PaperlayerView exact revision on arXiv

Source-linked AI summary

Computing Power Network: A Survey

Yukun Sun, Bo Lei, Junlin Liu, Haonan Huang, Xing Zhang, Jing Peng, Wenbo Wang

arXiv:2210.06080v2cs.NI

TL;DR

Distributed computing resources remain isolated, motivating CPN's coordinated scheduling. This survey synthesizes CPN architecture and research, and its testbed's offloading strategy cuts response delay by about 25% versus RR while maintaining uniform node load similar to RR.

  • Problem

    Ubiquitously distributed computing resources are difficult to coordinate because of the computing-power island effect.

    Source: §I · p. 2

  • Method

    The paper surveys CPN architecture and research on modeling, awareness, allocation, forwarding, and platforms, and evaluates a prototype with four edge computing nodes.

    Source: §8.2 · p. 23, §XII · pp. 30–31

  • Results

    About 25% lower response delay than RR, with uniform node load similar to RR and much better than GR.

    Source: §8.2 · p. 23

  • Takeaways & Limitations

    CPN is intended to provide lower transmission latency than cloud computing and lower computing latency than MEC, enabling ultra-low-latency task processing.

    Source: §I · p. 2

Abstract

from arXiv · show

With the rapid development of cloud computing, edge computing, and smart devices, computing power resources indicate a trend of ubiquitous deployment. The traditional network architecture cannot efficiently leverage these distributed computing power resources due to computing power island effect. To overcome these problems and improve network efficiency, a new network computing paradigm is proposed, i.e., Computing Power Network (CPN). Computing power network can connect ubiquitous and heterogenous computing power resources through networking to realize computing power scheduling flexibly. In this survey, we make an exhaustive review on the state-of-the-art research efforts on computing power network. We first give an overview of computing power network, including definition, architecture, and advantages. Next, a comprehensive elaboration of issues on computing power modeling, information awareness and announcement, resource allocation, network forwarding, computing power transaction platform and resource orchestration platform is presented. The computing power network testbed is built and evaluated. The applications and use cases in computing power network are discussed. Then, the key enabling technologies for computing power network are introduced. Finally, open challenges and future research directions are presented as well.

I. INTRODUCTION

The introduction motivates CPN as a response to scattered computing resources, weak coordination, and the computing-power island effect. It presents the survey as a broad review of CPN foundations, functions, and research directions.

I. INTRODUCTION

  • CPN coordinates ubiquitous computing resources to address the island effect, low utilization, high latency, and task-load imbalance in scattered, weakly coordinated nodes.The introduction frames CPN as connecting and scheduling resources across distributed nodes rather than leaving them isolated.Source: §I · p. 2

1.1 Related Work

  • Standards work spans IETF CFN documents and ITU-T CPN initiatives, while academic studies address architectures, collaboration, routing, and challenges.Source: §1.1 · pp. 2–4

1.2 Survey Novelty and Contributions

  • The survey claims a comprehensive CPN treatment spanning definition, architecture, advantages, functional models, platforms, enabling technologies, applications, and open directions.Source: §1.2 · p. 4
  • Its coverage surveys computing-power modeling, information awareness and announcement, resource management, and SRv6-based forwarding, and proposes transaction and resource-orchestration platforms.Source: §1.2 · p. 4
  • The survey also builds and evaluates a testbed, summarizes applications and use cases, and identifies enabling technologies including MEC, SDN, NFV, network slicing, Docker, SRv6, and AI.Source: §1.2 · p. 4
  • Its open-research agenda includes computing-power modeling, information synchronization, mobility, energy consumption, and deterministic ultra-low latency.Source: §1.2 · p. 4

1.3 Survey Structure

  • The roadmap moves from CPN concepts and architecture through modeling, information awareness, resource management, forwarding, and platform topics.Source: §1.3 · p. 4
  • Later sections cover a testbed, applications and use cases, enabling technologies, open challenges, and concluding ideas.Source: §1.3 · p. 4

II. OVERVIEW OF COMPUTING POWER NETWORK

Computing power network (CPN) connects distributed computing resources across end, edge, and cloud, and jointly perceives, allocates, and schedules computing and network resources. The overview describes its layered architecture and the resource coordination and service-scheduling advantages it aims to provide.

