Dynamic closed-loop optimization-based arrangement system for 6G computing network time delay sensitive service

Abstract

The invention provides a dynamic closed-loop optimization-based calculation and network integrated arrangement system oriented to time delay sensitive services. The system comprises a resource scheduling layer, a data layer and user equipment, the resource scheduling layer collects computing network resource information and user service demand information through a controller, constructs a computing network resource symbiosis diagram according to the collected information, and uniformly schedules switches and computing power nodes in the network; the data layer receives a user service request sent by the user equipment, provides a computing power service for the user service request through the computing power cluster by using the computing network resource symbiosis graph, and returns computing result data to the user equipment through a network; and the user equipment is used for sending a user service request to the data layer, receiving the calculation result data issued by the data layer and processing the calculation result data, and the user service request carries the priority information of the time delay sensitive service. According to the method, a real-time feedback mechanism and closed-loop dynamic optimization are adopted, so that dynamic service requirements and network environment changes are quickly adapted.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication networks, and in particular, to a computing-network integrated orchestration system based on dynamic closed-loop optimization for delay-sensitive services. Background Art

[0002] With the continuous development of communication technologies and the increasing demand for network services, the fifth-generation (5G) mobile communication technology has greatly expanded the service capabilities of mobile networks. Through the introduction of new radio access technologies, network architecture optimization, and service-oriented core network design, 5G networks have the characteristics of diverse scenarios, flexible deployment, and dynamic requirements, significantly reducing network latency and improving data transmission efficiency. Ultra-reliable low-latency communication (URLLC) is an important application scenario of 5G technology. In the future, 6G will go beyond the scope of traditional mobile communications. Based on new infrastructure such as computing networks, it will dynamically configure the required network functions on demand through software, realizing the integration of capabilities such as communication, computing, AI, perception, big data, and security, and achieving XaaS (everything as a service). In the 6G era, the development of communication and related technologies will far exceed that in the 5G era, including ubiquitous connection, ubiquitous computing, digital twin networks, and blockchain networks. With the deepening of digital transformation, emerging fields such as the Internet of Things, industrial automation, intelligent transportation systems, and telemedicine have put forward higher requirements for the real-time performance and reliability of networks. These application scenarios require fast and low-latency network connections to ensure service quality and user experience, especially in cases where immediate response is required, such as the decision-making process of autonomous vehicles and the operation of remote surgeries.

[0003] As a new type of network architecture, the computing network uniformly maintains the computing power of the cloud, edge, and end. By jointly allocating computing power resources and network resources, it aims to make computing power a resource that can be obtained and used immediately. The heterogeneous computing power resources and communication media in the computing network have diversity and performance differences. Existing service deployment solutions that model optimization objectives such as service response latency by comprehensively considering computing power resources and network resources are difficult to truly and deeply match the needs of computing power services and the performance of heterogeneous resources, resulting in the deployed services being unable to truly meet user needs. At the same time, existing solutions lack the control ability of the network and do not fully utilize network resources. In the case where the deployed services cannot effectively meet user needs due to fluctuations in computing power resources or network resources, existing solutions do not provide effective solutions.

[0004] Currently, due to the popularity of latency-sensitive services such as VR (Virtual Reality), AR (Augmented Reality), and autonomous driving, as well as 5G and 6G technologies, higher requirements are placed on network latency, bandwidth, and reliability. The scheduling and optimization of computing and network resources in the computing power network also face more arduous challenges.

[0005] With the continuous progress of communication technologies and the growing demand for network services, the fifth-generation (5G) mobile communication technology has significantly enhanced the service capabilities of mobile networks. By adopting new radio access technologies, optimizing network architectures, and using service-based core network designs, 5G networks exhibit characteristics such as diverse scenarios, flexible deployments, and dynamic demands, greatly reducing network latency and improving data transmission efficiency. Among them, ultra-reliable low-latency communication (URLLC) is an important application direction of 5G. In the future, the sixth-generation (6G) mobile communication technology will go beyond the scope of traditional communications. Based on new infrastructure such as the computing power network, it will achieve dynamic on-demand configuration of the network through software, integrating capabilities such as communication, computing, artificial intelligence, perception, big data, and security to realize "everything as a service" (XaaS). In the 6G era, the development of communication and related technologies will far exceed that of the 5G era, covering emerging fields such as ubiquitous connection, ubiquitous computing, digital twin networks, and blockchain networks.

