A 6G network multi-dimensional resource scheduling method for immersive communication service

Abstract

The application provides a 6G network multi-dimensional resource scheduling method for immersive communication services, belongs to the technical field of 6G communication network and intelligent computing collaborative optimization, and comprises the following steps: in response to a received immersive communication service request, decoupling the immersive communication service request into multiple sub-tasks with time sequence dependency, and determining a required multi-sub-layer micro-service instance set for each sub-task; extracting multi-dimensional attribute features of each micro-service instance in the multi-sub-layer micro-service instance set corresponding to each sub-task, clustering and reducing dimensions of the micro-service instances based on the multi-dimensional attribute features, and generating a plurality of micro-service clusters; by decomposing the immersive communication service into time sequence sub-tasks and matching corresponding micro-services, the complex service can be converted into a standardized scheduling unit, and the decision space can be effectively compressed and the scheduling efficiency can be improved through micro-service feature clustering and dimension reduction, and meanwhile, a logical and physical double-layer topology is built and a cross-layer mapping constraint is set.

Description

Technical Field

[0001] This invention belongs to the field of collaborative optimization technology of 6G communication network and intelligent computing, and specifically relates to a multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services. Background Technology

[0002] With the widespread adoption of 5G and the unfolding of the 6G network vision, immersive communication (such as extended reality XR, holographic projection, and metaverse social networking) is becoming the core form of the next-generation Internet. These services break down the boundaries between the virtual and the real, exhibiting significant characteristics of "high bandwidth, low latency, and strong interactivity," and placing extremely high demands on multi-dimensional Quality of Service (QoS) indicators such as frame rate stability and network jitter.

[0003] Immersive services are typically decoupled into a set of computational tasks with strict temporal dependencies and logical connections. These tasks constitute a complex directed acyclic graph (DAG) structure, requiring dynamic mapping and scheduling in heterogeneous network environments. However, this high-dimensional, fine-grained orchestration faces significant challenges in the complex 6G network environment. First, with the expansion of network scale and the refinement of task granularity, the action space for resource scheduling expands exponentially, making it difficult for traditional optimization algorithms to converge under millisecond-level real-time requirements. Second, immersive services have strict structural requirements for the connectivity and temporal dependencies of the logical topology. Existing graph-based or heuristic mapping methods often require the introduction of numerous additional constraints to enforce the connectivity and temporal dependencies of the logical topology at the physical layer. Existing scheduling schemes often focus on single end-to-end latency optimization, neglecting the multi-dimensional coupling relationship between immersive communication and QoS guarantee and load balancing, making it difficult to achieve optimal overall performance. Summary of the Invention

[0004] To address the problems in the background technology, this invention proposes a multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services includes: In response to a received immersive communication service request, the immersive communication service request is decoupled into multiple subtasks with temporal dependencies, and a set of multi-layer microservice instances required for each subtask is determined. Extract the multi-dimensional attribute features of each microservice instance in the multi-sub-layer microservice instance set corresponding to each subtask, and perform clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters; A logical functional topology is constructed based on the temporal dependencies of subtasks, and a physical network topology is constructed based on the physical network. A cross-layer mapping constraint is established from the logical functional topology to the physical network topology. The cross-layer mapping constraint is used to ensure the topological consistency of the logical execution order of the subtasks during the physical deployment process. Based on the microservice cluster and the cross-layer mapping constraints, a pre-trained multi-agent deep reinforcement learning model is used to match microservice instances from the microservice cluster for each subtask, and deploy the matched microservice instances to physical nodes of the physical network. At the same time, the link routes between the physical nodes are determined to obtain a multi-dimensional resource scheduling scheme.

[0006] Preferably, the step of decoupling the immersive communication service request into multiple sub-tasks with temporal dependencies, and determining the required set of multi-layered microservice instances for each sub-task, specifically includes: The service level protocol and QoS requirements of the immersive communication service request are parsed, and the immersive communication service request is decomposed into multiple sub-tasks according to the functional sub-layers. The resource requirement weight of each sub-task in each sub-layer of the functional sub-layer architecture is determined. The functional sub-layers include an algorithm sub-layer, a data sub-layer, a computing power sub-layer, and a connection sub-layer. Based on the resource requirement weight and functional description of each subtask, microservice instances are selected to form a multi-layered microservice instance set.

[0007] Preferably, the QoS requirements include: rendering quality and tracking stability of the algorithm sub-layer, frame rate and data jitter of the data sub-layer, computing power accuracy of the computing power sub-layer, and network jitter of the connection sub-layer.

[0008] Preferably, the step of extracting multi-dimensional attribute features of each microservice instance in the multi-layer microservice instance set corresponding to each subtask, and performing clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters, specifically includes: Construct a multi-view hypergraph that integrates performance metrics, latency, and resource overhead views; wherein the vertices of the multi-view hypergraph are each microservice instance, and the hyperedges are used to connect microservice instances with similar attributes under different views; Construct the adjacency matrix of the multi-view hypergraph, and perform spectral clustering analysis on the adjacency matrix to aggregate microservice instances with similar multidimensional attributes into the same microservice cluster, thus obtaining several microservice clusters.