2.1 What Is Computing Power Network

  • CPN is a network paradigm that perceives user demand and multidimensional resources, then jointly allocates computing and network resources to schedule tasks on demand.It is described as a deep integration of end, edge, cloud, and network, and as an evolution from edge toward ubiquitous computing.Source: §2.1 · p. 5

2.2 Functional Architecture of Computing Power Network

  • The architecture spans ubiquitous end-edge-cloud computing and network infrastructure, pooled computing, storage, and network resources, and dynamically deployed services.The pooling layer models and abstracts heterogeneous resources into computing, network, and service pools.Source: §2.2 · pp. 6–7
  • Resource announcement synchronizes available computing, storage, network, and service information; scheduling uses that information and user demands to allocate resources and optimize latency, utilization, management, and provider revenue.The survey describes centralized, distributed, and hybrid synchronization, and lists machine-learning methods among possible scheduling algorithms.Source: §2.2 · pp. 7–8
  • The service layer provides AI algorithms and platforms, while orchestration and management handle resource and service orchestration, operations, modeling, and security.The task procedure models requests, measures and announces node resources, deploys services, makes an offloading decision, routes the task, and returns its result.Source: §2.2 · p. 8

2.3 Advantages of Computing Power Network

  • CPN coordinates ubiquitous resources to support low-latency task processing, improve resource utilization, maintain consistent user experience, and dynamically deploy services according to network conditions and SLAs.The survey illustrates utilization gains by describing tasks moved from a loaded node to other nodes with no tasks.Source: §2.3 · p. 9

III. COMPUTING POWER MODELING IN COMPUTING POWER NETWORK

Computing power modeling addresses heterogeneous deployment and provides a basis for scheduling, management, and transactions. The section reviews measurement objectives, infrastructure and task-demand models, and computing-power grading.

3.1 Computing Power Measurement Objectives

  • Computing power is measured by hardware capability and service performance, including inference latency, network delay variation, throughput, and frames per second.Logical computing power is typically measured in TOPS, while parallel and neural-network computing power are typically measured in FLOPS.Source: §3.1 · pp. 9–10

3.2 Infrastructure Computing Power Modeling

  • Heterogeneous chips require a uniform mapping into a shared dimension so resource pools can be understood by the service layer.The infrastructure includes CPUs, GPUs, TPUs, NPUs, ARM processors, and FPGAs.Source: §3.2 · p. 10
  • A quantity model sums mapped logical, parallel, and neural-network chip capacities using scaling coefficients and redundant-power terms.The survey also describes black-box measurement using historical pool data and deep-learning algorithms.Source: §3.2 · pp. 10–11

3.3 Task Requirements Computing Power Modeling

  • Task-demand models map requirements to computing power using task type, computation type, data size, and tolerated processing latency.Task latency is difficult to calculate from operation counts alone, and node computing and storage resources vary over time.Source: §3.3 · p. 11
  • Function fitting relates task-processing latency to node performance and task properties, supporting allocation of computing power for customer tasks.Node performance includes computing, storage, and I/O resources; task load and size are also modeled.Source: §3.3 · pp. 11–12

3.4 Computing Power Grading

  • The survey grades computing power into four levels, from over 1 PFLOPS for super-computing applications to under 500 GFLOPS for all-computing applications.The intermediate levels are large computing power, about 10 TFLOPS to 1 PFLOPS, and medium computing power, about 500 GFLOPS to 10 TFLOPS.Source: §3.4 · p. 12

IV. INFORMATION AWARENESS AND ANNOUNCEMENT IN COMPUTING POWER NETWORK

The section surveys user, computing-node, and network-link information needed for on-demand computing-power scheduling, and compares centralized, distributed, and hybrid announcement approaches.