[0006] At the same time as network technology progresses, computing demands are also growing rapidly, and the computing power network emerges as an innovative network architecture. By integrating computing resources in the cloud, at the edge, and on terminals, this architecture enables unified scheduling and flexible management of computing capabilities, aiming to provide a computing power service that can be allocated on demand and is immediately available. However, in the process of achieving this goal, the current computing power network still faces many challenges. First, existing solutions have deficiencies in network control and fail to fully utilize the potential of network resources. In practical applications, once fluctuations occur in computing resources or network resources, resulting in the service being unable to effectively meet user needs, existing solutions often lack effective coping strategies. In addition, although existing service deployment solutions attempt to comprehensively consider computing resources and network resources to optimize key metrics such as service response time, these solutions usually struggle to fully meet the dynamically changing computing power service demands, resulting in the actual deployed service falling short of user expectations.

[0007] With the deepening of digital transformation, emerging fields such as the Internet of Things, industrial automation, intelligent transportation systems, and telemedicine have put forward higher requirements for the real-time performance and reliability of the network. These application scenarios require fast and low-latency network connections to ensure service quality and user experience, especially in situations that require immediate response, such as the decision-making process of autonomous vehicles and the operation of remote surgeries. At the same time, with the rapid development of 5G and 6G technologies, the network's requirements for latency, bandwidth, and reliability are increasing day by day, and the resource scheduling and optimization of the computing power network are facing more severe challenges.

[0008] To address these challenges, the present invention proposes a method for orchestrating latency-sensitive services in an integrated computing and network system with dynamic closed-loop optimization. By real-time monitoring and dynamically adjusting computing and network resources, this method aims to provide low-latency and high-reliability computing power services for latency-sensitive services, thereby enhancing user experience and service quality.

[0009] An integrated computing and network resource scheduling system in the prior art includes a computing and network orchestration module, a computing and network scheduling module, a computing and network perception module, and an intelligent decision-making module. 1) The computing and network perception module is used to interface with the managed computing and network resources and collect computing and network resource data; 2) The intelligent decision-making module is used to provide computing and network resource orchestration strategies for various service scenarios based on pre-set artificial intelligence algorithms; 3) The computing and network orchestration module is used to obtain service requirements, determine a target computing and network resource orchestration strategy from various computing and network resource orchestration strategies according to the service requirements, and perform computing and network resource scheduling and orchestration on the computing and network resource data based on the target computing and network resource orchestration strategy combined with the service requirements to obtain a scheduling plan; 4) The computing and network scheduling module is used to execute the scheduling plan.

[0010] The disadvantages of the above-mentioned integrated computing and network resource scheduling system in the prior art include: This system only focuses on the allocation of computing and network resources and the scheduling of computing power resources, does not treat computing power and the network as a whole, and ignores the optimization of network routing.

[0011] A deterministic control method for distributed training in the prior art includes an integrated control module, a computing power configuration module, and a network configuration module. In the integrated control module, a software or device for globally managing network and computing power resources is proposed, namely, a computing network resource integrated controller, which supports deterministic management of the distributed training process and can globally configure deterministic policies for each training task of the central server. The management scope of the controller includes industrial terminals, central servers, and network controllers. It can control the local model training of industrial terminals and the overall model training of central servers, and can also send deterministic configuration parameters to the central server, industrial terminals, and network controllers respectively. Among them, the computing power task of local model training is a computing task in which each industrial terminal device uses the real-time data collected by itself as the input of the local model and obtains a new state model through machine learning or artificial intelligence algorithms; the computing task of overall model training is that the central server of the distributed algorithm merges multiple local models in dimensions such as parameter sets, attribute sets, and model structures into an overall model.

[0012] The computing power configuration module is used to configure deterministic computing power metrics to the central server and industrial terminals.

[0013] The network configuration module is used to configure deterministic network metrics to the network controller and network devices.

[0014] The disadvantages of the above-mentioned deterministic control method for distributed training in the prior art include: This method can only handle the limitations of deterministic services, and its flexibility and adaptability are insufficient.