[0009] Preferably, the multidimensional attribute features include at least one of the following: real-time performance metrics, data processing latency, resource consumption overhead, and deployment success rate of each microservice instance.

[0010] Preferably, the establishment of cross-layer mapping constraints from the logical functional topology to the physical network topology specifically includes: Determine the logical functional scope and the physical scope; Construct a functor mapping from the logical functional category to the physical category, and use the functor mapping as the cross-layer mapping constraint.

[0011] Preferably, the functor mapping includes object mapping and morphic mapping; wherein, the object mapping is used to map the logical execution phase to physical nodes, and the morphic mapping is used to map the logical temporal dependencies between subtasks to link paths between physical nodes.

[0012] Preferably, the multi-agent deep reinforcement learning model includes a first agent and a second agent.

[0013] Preferably, based on the microservice cluster and the cross-layer mapping constraints, a pre-trained multi-agent deep reinforcement learning model is used to match microservice instances from the microservice cluster for each subtask, and the matched microservice instances are deployed to physical nodes of the physical network. Simultaneously, the link routes between the physical nodes are determined to obtain a multi-dimensional resource scheduling scheme, specifically including: Based on the performance requirements of the subtasks and the characteristic indicators of each microservice cluster after clustering, candidate microservice clusters that are suitable for the current subtasks are selected. A pre-trained first agent selects the corresponding microservice instance for each subtask from the candidate microservice cluster, and a pre-trained second agent allocates the matched microservice instance to the corresponding physical node based on cross-layer mapping constraints and physical node resource limitations. Based on the physical node deployment results, resource constraint pruning is performed on network links and link routing is planned. By integrating all decision results of microservice matching, node deployment, and link routing, a multi-dimensional resource scheduling scheme is obtained.

[0014] Preferably, the multi-agent deep reinforcement learning model is trained based on a hybrid reward function; wherein, the hybrid reward function includes a latency reward, a QoS satisfaction reward, and a load balancing reward.

[0015] The beneficial effects of this invention are: The method of this invention decomposes immersive communication services into time-series subtasks and matches them with corresponding microservices, which can transform complex services into standardized scheduling units. By clustering microservice features to reduce dimensionality, the decision space can be effectively compressed and scheduling efficiency improved. At the same time, a logical and physical two-layer topology is built and cross-layer mapping constraints are set to ensure that the physical deployment is consistent with the business logic topology and execution sequence, avoiding latency and logic problems caused by deployment misalignment.

[0016] The method of this invention completes the joint optimization of microservice matching, node deployment and link routing through a pre-trained multi-agent model, which can simultaneously meet the multi-dimensional requirements of latency, QoS, load balancing and other aspects, and can obtain a globally optimal scheduling scheme, thereby enhancing the service guarantee capability of 6G network for immersive services.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The physical network topology diagram of the present invention is shown; Figure 2 The normalized reward value graph of the present invention is shown; Figure 3 A graph showing the deployment success rate comparison (deployment algorithm) of the present invention is displayed; Figure 4 A load balancing diagram of the present invention is shown; Figure 5 The diagram shows a comparison of the deployment success rates of this invention (microservice matching algorithm); Figure 6 A comparison chart of the time delay constraint satisfaction rate of the present invention is shown; Figure 7 The hierarchical scheduling architecture diagram of the present invention is shown; Figure 8 The flowchart of the clustering and two-stage multi-agent algorithm of the present invention is shown; Figure 9 A flowchart of a 6G network multi-dimensional resource scheduling method for immersive communication services according to the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 9 As shown, a multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services specifically includes the following steps: Reference Figure 7 As shown, in S10, in response to the received immersive communication service request, the immersive communication service request is decoupled into multiple sub-tasks with temporal dependencies, and the required set of multi-sub-layer microservice instances is determined for each sub-task. S101. Analyze the service level protocol and QoS requirements of the immersive communication service request, decompose the immersive communication service request into multiple sub-tasks according to the functional sub-layer, and determine the resource requirement weight of each sub-task in the architecture of the functional sub-layer. Specifically, immersive communication services can be categorized into a set of services. ,in, Represents a collection of immersive communication services. Represents the set of elements. An immersive communication service. This indicates the total number of services within the service set.

[0022] The single immersive communication service currently being processed can be represented as ,in, This indicates the subtasks that need to be executed sequentially after the business process is broken down. This indicates the bandwidth requirements of the service. This indicates the amount of data the service provides on the source node. These represent the source node and destination node of the service, respectively. This is the maximum end-to-end latency that the service can tolerate.

[0023] All subtasks corresponding to this business can be organized as follows: ,in, This indicates the first [number] of the service. Sub-tasks This represents the total number of subtasks included in the business logic. All subtasks need to be executed sequentially according to the given order. Each subtask to be run contains various types of runtime information, such as execution dependencies, performance requirements, and metric weights, which can be represented as follows: ,in, This indicates the execution order of microservices in each sub-layer of this subtask. Represents the performance indicator requirement matrix. This indicates the weight of the performance indicators within each sublayer. This represents the overall weight of each functional sublayer.