4.1.1 Mobility Awareness

  • Mobility awareness supports offloading decisions and task or result migration, while predicted user movement can inform resource scheduling and load balancing.User requirements also include task-resource needs; task features include data size and type, latency tolerance, and flow stickiness.Source: §4.1.1 · p. 13, §4.1 · p. 13, §4.1.2 · p. 13

4.1.2 Task Requirements Awareness

  • Task requirements span data features and latency or flow stickiness, computing capacity and type, storage capacity and quality, network performance, and algorithm objectives.The surveyed examples include roughly 2 PFLOPS to train COCO once with ResNet50 versus 6 PFLOPS with VGGNet.Source: §4.1.2 · pp. 13–14

4.2 Aware Computing Node Information

  • Computing-node awareness covers location, resource status, interface bandwidth, available services and identifiers, and node load across heterogeneous end, edge, and cloud deployments.The resource information includes computing-chip type and capacity, plus storage capacity, ability, and quality.Source: §4.2 · pp. 14–15

4.3 Aware Network Link Information

  • Real-time link awareness tracks connectivity, transmission rate, delay jitter, packet loss, spectrum, and base-station transmit power to construct a weighted graph for offloading-path selection.Predictive user, node, and link information is also identified as support for proactive allocation and scheduling.Source: §4.3 · p. 15

4.4 Information Announcement in Computing Power Network

  • Centralized announcement gives a controller a global resource view for route optimization and has low deployment impact, but reporting node information to it adds latency.The approach extends the northbound interface without requiring protocol extensions.Source: §4.4 · p. 15, Figure 4 · p. 16
  • Distributed announcement uses IGP and BGP among computing routing nodes, avoiding central interaction for faster, lower-latency convergence but requiring costly, lengthy device extensions.The cited comparison table covers the different announcement approaches, but its values are not supplied here.Source: §4.4 · p. 16, Figure 5 · p. 16, Table 4 · p. 17
  • Hybrid announcement uses distributed IGP/BGP within groups and reports selected nodes’ converged information to a controller for a global view with shorter latency.The survey presents this mode as balancing the cost and response-speed trade-offs of centralized and distributed approaches.Source: §4.4 · p. 16, Figure 6 · p. 16

V. INTELLIGENT RESOURCE MANAGEMENT IN COMPUTING POWER NET-

This section surveys resource management for heterogeneous, dynamic computing-power networks, from node registration and monitoring to offloading, service orchestration, and AI-supported management.

5.1 Computing-networking Resources Management

  • Node registration distributes resource parameters and scheduling strategies to smart routers, while OAM monitors performance, replans routes or node selection, and detects faults for traffic switching.Nodes update their parameters over time and are logged out when they stop providing resources.Source: §5.1 · p. 17

5.2 Computing Offloading

  • Offloading must address limited edge capacity, heterogeneous node capabilities, and changing loads, while coordinating task selection and division across nodes and domains.The surveyed objectives include reducing latency, balancing energy use, and improving resource utilization according to application or system needs.Source: §5.2 · pp. 17–18, §5.2.1 · p. 18

5.2.2 Optimization strategies

  • Offloading strategies vary by source and destination and by whether tasks are indivisible or split across nodes; optimization approaches include convex optimization, game theory, and deep reinforcement learning.The survey describes both full and partial offloading, including parallel processing of splittable tasks.Source: §5.2.2 · pp. 18–19

5.3 Service Orchestration

  • Service orchestration can switch between containerized services as user needs change, activating the requested service and closing another when node resources permit only one at a time.The section frames this as a way to allocate resources dynamically and improve utilization.Source: §5.3 · p. 19

5.4 Resource Management With Intelligence

  • The survey proposes AI support for resource awareness, demand prediction, fault detection, and distributed offloading decisions, including GNNs, RNNs or LSTMs, and multi-agent deep reinforcement learning.The described uses include modeling dynamic global resource information and predicting user requests to orchestrate services in advance.Source: §5.4 · pp. 19–20

VI. INTELLIGENT COMPUTING POWER ROUTING AND FORWARDING IN COMPUTING POWER NETWORK

The section describes SRv6-based forwarding that maps application service requirements to network paths, then identifies AI-assisted path selection as a research direction.