[0015] A computing network integrated resource scheduling scheme for delay-sensitive services in the prior art includes: aiming to achieve low-latency and high-reliability computing power services by optimizing service deployment and path selection. This scheme is mainly divided into the following three parts:

[0016] The computing network integrated resource scheduling architecture for delay-sensitive services. Objective: Provide low-latency and reliable computing power services for delay-sensitive services while maximizing the benefits of service providers. Method: Among the computing power nodes that meet the requirements of the service, several computing power nodes are selected for service deployment according to the user distribution and the usage status of computing power resources, which can ensure that most users can quickly connect to the appropriate computing power service when initiating service requests without a complex decision-making process. During the computing power service process, if the fluctuations in computing power or network status cause an increase in service response latency and cannot meet the user's needs, the scheme will utilize the advantages of network control to actively allocate network resources. Specifically, by finding a path with better network transmission performance between the user and the computing power service node, the network transmission latency is reduced, thereby meeting the user's demand for low latency.

[0017] Computing power network service deployment mechanism optimized based on delay prediction. Objective: To solve the problem of service deployment failure caused by user mobility and achieve accurate prediction of the delay in the future period after service deployment. Method: Model the service deployment problem as a multi-constraint optimization problem with computing power resources as constraints and maximizing the number of service users as the goal. Use the KNN algorithm to detect abnormal delays, the RouteNet algorithm to correct abnormal delays, and the LSTM algorithm to predict delays. Based on the accurately predicted delays, use the simulated annealing algorithm for service deployment to ensure the effectiveness of service deployment decisions.

[0018] Network active allocation mechanism adapted to service performance fluctuations: Objective: To address the problem of service performance degradation caused by fluctuations in computing power resources and network resources and ensure that users' delay requirements are met. Method: Model the problem as multi-QoS constraint path optimization, and find the path with the shortest delay between users and computing nodes in the network under constraints such as bandwidth, packet loss rate, and jitter. To address the problem that traditional heuristic algorithms are prone to falling into local optimal solutions, propose the FLGA algorithm for optimization.

[0019] This technology can improve the scheduling efficiency of delay-sensitive services in the computing power network.

[0020] The disadvantages of the above-mentioned existing technology's computing-network integrated resource scheduling scheme for delay-sensitive services in the computing power network include:

[0021] This scheme adopts a lightweight delay prediction model based on reinforcement learning. Compared with the computing power network service deployment mechanism optimized based on delay prediction in Technology Three, delay prediction depends on the combination of algorithms such as KNN, RouteNet, and LSTM. In practical applications, it can avoid being affected by factors such as data noise and model overfitting, and can improve the response speed.

[0022] This technology constructs a dynamic and real-time "computing-network resource symbiotic graph", which can use traditional graph optimization algorithms to quickly generate the optimal path that meets delay and resource requirements and send it down. In the network active allocation mechanism adapted to service performance fluctuations in Technology Three, it one-sidedly focuses on optimizing the network under constraints such as bandwidth, packet loss rate, and jitter, without considering the dynamic changes of computing node resources, and has a high computational complexity, which may increase the computational overhead and lead to a long optimization time. Summary of the Invention

[0023] Embodiments of the present invention provide a computing-network integrated orchestration system based on dynamic closed-loop optimization for delay-sensitive services to effectively improve the scheduling efficiency of computing-network resources.

[0024] To achieve the above object, the present invention adopts the following technical solutions.

[0025] An arithmetic-network integrated orchestration system for delay-sensitive services based on dynamic closed-loop optimization, comprising: a resource scheduling layer, a data layer, and user equipment;

[0026] The resource scheduling layer is used to collect arithmetic-network resource information and user service demand information through a controller, construct an arithmetic-network resource symbiotic graph based on the collected information, and perform unified scheduling on switches and computing nodes in the network;

[0027] The data layer is used to receive user service requests sent by user equipment, provide computing power services for the user service requests through a computing power cluster by using the arithmetic-network resource symbiotic graph, and return the calculation result data to the user equipment through the network;

[0028] The user equipment is used to send user service requests to the data layer. The user service requests carry priority information of delay-sensitive services, receive the calculation result data sent by the data layer, and process the calculation result data.

[0029] Preferably, the resource scheduling layer includes an arithmetic-network resource symbiotic graph construction module, a lightweight prediction and regulation module, a service orchestration and path planning module, and a resource configuration and distribution module;

[0030] The arithmetic-network resource symbiotic graph construction module is used to monitor arithmetic-network resource information in real time through a controller, represent resources and overheads with nodes and edges, and construct and update the arithmetic-network resource symbiotic graph;

[0031] The lightweight prediction and regulation module is used to predict future service traffic and resource demands by using a built-in reinforcement learning model, pre-configure arithmetic-network resources, and feedback real-time monitoring indicators to the arithmetic-network resource symbiotic graph construction module;

[0032] The service orchestration and path planning module is used to collect user service demands, plan paths for services through the Dijkstra algorithm by using the arithmetic-network resource symbiotic graph and user service demands, orchestrate services through a service function chain (SFC), and allocate resources for each level of service;

[0033] The resource configuration and distribution module is used to, according to the service orchestration result, use the arithmetic-network resource symbiotic graph and user service demands, uniformly schedule the computing power resources and network bandwidth resources of user equipment through network function virtualization network slicing and the docker method, generate resource scheduling instructions, and return the resource scheduling instructions to the user equipment along the optimal path.