[0024] The performance requirement matrix for the subtask is as follows In this context, the superscripts 1, 2, 3, and 4 correspond to the algorithm sublayer, data sublayer, computing power sublayer, and connection sublayer, respectively. These represent performance metrics related to algorithm resources, including rendering quality and tracking stability. These represent performance metrics related to data resources, including frame rate and data jitter. These represent performance metrics related to computing resources, including metrics such as computing accuracy. This indicates performance metrics related to connectivity resources, including metrics such as network jitter.

[0025] The performance indicators of a single sublayer can be expressed as follows: The complete requirements for each indicator within the sub-layer are provided by A common description, in which, Indicates the subtask number Sub-layer The required values ​​for each performance indicator This indicates the tolerance threshold for this performance indicator. This is the direction vector of the index.

[0026] The weighted combination of indicators within each sub-layer is as follows:

[0027] in, Representing the Sub-layer The weights of performance metrics, and the weights of all metrics within the same sub-layer satisfy the following: The overall weight of each functional sublayer is expressed as follows: ,in, Indicates the first The weights of each functional sublayer, and the weights of all sublayers satisfy the following: .

[0028] The execution order of different sub-layer microservices is controlled by execution priority, and the relevant parameters are expressed as follows: ,in, Indicates the first The execution priority of each sub-layer microservice is a positive integer. The higher the priority value, the higher the execution priority of the corresponding sub-layer microservice. If multiple sub-layer microservices have the same priority value, these sub-layer microservices will be executed in parallel on the same physical node.

[0029] S102. Based on the resource requirement weight and functional description of each subtask, select microservice instances that can provide the corresponding functions and meet the lower limit of resource requirements of each sublayer, and form a multi-sublayer microservice instance set.

[0030] Specifically, based on the operational requirements, performance indicators, and execution order rules of each subtask, the microservice resource pools corresponding to each functional sublayer are invoked sequentially. The set of all microservice instances in each sub-layer is ,in, Sub-layer The first in A microservice instance, This represents the total number of microservice instances in this sublayer. The runtime attributes of each microservice instance can be represented as follows: ,in, The performance metrics provided for this microservice instance This indicates the processing latency required to execute this microservice instance. Indicates the coefficient of variation in data volume. This indicates the demand for computing and storage resources. This indicates the computing resource requirements of the microservice instance. This indicates the storage resource requirements of the microservice instance.

[0031] Define a selection variable to indicate whether a subtask uses a microservice instance. ,when At that time, it represents a subtask. Select the The first sub-layer Each microservice instance. Simultaneously, it adheres to operational constraints. This ensures that only one microservice instance is selected within the same sublayer.

[0032] To determine the fit between microservice instances and subtask requirements, the normalized deviation of the metric is first calculated, and its expression is: ,in, This represents the normalized deviation between the subtask metric requirements and the selected microservice instance metrics. Indicates the subtask number Sub-layer The required value of the indicator, This indicates the actual value of the metric corresponding to the selected microservice instance.

[0033] The formula for calculating the matching degree of a single index based on normalized deviation is as follows: ,in, The matching degree of a single performance indicator, with a value range of [value range missing]. , It is a natural exponential function.

[0034] The overall fit level of a single sublayer is obtained by summarizing the indicator weights. The formula for calculating the performance utility of a sublayer is: ,in, Indicates the first The performance utility of a sublayer is used to measure the extent to which the microservices in that sublayer meet the requirements of the subtask. This indicates the weight of a single indicator within a sub-layer. This represents the matching degree of a single indicator. Then, combining the weights of each sub-layer, the overall matching effect of the sub-task is calculated. The overall performance utility formula is: ,in, This indicates the overall matching degree between the subtask and the microservices in each sublayer.

[0035] Finally, based on the calculated index normalization deviation, single index matching degree, single sub-layer performance utility, and sub-task comprehensive performance utility, and combined with the microservice execution priority rules and resource consumption, the microservice instances in each sub-layer are selected, and finally, microservice instances of the four sub-layers of algorithm, data, computing power, and connection are configured for each sub-task, forming a complete set of multi-sub-layer microservice instances.

[0036] Reference Figure 8 As shown in step S20, extract the multi-dimensional attribute features of each microservice instance in the multi-sub-layer microservice instance set corresponding to each subtask, and perform clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters. S201. Construct a multi-view hypergraph that integrates performance metrics, latency, and resource overhead views; wherein, the vertices of the multi-view hypergraph are each microservice instance, and the hyperedges are used to connect microservice instances with similar attributes under different views; Specifically, regarding the first Each sub-layer constructs a hypergraph with all microservice instances in that layer as vertices. The hypergraph expression is: ,in, This represents a set of vertices composed of microservice instances. Denotes the set of superedges. This represents the set of superedge weights. First, build the core performance metric view, then calculate the extreme values ​​of various metrics for all microservice instances within this sub-layer, defining the first... Sub-layer Maximum value of the item indicator and minimum value Combined with indicator direction vector The performance metrics are polarity aligned and normalized. When, the normalization calculation formula is: ,when hour, After normalization, based on the normalized vector, the k-nearest neighbor algorithm (kNN) is used to connect the vector of each node to its nearest Euclidean neighbors to form a hyperedge. Hyperedge weight It is determined by the Gaussian kernel similarity between nodes.