VI. INTELLIGENT COMPUTING POWER ROUTING AND FORWARDING IN COMPUTING POWER NETWORK

  • SRv6 forwarding is used to carry both IP routing and computing-service routing, with application-to-path matching completed at the entrance gateway.The survey describes SRv6 as using the IPv6 forwarding plane and source-routing capabilities.Source: §VI · p. 20
  • The forwarding process announces computing-network information, maps a service ID to a service IP and exit node at the entry router, and then forwards subsequent messages using the FIB.The computing node processes the client task before the described subsequent-message forwarding.Source: §VI · p. 20, Figure 7 · p. 20
  • The survey identifies AI-based path selection as promising, citing traffic prediction, online policy updates from forwarding data, and federated learning for privacy protection.The stated challenge is selecting paths among many connections to ubiquitous computing nodes.Source: §VI · p. 20

VII. COMPUTING POWER NETWORK PLATFORM

The section introduces a transaction platform connecting computing-power consumers, providers, and managers, and a resource-orchestration platform for flexible service deployment and scheduling.

7.1 Computing Power Network Transaction Platform

  • The transaction platform brings consumers requesting computing and network services together with authenticated providers and a manager who handles service contracts and usage-based billing.Providers may include edge or cloud nodes, supercomputing centers, operators, enterprises, or individual devices.Source: §7.1 · pp. 20–21
  • The survey identifies a trusted blockchain-based transaction platform as significant for commercialization and cites a framework for ubiquitous AI.The cited framework aims to help parties access optimal strategies and obtain expected benefits.Source: §7.1 · p. 21

7.2 Computing Power Network Resource Orchestration Platform

  • The orchestration platform is described as supporting distributed, adaptive, jointly optimized resource scheduling; the survey presents container-cluster management, including Kubernetes and K3S, as a direction for implementation.Kubernetes can manage multiple clusters, while K3S is identified for managing limited computing nodes.Source: §7.2 · pp. 21–22, Figure 9 · p. 22

VIII. COMPUTING POWER NETWORK TESTBED

The testbed combines computing-resource awareness, modeling, and graph-based offloading to coordinate edge resources. Its evaluations report lower response delay and stronger performance as network topology changes, while identifying the prototype’s network-architecture scope as a limitation.

8.1 Computing Power Network Overview

  • The prototype uses ComNet to represent computing resources and link quality, models edge-node computing power, and applies graph matching to minimize request response delay.The testbed uses Kubernetes to deploy microservices and manage distributed resources; its applications include face recognition, traffic detection, image rendering, and distributed model training.Source: §VIII · p. 22, §8.1 · pp. 22–23

8.2 Computing Power Network Test Results

  • The testbed is evaluated on four edge nodes using video-stream tasks and common local, round-robin, and greedy scheduling strategies for comparison.The terminal uploads preprocessed multi-channel video to an edge node over gRPC.Source: §8.2 · p. 23
  • About 25% lower response delay than RR, with node load similar to RR and much better than GR, is reported for the proposed offloading strategy.The comparison is under constant network topology, with only node 2 connected to multiple devices.Source: Figure 11 · p. 23, §8.2 · p. 23
  • After a node rejoins the cluster, throughput recovers within fewer time slots, and the strategy maintains the best performance throughout the topology change.Node 3 is withdrawn at time slot 50 and rejoins after time slot 200; computing awareness corrects offloading decisions as topology changes.Source: §8.2 · pp. 23–24, Figure 12 · p. 24
  • The prototype schedules distributed computing resources based on Kubernetes without updating the network architecture.The paper describes programmable network devices for network-layer scheduling as a longer-term direction.Source: §8.2 · p. 24

IX. APPLICATIONS AND USE CASES

The survey describes applications that use CPN to coordinate computing resources across terminals, edge nodes, and clouds. The use cases include terminal resource sharing, cloud gaming, connected vehicles, smart cities, and collaborative AI training.