[0034] Preferably, the lightweight prediction and control module monitors the operating status of the computing power cluster and the link load information in the data network, and feeds back the monitored information to the computing-network resource symbiotic graph construction module and the service orchestration and path planning module. The computing-network resource symbiotic graph construction module re-optimizes the computing-network resource symbiotic graph according to the feedback information and adjusts the resource scheduling strategy, and the service orchestration and path planning module re-optimizes the path planning scheme according to the feedback information.

[0035] Preferably, the data layer includes a core network, a data network, and a computing power cluster. The computing power cluster includes GPU resources, CPU resources, and storage resources. The computing power cluster provides computing power services, receives user service requests sent by user equipment, and coordinates the computing power cluster, the data network, and the core network to complete service processing and data transmission according to the resource scheduling information issued by the resource scheduling layer and the information of the service orchestration and path planning module, and returns the calculation result data to the user equipment through the network. The data network includes a routing network composed of programmable switches, and completes the forwarding task by receiving controller instructions; the core network includes a 5G or 6G core network and some of its network elements, which are used for user access to the core network, opening data channels, and allocating network slice processing, and transmits the received user service requests to the resource scheduling layer.

[0036] Preferably, the data sent by the data layer to the resource scheduling layer includes: computing power cluster status data, data network link status data, and core network user access related data. The data received by the data layer from the resource scheduling layer includes: resource scheduling instructions, network slice configuration instructions, and service orchestration and path planning results.

[0037] Preferably, after the user equipment receives the calculation result data sent by the data layer, it calls the corresponding parsing program to parse the received data, and presents the parsed data in a form that can be perceived by the user; for the image recognition result data, according to the established data format, extract the object category and location information. For the image recognition result, display the text description of the recognition result on the display screen of the user equipment or mark the recognized object on the original image; for the video file after transcoding, play it through the video playback software of the device.

[0038] It can be seen from the technical solutions provided by the embodiments of the present invention described above that the method of the present invention adopts a real-time feedback mechanism and closed-loop dynamic optimization, so as to quickly adapt to dynamic service requirements and network environment changes. The computing power resources and network resources are constructed into a "symbiotic graph", and the optimal resource combination is determined in real time through graph optimization algorithms (such as the shortest path and minimum weight matching) to ensure the minimum delay. A lightweight delay prediction model (such as based on reinforcement learning or graph neural network) is built in to predict future service traffic and resource requirements, and realize early scheduling. Pre-configure computing-network resources before the service arrives to reduce the additional delay caused by dynamic allocation.

[0039] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent from the following description, or will be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0041] Figure 1 It is a processing flow chart of a computing and network integrated orchestration system based on dynamic closed-loop optimization for delay-sensitive services provided by an embodiment of the present invention;

[0042] Figure 2 It is a flow chart for implementing closed-loop optimization within a controller provided by an embodiment of the present invention;

[0043] Figure 3 It is a "symbiotic diagram" of computing and network resources provided by an embodiment of the present invention;

[0044] Figure 4 It is a reinforcement learning training flow chart of a lightweight prediction and regulation module provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0046] Those skilled in the art of the present technology can understand that unless specifically stated otherwise, the singular forms "a", "an", "the" and "said" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present invention means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The phrase "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0047] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the technical field to which the present invention pertains. It should also be understood that terms defined in common dictionaries should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0048] For ease of understanding the embodiments of the present invention, the following will further explain with several specific embodiments in conjunction with the drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0049] The embodiments of the present invention are directed to delay-sensitive services in a computing power network, and design a dynamic closed-loop optimized integrated computing and network delay-sensitive service orchestration method, aiming to provide low-latency and reliable computing power services for delay-sensitive services.

[0050] A schematic diagram of the deployment of delay-sensitive services provided by the embodiments of the present invention is as Figure 1 shown, including a resource scheduling layer, a data layer, and a user UE (User Equipment).