[0037] Building upon the performance metric view, a latency view is constructed. This view characterizes the time-dependent features of microservices, and the corresponding metric calculation formula is as follows: , where, define function , , These are the weight coefficients for the corresponding dimensions. This indicates the processing latency of a microservice instance. This represents the data volume variation coefficient of a microservice instance. Based on this metric, a corresponding hyperedge is constructed, and the hyperedge weight under the latency view is calculated. .

[0038] Further, a resource overhead view is constructed. This view is used to characterize the resource consumption level of microservices, and the corresponding indicator calculation formula is as follows: , , These are the weight coefficients for the corresponding dimensions. This represents the storage resource requirements of a microservice instance. Hyperedges are also constructed based on this metric, and their weights are calculated within the resource overhead view. .

[0039] Finally, the weights of the hyperedges in the three views are calculated together. The formula for the comprehensive weight of a single hyperedge is as follows: The weighting coefficients satisfy the following conditions: This means that the performance metrics view is the core dimension, and the latency view and resource overhead view are auxiliary dimensions. After integrating all vertices, hyperedges and comprehensive weights, a complete multi-view hypergraph can be obtained.

[0040] S202. Construct the adjacency matrix of the multi-view hypergraph and perform spectral clustering analysis on the adjacency matrix to aggregate microservice instances with similar multidimensional attributes into the same microservice cluster, thus obtaining several microservice clusters. During matrix construction, the association matrix is ​​constructed sequentially for the multi-view hypergraph. Calculate the node degree matrix and hypermarginality matrix Based on the three types of matrices, the normalized hypergraph Laplacian matrix is ​​derived, and its expression is: ,in, It is an identity matrix.

[0041] Perform eigenvalue decomposition on the obtained normalized hypergraph Laplacian matrix, and select the smallest eigenvalue from the decomposition results. A set of eigenvectors, which are combined to form the spectral embedding space. The K-means algorithm is used to cluster the data in the spectral embedding space, ultimately obtaining the [number of clusters]. A collection of microservice clusters in each sub-layer within the set This represents a single microservice cluster. To facilitate rapid matching of microservices in subsequent subtasks, the comprehensive performance index of each microservice cluster is calculated offline, and its expression is: ,in, This indicates the total number of microservice instances contained within the current microservice cluster. This represents the performance metrics of a single microservice instance within a cluster.

[0042] After spectral clustering, the originally large number of microservice instances were divided into multiple microservice clusters according to attribute similarity, achieving the goal of clustering dimensionality reduction and effectively reducing the decision space in the subsequent microservice matching process.

[0043] S30. Construct a logical functional topology based on the temporal dependencies of subtasks, and construct a physical network topology based on the physical network. Establish cross-layer mapping constraints from the logical functional topology to the physical network topology. The cross-layer mapping constraints are used to ensure the topological consistency of the logical execution order of subtasks during the physical deployment process. Based on the microservice clusters obtained by clustering and dimensionality reduction, a complete logical functional topology is built by combining the execution sequence dependencies between subtasks and between sub-layer microservices within subtasks. At the same time, a physical network topology is built based on the actual network architecture composed of terminals, edge nodes, cloud control centers, and heterogeneous communication links. On this basis, cross-layer mapping constraints between the two types of topologies are established. According to the constraints, the original logical execution order of subtasks and the topological connectivity structure are guaranteed to remain unchanged in the physical deployment scenario. At the same time, the quantitative calculation of business latency and network load balancing is completed, and a multi-objective joint optimization problem is constructed by combining various resource boundary conditions.

[0044] S301. Define logical functional categories and physical categories, and construct functor mappings from logical functional categories to physical categories; wherein, functor mappings include object mappings and morphic mappings. Object mappings are used to map logical execution phases to physical nodes, and morphic mappings are used to map logical temporal dependencies between subtasks to link paths between physical nodes.

[0045] For immersive communication services First, construct the corresponding service scope. The set of objects in this category is This service consists of all logical execution phases and virtual source nodes. and destination node Composition, category of morphism This indicates the logical execution order (existing only between objects in adjacent phases), determined by the execution order within and between subtasks. For any object within the category... All of them are capable of handling latency Computing resource requirements Storage resource requirements Coefficient of change in data volume Four properties; for any state, Its attribute is link bandwidth requirement. And the value is equal to the bandwidth requirement of the corresponding service. ,Right now Since the multiple services across the network are independent of each other, the overall logical functional scope is the product of the scopes of all individual services, and its expression is: The corresponding set of objects satisfies The set of states satisfies .

[0046] Furthermore, the actual physical network is modeled as an undirected graph. And define the physical domain based on this network. The collection of objects in this category For all physical nodes, i.e. Each physical node All have computing resource capacity , Two attributes; each physical link With bandwidth capacity Propagation delay Two attributes. In the physical domain, a morphism is defined as an effective transmission path between nodes. If two physical nodes are different nodes, the morphism is composed of multiple physical links in sequence; if they are the same physical node, an identical zero-length path is defined as the morphism, corresponding to the scenario where the logical stages are deployed on the same node.

