minimize expected cost.
  • Placement: Hierarchical placement can iterate from fog toward cloud, placing computation modules on available fog nodes using CPU, memory, and network bandwidth attributes.
  • Placement: The Fog Service Placement Problem optimally shares fog-node resources among IoT services while considering latency or deadline QoS constraints.
  • Conclusion: Technical challenges in managing limited fog/edge resources have been addressed to a high degree, although further improvements in capability and performance remain.
  • Conclusion: Resource-limited devices cannot run heavyweight tools such as Apache Spark and deep-learning libraries, while lightweight alternatives lack advanced analytics functions.
  • Conclusion: Efficient GPU resource management in containers remains an open research challenge despite widespread container use for lightweight virtualization.
  • Conclusion: Fog/edge resource-management research is classified across architectures, infrastructure, and algorithms to study contributions near mobile and IoT devices.
Loading 1810.00305v1…