[0051] The resource scheduling layer includes a controller connected to the computing power cluster of the data network through a switch. The controller collects computing and network resource information and user service demand information, constructs a computing and network resource symbiotic graph based on the collected information, and uniformly schedules switches and computing power nodes in the network. The resource scheduling layer includes a computing and network resource symbiotic graph construction module, a lightweight prediction and regulation module, a service orchestration and path planning module, and a resource configuration and distribution module.

[0052] The computing and network resource symbiotic graph construction module is used to monitor computing and network resource information in real time through the controller, represent resources and overheads with nodes and edges, and construct or update a computing and network resource status model. The service orchestration and path planning module collects user service demands, combines with the "symbiotic graph" of computing and network resources, plans paths using the Dijkstra algorithm, and orchestrates services with SFC (Service Function Chain), and allocates resources for high-priority services. The resource configuration and distribution module constructs a network and deploys functional modules based on the orchestration result using NFV (Network Functions Virtualization) network slicing and docker technology. The lightweight prediction and regulation module uses a built-in reinforcement learning model to predict future service traffic and resource demands, pre-configure computing and network resources, and feedback real-time monitoring indicators to dynamically adjust resource allocation.

[0053] The resource scheduling layer uniformly schedules two types of key resources of user equipment. Computing resources: Flexibly determine when to enable local CPU (Central Processing Unit) and GPU (Graphics Processing Unit) resources during local computing of user equipment, and when to request additional computing power support from the computing power cluster to meet the computing requirements of different services. For example, when processing complex tasks such as high-definition video rendering, if local GPU resources are insufficient, some tasks will be allocated to high-performance GPU nodes in the computing power cluster; Network bandwidth resources: According to the latency sensitivity and data traffic requirements of services, allocate network channels with different priorities and bandwidth sizes for user equipment. For example, for services such as real-time video calls that are extremely sensitive to latency, high-priority and large-bandwidth channels will be allocated to ensure the smoothness of the call process.

[0054] The resource scheduling layer plans the complete data transmission path of user equipment from the initiation of a request, how data accesses through the core network, is routed by programmable switches in the data network, reaches the computing power cluster, and then returns the calculation results to the user equipment along the optimal path.

[0055] The scheduling information and planning information do not need to be sent to the user equipment. The user does not need to know the orchestration results of resources and paths, but only needs to receive and process the computing power. When the user equipment initiates a computing request, the resource scheduling layer allocates computing power resources for the user according to the computing-network resource symbiosis graph and service requirements. After the computing power cluster completes the calculation, it directly returns the results to the user equipment.

[0056] The collection of feedback information for the system's closed-loop dynamic optimization does not rely on the user equipment to feedback information such as resource usage status, but monitors internal system metrics such as the operating status of the computing power cluster and link loads in the data network, and uses this feedback information to re-optimize the computing-network resource symbiosis graph, and then adjusts the resource scheduling strategy and path planning scheme to improve the service quality of latency-sensitive services and resource utilization efficiency.

[0057] When a service is in an execution state, such as providing high-definition video rendering services for a user, the controller will monitor key performance indicators such as latency, packet loss rate, and computing power occupancy rate in real time, feedback to the computing-network resource symbiosis graph construction module for real-time update, and then enter the optimization engine in the path planning and service orchestration module to form a closed-loop optimization, so as to more precisely dynamically adjust the resource allocation strategy and quickly adapt to dynamic service requirements and network environment changes.

[0058] The data layer includes the core network, the DN (Data Network), and the computing power cluster. The computing power cluster includes GPU resources, CPU resources, and storage resources. The computing power cluster provides computing power services, receives user computing requests, and returns the computing results to the user through the network. The data network includes a routing network composed of programmable switches, which completes forwarding tasks by receiving controller instructions; the core network includes the 5G (6G) core network and some of its network elements, which are used for user access to the core network, opening data channels, allocating network slices, etc.

[0059] The data sent by the data layer to the resource scheduling layer includes:

[0060] 1 Computing power cluster status data: The real-time usage of GPU, CPU resources, and storage resources in the computing power cluster, such as information on resource utilization rate, idle resource quantity, etc., so that the resource scheduling layer can reasonably allocate computing power for user services. For example, when the GPU resource utilization rate is too high, the resource scheduling layer can consider allocating some computing tasks to other idle GPU nodes or using CPU resources for auxiliary computing.