[0047] Constructing functor mappings between logical functional categories and physical categories. This mapping consists of two parts: object mapping and morphic mapping. Object mapping is responsible for mapping objects (logical subtasks, execution phases, virtual nodes) within the logical functional scope to physical nodes in the physical scope. There are fixed mapping constraints between virtual source nodes and virtual destination nodes, with constraint C2 being... Constraint C3 is ,in, , These are the source and destination physical nodes preset by the business logic. The morphism mapping is responsible for mapping the morphisms representing time-series dependencies within the logical scope to the morphisms representing data transmission paths within the physical scope. The start and end points of the physical path must be consistent with the physical nodes deployed on the corresponding logical objects. At the same time, the connectivity between subtasks deployed on the physical nodes must be consistent with the logical topology. The number of hops in the physical link path and the number of steps in the logical execution order follow preset constraints.

[0048] Furthermore, to ensure that physical network resources do not exceed limits, several hard resource constraints are set, specifically: Constraint C4 is This means that the total computing resource requirements of all logical stages deployed on the same physical node do not exceed the computing capacity of that node.

[0049] Constraint C5 is This means that the total storage resource requirements of all logical stages on the same physical node do not exceed the storage capacity of that node.

[0050] Constraint C6 is This means that the total bandwidth of all logical flows carried by a single physical link does not exceed the bandwidth capacity of the link itself.

[0051] The end-to-end latency of a service is broken down and calculated based on the functor mapping results. The total service latency consists of three parts: processing latency, transmission latency, and propagation latency. Each service... The expression for the total processing latency is: The processing latency of the virtual source node and the virtual destination node is set to 0. Define any logical object. The pre-execution phase is The amount of output data corresponding to the virtual source node The output data volume of the remaining logical objects meets the requirements. The transmission latency of a single logical stage is The total service transmission delay expression is as follows: Physical path The propagation delay is the sum of the propagation delays of all links within the path, i.e. The total service propagation delay expression is: .

[0052] Therefore, the total server-to-end latency is At the same time, set delay constraints This ensures that the runtime of the service does not exceed the maximum tolerable latency.

[0053] The resource load variance is used to quantify the overall network load balancing level, physical nodes. The cumulative computing load is The cumulative storage load is The node computing resource utilization rate is Storage resource utilization rate Introducing weighting coefficients and ,and The overall load of the nodes can be obtained. .

[0054] physical link The cumulative bandwidth load is Link resource utilization rate The average load across all network nodes is The node load variance is The average load of the entire network links is The link load variance is The final global load balancing function is .

[0055] Finally, combining microservice selection variables Using the deployment functor and the constraints C1 to C7 mentioned above, we construct a multi-objective joint optimization problem. The expression for the objective function is as follows: The optimization objectives are to maximize the overall matching utility of microservices, minimize end-to-end business latency, and minimize global load variance. The entire optimization problem needs to satisfy constraints C1 to C7 simultaneously, which clarifies the optimization direction and boundary conditions for subsequent intelligent decision-making.

[0056] S40. Based on microservice clusters and cross-layer mapping constraints, a pre-trained multi-agent deep reinforcement learning model is used to match microservice instances from the microservice clusters for each subtask, and the matched microservice instances are deployed to physical nodes of the physical network. At the same time, the link routes between physical nodes are determined to obtain a multi-dimensional resource scheduling scheme. S401. Based on the performance requirements of the subtask and combined with the characteristic indicators of each microservice cluster after clustering, select candidate microservice clusters that are suitable for the current subtask. Before making decisions for the intelligent agent, the performance requirements vector of the subtasks should be considered. Comprehensive performance metrics of microservice clusters obtained from offline computation The optimal microservice cluster is selected by Euclidean distance, and its expression is: This serves as a subtask to identify four candidate microservice clusters corresponding to the sub-layers, thus narrowing down the subsequent selection range.

[0057] Furthermore, the multi-agent deep reinforcement learning model includes a first agent and a second agent.

[0058] S402. Using a pre-trained first agent, select the corresponding microservice instance for each subtask from the candidate microservice cluster, and using a pre-trained second agent, allocate the matched microservice instance to the corresponding physical node according to the cross-layer mapping constraints and physical node resource limitations. Specifically, a two-stage multi-agent deep reinforcement learning architecture is adopted, that is, for each service request... Each subtask A pair of intelligent agents is dynamically generated—a first intelligent agent (microservice matching agent) and a second intelligent agent (microservice deployment agent). The first intelligent agent is responsible for selecting sub-tasks from candidate microservice clusters. The four logical sub-layers filter specific microservice instances; the second intelligent agent is responsible for mapping the matched microservice instances to physical nodes and planning link routes.

[0059] It should be further noted that when the number of concurrent requests is At that time, there exists Individual agents ( The first intelligent agent and (A second agent), where all agents share training parameters but make independent decisions.

[0060] For the first intelligent agent, the three elements of deep reinforcement learning are: State space: describes the service The current task's performance requirements and the characteristics of the candidate microservice clusters are defined as follows: ,in, This is the performance requirement vector for the subtask. , This is the feature matrix of the candidate microservice cluster.

[0061] Action space: defined as a multidimensional discrete space This is used to select a specific instance for each of the four sub-layers at once. A validity mask is applied to the policy network output layer to set the probability of invalid actions pointing to the virtual instances to zero, ensuring the legality of the decision.