9.1 Terminal Computing Power Network

  • TCAN brings terminals into CPN to share spare computing power for intensive tasks, while accounting for limited battery capacity and widely distributed devices.Its goals include terminal-side resource perception, collaboration, scheduling, and trusted transactions.Source: §9.1 · p. 24

9.2 Cloud Game

  • For large-scale cloud games, CPN can coordinate multiple edge clouds and schedule tasks to idle computing nodes or nearby terminals when the closest edge cloud is busy.A single edge cloud handles small-scale cloud game tasks; the described data flow sends control commands to the edge cloud and rendered images back to end devices.Source: §9.2 · p. 25

9.3 Connected Vehicles

  • Connected-vehicle tasks can be offloaded according to delay needs and node availability, while CPN shares vehicle and environmental information across edge clouds for broader analysis.The example sends delay-insensitive assisted-driving work to an available edge cloud and delay-tolerant in-car entertainment to the cloud.Source: §9.3 · pp. 25–26

9.4 Smart City

  • In smart cities, CPN maps tasks from devices such as monitors and delivery robots to service instances and can update microservices in real time.The survey contrasts this with edge computing nodes commonly treated as IaaS and describes CPN as spanning IaaS, PaaS, and SaaS.Source: §9.4 · p. 26

9.5 AI Model Training

  • CPN can distribute large-scale VGGNet training across computing nodes by splitting the model and transmitting intermediate results, whereas edge computing is described as allowing only two segments.The example describes cooperative training across EC1, EC2, and EC3 with shorter latency.Source: §9.5 · p. 26

X. ENABLING TECHNOLOGIES

The survey reviews technologies for managing, connecting, and deploying CPN resources and services. It presents these technologies as supporting intelligent management, flexibility, scalability, and security.

10.1 Multi-access Edge Computing

  • Multi-access edge computing provides storage and computing at the edge, while efficient collaboration between MEC and cloud computing remains a CPN challenge.The cited examples include partial data offloading and cloud-assisted feature extraction on a low-earth-orbit satellite server.Source: §10.1 · p. 27

10.2 Software Defined Network

  • SDN separates control and data planes, enabling centralized perception of computing resources and task scheduling based on resource and link state.The survey also describes SDN for coordinating communication and computation resources across operators and technologies.Source: §10.2 · p. 27

10.3 Network Function Virtualization

  • NFV can support on-demand CPN computing-power orchestration and scheduling by virtualizing network functions on general-purpose computing platforms.The survey reviews work on joint MEC-cloud orchestration and VNF placement and resource allocation.Source: §10.3 · pp. 27–28

10.4 Network Slicing

  • Network slicing partitions shared infrastructure into end-to-end logical networks with dedicated virtualized resources and adjustable bandwidth for particular applications.The survey presents slicing as a way to share CPN computing and networking resources while isolating services.Source: §10.4 · p. 28

10.5 Docker Container

  • Docker containers and orchestration support service decomposition, deployment, scaling, and lifecycle management of microservices in CPN.The survey notes that containers isolate resource-use groups and that orchestration coordinates otherwise uncoupled components.Source: §10.5 · p. 28

10.6 Segment Routing IPv6

  • SRv6 uses IPv6 extension headers for network programming and can support flexible forwarding and resource scheduling in CPN.The survey notes that traditional routers need hardware support and updates for SRv6.Source: §10.6 · p. 28

10.7 Artificial Intelligence

  • AI both provides services that need CPN resources and offers algorithms such as DQN, DDPG, and federated learning for optimizing resource scheduling.The survey anticipates more intelligent and flexible CPN as AI progresses.Source: §10.7 · pp. 28–29

XI. OPEN CHALLENGES AND FUTURE DIRECTIONS

The survey identifies open challenges in modeling computing power, scaling information announcement, supporting user mobility, and addressing communication bottlenecks. It presents these as directions for future CPN research.