[0061] 2 Data network link status data: The link load conditions of the routing network composed of programmable switches, including real-time traffic, bandwidth occupancy rate, delay, etc. of each link. Based on these data, the resource scheduling layer can select a better link during path planning to ensure the high efficiency of data transmission. If the bandwidth occupancy rate of a certain link continues to be high and the delay increases, the resource scheduling layer will re-plan the data transmission path to avoid this link.

[0062] 3 Core network user access-related data: Information about user access in the core network, such as changes in the number of users, types and distributions of user devices, and the usage status of network slices. These data help the resource scheduling layer understand the overall situation on the user side and better allocate network resources and computing power resources for different user services to achieve precise resource scheduling.

[0063] The data received by the data layer from the resource scheduling layer includes:

[0064] 1 Computing-network resource scheduling instructions: Resource scheduling instructions generated by the resource scheduling layer based on the computing-network resource symbiotic graph and user service requirements, including instructions for allocating resources in the computing power cluster, such as allocating a specific computing task of a certain user to a specified GPU or CPU resource for processing; forwarding instructions for programmable switches in the data network, specifying which links the data should be transmitted through to achieve optimal path planning.

[0065] 2 Network Slicing Configuration Instructions: According to the priority and requirements of services, the resource scheduling layer sends network slicing configuration instructions to the core network to adjust the resource allocation of network slices. For high-priority services sensitive to latency, it is required that the core network allocate high-bandwidth and low-latency network slice resources for them to ensure the quality of service of the services.

[0066] 3 Service Orchestration and Path Planning Results: The resource scheduling layer sends the specific scheme for user service orchestration and the planned data transmission path information to the data layer. Based on this information, the data layer coordinates the computing power cluster, data network, and core network to complete service processing and data transmission, and accurately returns the calculation results to the user device.

[0067] The input data of the computing power cluster comes from the computing requests initiated by the user device, including various service-related data. For example, in the image recognition service, the input data is the image file to be recognized by the user; in the video editing service, the input is the original video segment uploaded by the user. These data cover different formats and types, and the data volume is determined by the user service requirements. The output data is the processed calculation result. For the image recognition service, the output is the recognized object category, location information, etc.; the output of the video editing service may be the video finished product after editing and adding special effects.

[0068] The calculations performed by the computing power cluster depend on the service type. In deep learning model training, GPU resources are used for large-scale matrix operations to optimize model parameters; in scientific computing tasks, CPU resources are used for complex numerical calculations, such as solving partial differential equations. In the data storage and retrieval service, storage resources are used for data reading, writing, index construction, etc. Through the collaboration of various computing resources, the computing needs of various services are met.

[0069] After receiving the user service request sent by the user UE, the data layer directly sends it to the resource scheduling layer. The resource scheduling layer has the ability to construct a symbiotic graph of computing and network resources, and can make overall arrangements by integrating computing and network resource information and user service requirements. The core network in the data layer is responsible for user access and transfers the service request to the resource scheduling layer. After receiving the request, the resource scheduling layer uses its modules such as the symbiotic graph construction module of computing and network resources, service orchestration and path planning module, etc., analyzes the request content, combines with computing and network resource status data, link status data, etc., determines how to schedule resources and plan paths, and then sends corresponding instructions to the data layer to collaborate on service processing, rather than having the data layer process the service request first.

[0070] A user UE is used to send a user service request to the data layer, and the user service request carries the priority information of the latency-sensitive service. The user UE receives the data sent by the data layer, which is the result data after the computing cluster has completed the calculation. The format and content of this data are related to the service request initially initiated by the user. If the user initiates an image recognition request, the data sent by the data layer may be information such as the recognized object category and location coordinates; if it is a video transcoding service, the data sent is the transcoded video file.

[0071] The user UE performs subsequent processing procedures: 1. Data parsing: The user UE calls the corresponding parsing program to parse the received data according to the service type. For the image recognition result data, according to the established data format, key information such as the object category and location is extracted; for the video file data, decoding operations are performed according to the video coding format so that it can be normally played on the user device. 2. Service presentation: The parsed data is presented in a form that can be perceived by the user. For the image recognition result, the text description of the recognition result is displayed on the display screen of the user device or the recognized object is marked on the original image; for the transcoded video file, it is played through the video playback software of the device to meet the user's service requirements.