[0062] Reward Mechanism: Combining the overall matching utility of microservices to measure QoS satisfaction, its single-step internal reward formula is as follows: ,in, The overall matching utility of the microservice instance selected for the current subtask (calculated by weighting QoS metrics such as response time, reliability, and throughput). As a utility weighting coefficient, this reward directly reflects the degree to which the matching result fits the performance requirements of the subtask.

[0063] For the second intelligent agent, the three elements of deep reinforcement learning are: State space: Describes the global state of the physical network and the current microservice deployment context, defined as: ,in, To utilize the physical network topology-aware features extracted by GCN, These are the source node and destination node of the service, respectively. For bandwidth requirements, These are the attributes of the microservice execution phase (processing latency, data volume transformation coefficient, and resource requirements). For the context of this execution phase, The physical nodes deployed in the previous microservice execution phase. These represent the total computing and storage resources required by subsequent nodes, and the actions. For the physical network node set Select a deployment node.

[0064] Action space: defined as the set of physical network nodes Select a node as the deployment location for the current microservice instance, i.e. .

[0065] Reward Mechanism: Centered on resource constraint compliance, the single-step internal reward formula is as follows:

[0066] in, Rewards for successfully placing microservices, Penalties for violating resource constraints.

[0067] Furthermore, single-step internal rewards only measure the local rationality of the current decision step. To achieve multi-objective global optimization such as minimizing end-to-end latency and balancing network load, it is necessary to further integrate joint rewards and latency rewards.

[0068] Specifically, the global joint reward covers three optimization objectives: microservice matching utility, end-to-end total latency, and network load balancing. The expression is:

[0069] in, For service The sum of the microservice matching utility of all subtasks. For the current deployment plan The end-to-end total delay is below. This is a load balancing metric for physical nodes. , , These are the weighting coefficients for each objective.

[0070] Latency and Link Constraint Rewards: When the service After all microservices have been deployed and the connections have been established, calculate the global latency bonus:

[0071] in, This is a penalty for latency or link violations.

[0072] Total agent reward: The total reward for a single step is obtained by fusing the local internal rewards and global rewards of the two agents.

[0073]

[0074] in, , , These are the weighting coefficients for joint rewards and time-delay rewards, respectively, to achieve a balance between local decision-making and global optimization.

[0075] During the training phase, the Proximal Policy Optimization (PPO) algorithm combined with General Advantage Estimation (GAE) is used to achieve collaborative iterative optimization among multiple agents. First, the temporal difference error is calculated. Then derive the advantage function Through the policy probability ratio To control the policy update magnitude, the Actor network truncation loss function is: Avoid sudden policy changes.

[0076] Furthermore, a policy entropy regularization term is introduced. To avoid the model getting trapped in local optima, the final overall loss function is:

[0077] After multiple iterations until the model converges, training is complete, and a pre-trained multi-agent model is obtained.

[0078] During the inference phase, parameter updates are no longer performed; the pre-trained model is loaded directly. The first agent completes the microservice instance selection and matching within the candidate microservice cluster. The second agent, combining the logical-physical cross-layer mapping constraints established earlier and the physical node computing and storage resource capacity limitations, completes the allocation and deployment of microservice instances to physical nodes.

[0079] S403. Based on the physical node deployment results, perform resource constraint pruning on network links and complete link routing planning. Integrate all decision results of microservice matching, node deployment, and link routing to output the final multi-dimensional resource scheduling scheme.

[0080] After the physical nodes are deployed, a hard-constraint pruning operation is first performed on all physical links in the network to remove invalid links whose remaining bandwidth cannot support the service demand, and a new auxiliary topology is constructed. The routing cost is calculated based on the effective links, using the following formula: ,in, This represents the remaining bandwidth of the link. For propagation delay weighting coefficient, To account for link propagation delay, the optimal physical transmission path is selected based on this cost. After link planning is completed, the microservice instance selection results, physical node allocation results, and inter-node link routing results are summarized and integrated to form a complete multi-dimensional resource scheduling scheme, which is then output externally to complete the entire resource scheduling process for the immersive communication service.

[0081] The solution method of this invention can gradually converge during training and maintain a high service success rate. By reducing the microservice matching space through multi-view hypergraph clustering, ensuring structural consistency between the logical functional layer and the physical layer through category theory mapping, and co-optimizing microservice selection and physical node deployment through a two-stage multi-agent deep reinforcement learning model, it can improve the resource scheduling efficiency, service success rate, and network load balancing performance of immersive communication services.

[0082] In a specific embodiment, to verify the actual operating performance of the 6G multi-dimensional resource scheduling scheme for immersive communication services proposed in this invention, this embodiment builds a cloud-edge-device collaborative computing simulation environment, completes full-dimensional parameter configuration in combination with the service characteristics of immersive applications, and designs multiple sets of comparative experiments to conduct tests from multiple dimensions such as algorithm convergence efficiency, high concurrency robustness, network load balancing, and latency guarantee capability, so as to comprehensively evaluate the overall performance and application value of the scheme.