11.1 Computing Power Modeling

  • No general computing-power metric exists, and modeling needs broader characteristic factors and more accurate fitted functions.The survey suggests machine learning may help fit more accurate models.Source: §11.1 · p. 29

11.2 Information Synchronization

  • Current distributed-node information announcement limits CPN scale, leaving open the design of faster-converging protocols and possible decoupling from network routing.The survey notes that node information can be announced through IGP and BGP protocols.Source: §11.2 · p. 29

11.3 Mobility Enhancement

  • User mobility can increase the distance between users and allocated resources, raising latency and underutilizing resources; the survey calls for mobility enhancement across horizontal and vertical movement.The stated goal is to allocate optimal computing resources as users move.Source: §11.3 · p. 29

11.4 Communication Bottleneck

  • Growing data volumes challenge CPN communication capacity, motivating data compression, load balancing to avoid link congestion, and in-network processing.The survey describes in-network computing as using routers and switches to process data during communication.Source: §11.4 · p. 29

11.5 Energy Consumption

The survey identifies energy efficiency, deterministic low-latency forwarding, trusted transactions and management, and pricing as CPN research concerns.

11.5 Energy Consumption

  • Energy reduction spans lower-energy, higher-performance chips and task partitioning/offloading; prioritizes offloading energy, while minimizes total energy under latency constraints.The latter considers both partial and binary offloading.Source: §11.5 · p. 30

11.6 Deterministic Ultra-low Latency

  • DetNet's reservation, cycle mapping, path binding, and aggregation scheduling offer deterministic forwarding; combining DetNet with CPN is proposed for future ultra-low latency.The cited motivation includes rapid, accurate transmission of command information for time-sensitive tasks.Source: §11.6 · p. 30

11.7 Security

  • Blockchain smart contracts can encode trading rules and prices, while distributed trusted management can support access and services across heterogeneous nodes.The discussion presents trusted management as important for the security of computing-power providers and consumers.Source: §11.7 · p. 30

11.8 Pricing Policy

  • Pricing remains distinct because users prioritize price and providers profit; cited approaches use backward induction and auctions for cloud/fog resources.The survey notes that pricing policy significantly influences CPN operation profit.Source: §11.8 · p. 30

XII. CONCLUSION

The conclusion presents CPN as a collaborative, dynamic scheduling paradigm and recaps the survey's coverage from architecture and operational issues to testbed, applications, technologies, and future challenges.

XII. CONCLUSION

  • CPN is presented as a paradigm for dynamically, flexibly, collaboratively, and on-demand scheduling computing, storage, and network resources.The conclusion also presents its concept and functional architecture.Source: §XII · p. 30
  • The survey reviews modeling, information awareness and announcement, resource allocation, SRv6-based routing and forwarding, and transaction and orchestration platforms.Resource allocation is organized around resource management, computing offloading, and service orchestration.Source: §XII · p. 31
  • The conclusion also covers the CPN testbed, applications and use cases, enabling technologies, and open challenges and future directions.The testbed is based on the key technologies discussed in the survey.Source: §XII · p. 31

Biographies

The biographies identify the authors' academic and industry roles and research interests.

Biographies

  • Sun, Liu, and Huang are BUPT researchers whose listed interests include CPN, offloading and resource allocation, and edge computing.Sun is pursuing a Ph.D., Liu an M.E., and Huang a master's degree.Source: Biographies · p. 37
  • Lei directs China Telecom's future-network research center, while Peng leads an innovation department at China Telecom Beijing Branch.Source: Biographies · pp. 37–38
  • Zhang and Wang hold professorial roles at BUPT; their stated interests span 5G/6G, IoT, wireless transmission, and software radio.Source: Biographies · p. 38
Loading 2210.06080v2…