[0072] The flowchart of implementing closed-loop optimization within a controller provided by an embodiment of the present invention is as Figure 2 shown. The computing and network resource symbiotic graph construction module uses INT (In-band Network Telemetry) to insert information into data packets in sequence through the path intermediate switching nodes in real time to monitor and collect the real-time status information of computing and network resources. The computing and network resource information includes node resource utilization rate and network link load. The controller integrates the collected data and constructs in real time a computing and network resource "symbiotic graph" representing the global network and computing resource view. A computing and network resource "symbiotic graph" provided by an embodiment of the present invention is as Figure 3 shown, Figure 3 where nodes (such as computing nodes, storage nodes, and transmission nodes) represent resources, and edges represent latency overhead and traffic bandwidth, Figure 3 and the transmission node represents the network between all types of computing nodes and the user for data exchange.

[0073] In the service orchestration and path planning module, first, the service requirements of users are collected in real time through access switches. In the path planning stage, using the "symbiotic graph" model of computing and network resources, combined with the collected service requirements of users, the graph optimization algorithm - Dijkstra algorithm is used. Taking the latency overhead and traffic bandwidth as the weights of the edges, with the user as the starting point and the computing power nodes that can meet the requirements as the ending points, several resource paths that meet the latency requirements can be quickly generated. In the service orchestration stage, SFC (Service Function Chaining) is used for orchestration, which can optimize the use of network resources and increase the control of the network. For user services with strict latency requirements, a latency-sensitive service priority engine is enabled to allocate the resource paths and computing power pools with the lowest latency for high-priority user services, ensuring that the response time of high-priority user services meets the strict latency requirements. Combining the optimization results of the above two stages, the service orchestration and path planning module allocates the optimal computing power resources and network paths for user services.

[0074] User service requirements include data volume, computational complexity, priority, and expected response time. The information data volume is the data scale and data type expected to be processed by the user; the computational complexity is the resource metric required by the service, including time complexity and space complexity; the priority is the importance of the service requirement, used to determine the priority order of resource allocation; the expected response time is the urgency of the service requirement, which affects resource scheduling and path selection; the latency-sensitive service priority engine defines multi-dimensional priority parameters for latency-sensitive services, including computational urgency, data transmission volume, and network link load conditions. From these parameters and the weighting algorithm, the priority sorting result is determined.

[0075] In the resource configuration and distribution module, according to the real-time orchestration results of the path planning and service orchestration module, the network is constructed using NFV (Network Functions Virtualization) and network slicing technologies, and the virtualized deployment of each functional module in the node cache system is realized using docker technology. After the deployment is completed, the service of each node is started to complete the corresponding functions. NFV technology can dynamically deploy virtual network functions according to the service requirements that change over time. In the application of network slicing, elastic micro-slice scheduling is started, that is, the allocation granularity of network resources is dynamically adjusted to improve the system response speed. Docker technology can automatically deploy application programs in lightweight containers and make the application programs in different containers isolated from each other and work efficiently. Docker is an open-source application container engine that allows developers to package their applications and dependent packages into a portable image, and then publish it to any machine with a popular Linux or Windows operating system, and virtualization can also be achieved. Containers use a sandbox mechanism completely and there will be no interfaces between them.

[0076] In the lightweight prediction and control module, a reinforcement learning model for lightweight latency prediction is built in to predict future service traffic and resource requirements. The prediction results are directly involved in the function deployment in the resource allocation and distribution module, pre-configuring computing network resources before the service arrives to achieve early scheduling. Figure 4 This is a flowchart of the reinforcement learning training for a lightweight prediction and control module provided by an embodiment of the present invention. The agent in the reinforcement learning model is completed with a process as Figure 4 follows for reinforcement learning training. Among them, the reward is comprehensively determined by the key performance indicators of the computing network state (such as latency, packet loss rate, computing power occupancy rate). When a service is in the execution state, the controller will also monitor the above key performance indicators in real time and feedback them to the optimization engine in the path planning and service orchestration module to more finely dynamically adjust the resource allocation strategy.

[0077] The computing network integrated latency-sensitive service orchestration method with dynamic closed-loop optimization designed by the present invention is proposed based on the "symbiotic graph" model of computing network resources. On the basis of meeting the traditional resource orchestration and path optimization of the computing network, it also focuses on the unique requirements of latency-sensitive services and further performs dynamic optimization during the service execution process.