[0083] Simulation parameter settings: The simulation is based on a heterogeneous cloud-edge-device network architecture for modeling. To realistically reproduce the operating characteristics of 6G networks, parameterized configurations are performed for the computing power and storage capacity of nodes at the device, edge, and cloud layers, as well as the bandwidth and propagation latency of different types of network links. Simultaneously, considering the operational requirements of immersive services, fine-tuned settings are achieved for various performance indicators of service requests, subtasks, and microservice instances.

[0084] Table 1 Simulation Parameter List

[0085] The table above summarizes all parameters and their corresponding value ranges for this simulation, covering both network hardware and business applications. Hardware parameters differentiate between the computing and storage resources of end, edge, and cloud nodes, and set bandwidth value ranges according to the link types of end-to-edge, edge-to-edge, and edge-to-cloud. Business and microservice parameters include the number of subtasks, multi-level latency requirements, bandwidth requirements, source node data volume, and QoS indicators for the four sub-layers of algorithm, data, computing power, and connection. It also defines the performance parameters, resource overhead, processing latency, and data volume variation coefficients of microservices, comprehensively constructing a simulation scenario that closely matches real-world applications.

[0086] Algorithm Result Comparison and Analysis: All control experiments were executed within a unified network topology, node resource configuration, and service request sequence to ensure the fairness of the experimental results. The physical network in this simulation is an end-edge-cloud architecture with 24 nodes, comprising 10 end-side nodes, 10 edge nodes, and 4 cloud-side nodes. Nodes of each type are deployed in a regionally distributed manner. The physical network topology is shown in the attached diagram. Figure 1 As shown.

[0087] First, ablation experiments were conducted to verify the performance improvement effects of the multi-view hypergraph clustering mechanism and the two-stage collaborative decision-making architecture on the algorithm. Normalized reward convergence curves for different algorithms were collected during the experiments. The normalized reward variation curves are shown in the attached figure. Figure 2 As shown in the figure, three comparison schemes were set up in the experiment: the multi-view hypergraph clustering + two-stage multi-agent deep reinforcement learning scheme used in this scheme, the clustering combined with a single-stage decision algorithm, and the two-stage algorithm without clustering. The experimental results show that the proposed scheme can converge quickly in only about 3000 iterations and maintain a high reward value; the algorithm without hypergraph clustering suffers from an explosion in the scheduling action space, resulting in low agent exploration efficiency and violent oscillations in the curve; although the single-stage decision algorithm has a reasonable learning speed in the early stage, the decision load is too large due to the synchronous processing of matching and deployment tasks, and the final convergence performance is not as good as the proposed scheme, which fully verifies the technical advantages of the dimensionality reduction mechanism and the two-stage decoupled decision.

[0088] To test the robustness of the algorithm in a 6G high-concurrency scenario, this experiment introduced an average request intensity parameter to simulate different service loads, with the number of service requests modeled according to a Poisson distribution. First, deployment success rate was used as the core metric, comparing two mainstream scheduling schemes, DTS-SFC and Kolin. The deployment success rates under different request intensities are shown in the appendix. Figure 3 As shown, in the low request intensity range, the deployment success rate of all three solutions is close to 100%. As the business load continues to increase, the deployment success rate of the two comparison algorithms drops significantly, while this solution consistently maintains a higher deployment success rate. The higher the load, the more prominent the performance advantage becomes.

[0089] Based on this, the load balancing variance is used to evaluate the resource allocation status of the entire network. The smaller the variance value, the more balanced the resource utilization of nodes and links. The load balancing variance comparison results are shown in the appendix. Figure 4 The experimental data shows that, across the entire load range, the load balancing variance of this solution is consistently lower and fluctuates less, effectively preventing resources from being concentrated on a few nodes and links, reducing the risk of network congestion, and improving the overall stability of the system.

[0090] Subsequently, comparative tests were conducted on different microservice matching strategies. Three classic heuristic strategies were added: Min-ResourceMatching, Min-DelayMatching, and Max-UtilityMatching. The deployment success rates of different matching strategies are shown in the appendix. Figure 5 As shown, the performance differences between various strategies are small under low load scenarios; however, under high load conditions, the Min-ResourceMatching strategy, which prioritizes low resource consumption, performs best. This solution can adaptively balance resource consumption and service performance, resulting in a more gradual decline in deployment success rate.

[0091] Finally, a latency constraint satisfaction rate metric is introduced to evaluate the ability of each solution to guarantee the latency requirements of immersive services. The comparison results of latency constraint satisfaction rates are shown in the appendix. Figure 6 As shown, the latency satisfaction rate of all schemes decreased with increasing request intensity, but this scheme maintained the optimal level throughout. Min-DelayMatching is prone to forming network hotspots under high load, resulting in a significant degradation in latency performance, and other heuristic strategies also suffer from insufficient latency optimization. Based on multiple simulation results, the scheme of this invention can balance multiple objectives such as deployment success rate, load balancing, and latency constraints, and possesses excellent robustness, scalability, and practical value in 6G large-scale immersive communication service scenarios.