[0078] In summary, the closed-loop optimization of the embodiment of the present invention: Different from the static solution, it improves adaptability and efficiency through real-time feedback and adjustment. Graph optimization method: Introduce the "symbiotic graph" of computing network resources to provide a new modeling method for complex multi-dimensional resource allocation problems. Priority mechanism: Provide special optimization for latency-sensitive services, not taking the global optimum as the single goal, but giving priority to ensuring the latency requirements of key services. Lightweight prediction: Different from complex scheduling models, use lightweight algorithms to predict future states, reducing model complexity and computational overhead. Elastic slicing: Dynamically adjust the network resource allocation granularity, taking into account flexibility and efficiency, and improving the system response speed.

[0079] Those of ordinary skill in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or processes in the drawings are not necessarily essential for implementing the present invention.

[0080] From the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0081] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for the relevant parts, reference can be made to the description of the method embodiments. The device and system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0082] As mentioned above, the above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A computing-network integrated orchestration system for delay-sensitive services based on dynamic closed-loop optimization, characterized in that Including: A resource scheduling layer, a data layer, and user equipment; The resource scheduling layer is used to collect computing network resource information and user service requirement information through a controller, construct a computing network resource symbiotic graph according to the collected information, and uniformly schedule switches and computing nodes in the network; The data layer is used to receive user service requests sent by user equipment, provide computing power services for the user service requests through a computing power cluster by using the computing network resource symbiotic graph, and return the calculation result data to the user equipment through the network; The user equipment is used to send user service requests to the data layer. The user service requests carry priority information of delay-sensitive services, receive the calculation result data sent by the data layer, and process the calculation result data.

2. The system according to claim 1, wherein The resource scheduling layer includes a computing network resource symbiotic graph construction module, a lightweight prediction and regulation module, a service orchestration and path planning module, and a resource configuration and distribution module; The computing network resource symbiotic graph construction module is used to monitor computing network resource information in real time through a controller, represent resources and overheads with nodes and edges, and construct and update the computing network resource symbiotic graph; The lightweight prediction and regulation module is used to predict future service traffic and resource requirements by using a built-in reinforcement learning model, pre-configure computing network resources, and feedback real-time monitoring metrics to the computing network resource symbiotic graph construction module; The service orchestration and path planning module is used to collect user service requirements, plan paths for services by using the computing network resource symbiotic graph and user service requirements through the Dijkstra algorithm, orchestrate services through a service function chain (SFC), and allocate resources for each level of service; The resource configuration and distribution module is used to uniformly schedule the computing power resources and network bandwidth resources of user equipment by using the computing network resource symbiotic graph and user service requirements according to the service orchestration result through network function virtualization network slicing and the docker method, generate a resource scheduling instruction, and return the resource scheduling instruction to the user equipment along the optimal path; 3. The system according to claim 2, wherein The lightweight prediction and regulation module monitors the operating status of the computing power cluster and the link load information in the data network, and feeds back the monitored information to the computing network resource symbiotic graph construction module and the service orchestration and path planning module. The computing network resource symbiotic graph construction module re-optimizes the computing network resource symbiotic graph according to the feedback information, adjusts the resource scheduling strategy, and the service orchestration and path planning module re-optimizes the path planning scheme according to the feedback information.

4. The system according to claim 1, wherein The data layer includes a core network, a data network, and a computing power cluster. The computing power cluster includes GPU resources, CPU resources, and storage resources. The computing power cluster provides computing power services, receives user service requests sent by user devices, and collaborates with the computing power cluster, the data network, and the core network to complete service processing and data transmission according to the resource scheduling information sent by the resource scheduling layer and the information of the service orchestration and path planning module, and returns the calculated result data to the user device through the network. The data network includes a routing network composed of programmable switches, which completes forwarding tasks by receiving controller instructions; the core network includes a 5G or 6G core network and some of its network elements, which are used for user access to the core network, opening data channels, and allocating network slice processing, and transfers the received user service requests to the resource scheduling layer.

5. The system according to claim 4, characterized in that, The data sent by the data layer to the resource scheduling layer includes: computing power cluster status data, data network link status data, and core network user access-related data. The data received by the data layer from the resource scheduling layer includes: resource scheduling instructions, network slice configuration instructions, and service orchestration and path planning results.

6. The system according to claim 1, wherein After the user device receives the calculated result data sent by the data layer, it calls the corresponding parsing program to parse the received data and presents the parsed data in a form perceptible to the user. For the image recognition result data, according to the established data format, the object category and location information are extracted. For the image recognition result, the text description of the recognition result is displayed on the display screen of the user device or the recognized object is marked on the original image; for the video file after transcoding, it is played through the video playback software of the device.

Images