[0092] Based on the same inventive concept as the above method, this invention also proposes a 6G network multi-dimensional resource scheduling system for immersive communication services, comprising: The business request decoupling module is used to respond to the received immersive communication business request, decouple the immersive communication business request into multiple sub-tasks with temporal dependencies, and determine the required set of multi-sub-layer microservice instances for each sub-task. The microservice clustering and dimensionality reduction module is used to extract the multi-dimensional attribute features of each microservice instance in the multi-sub-layer microservice instance set corresponding to each subtask, and perform clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters. A cross-layer mapping constraint construction module is used to construct a logical functional topology based on the temporal dependencies of subtasks, and to construct a physical network topology based on the physical network, establishing cross-layer mapping constraints from the logical functional topology to the physical network topology; wherein, the cross-layer mapping constraints are used to ensure the topological consistency of the logical execution order of the subtasks during the physical deployment process. The multi-agent collaborative scheduling module is used to match microservice instances from the microservice cluster for each subtask based on the microservice cluster and the cross-layer mapping constraints, and deploy the matched microservice instances to physical nodes of the physical network. At the same time, it determines the link routes between the physical nodes to obtain a multi-dimensional resource scheduling scheme.

[0093] Based on the same inventive concept as the above method, the present invention also proposes a device including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and the processor executes the above-described 6G network multi-dimensional resource scheduling method when executing the computer instructions.

[0094] Based on the same inventive concept as the above method, the present invention also proposes a computer-readable storage medium storing computer instructions thereon, which, when executed, can realize the above-mentioned multi-dimensional resource scheduling method for immersive communication services in 6G networks.

[0095] Any references to memory, storage, database, or other media used in the embodiments provided in this invention may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services, characterized in that, include: In response to a received immersive communication service request, the immersive communication service request is decoupled into multiple sub-tasks with temporal dependencies, and a set of multi-layer microservice instances required for each sub-task is determined. Extract the multi-dimensional attribute features of each microservice instance in the multi-sub-layer microservice instance set corresponding to each subtask, and perform clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters; A logical functional topology is constructed based on the temporal dependencies of subtasks, and a physical network topology is constructed based on the physical network. A cross-layer mapping constraint is established from the logical functional topology to the physical network topology. The cross-layer mapping constraint is used to ensure the topological consistency of the logical execution order of the subtasks during the physical deployment process. Based on the microservice cluster and the cross-layer mapping constraints, a pre-trained multi-agent deep reinforcement learning model is used to match microservice instances from the microservice cluster for each subtask, and deploy the matched microservice instances to physical nodes of the physical network. At the same time, the link routes between the physical nodes are determined to obtain a multi-dimensional resource scheduling scheme.

2. The 6G network multi-dimensional resource scheduling method for immersive communication services according to claim 1, characterized in that, The step of decoupling the immersive communication service request into multiple sub-tasks with temporal dependencies, and determining the required set of multi-layered microservice instances for each sub-task, specifically includes: The service level protocol and QoS requirements of the immersive communication service request are parsed, and the immersive communication service request is decomposed into multiple sub-tasks according to the functional sub-layers. The resource requirement weight of each sub-task in each sub-layer of the functional sub-layer architecture is determined. The functional sub-layers include an algorithm sub-layer, a data sub-layer, a computing power sub-layer, and a connection sub-layer. Based on the resource requirement weight and functional description of each subtask, microservice instances are selected to form a multi-layered microservice instance set.

3. The 6G network multi-dimensional resource scheduling method for immersive communication services according to claim 2, characterized in that, The QoS requirements include: rendering quality and tracking stability of the algorithm sub-layer, frame rate and data jitter of the data sub-layer, computing power accuracy of the computing power sub-layer, and network jitter of the connection sub-layer.

4. The 6G network multi-dimensional resource scheduling method for immersive communication services according to claim 1, characterized in that, The step of extracting multi-dimensional attribute features of each microservice instance in the multi-layer microservice instance set corresponding to each subtask, and performing clustering and dimensionality reduction on the microservice instances based on the multi-dimensional attribute features to generate several microservice clusters, specifically includes: Construct a multi-view hypergraph that integrates performance metrics, latency, and resource overhead views; wherein the vertices of the multi-view hypergraph are each microservice instance, and the hyperedges are used to connect microservice instances with similar attributes under different views; Construct the adjacency matrix of the multi-view hypergraph, and perform spectral clustering analysis on the adjacency matrix to aggregate microservice instances with similar multidimensional attributes into the same microservice cluster, thus obtaining several microservice clusters.

5. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 4, characterized in that, The multidimensional attribute features include at least one of the following: real-time performance metrics, data processing latency, resource consumption overhead, and deployment success rate of each microservice instance.

6. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 1, characterized in that, The establishment of cross-layer mapping constraints from the logical functional topology to the physical network topology specifically includes: Determine the logical functional scope and the physical scope; Construct a functor mapping from the logical functional category to the physical category, and use the functor mapping as the cross-layer mapping constraint.

7. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 6, characterized in that, The functor mapping includes object mapping and morphic mapping; wherein, the object mapping is used to map the logical execution phase to physical nodes, and the morphic mapping is used to map the logical temporal dependencies between subtasks to link paths between physical nodes.

8. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 1, characterized in that, The multi-agent deep reinforcement learning model includes a first agent and a second agent.

9. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 8, characterized in that, Based on the microservice cluster and the cross-layer mapping constraints, a pre-trained multi-agent deep reinforcement learning model matches microservice instances from the microservice cluster for each subtask, deploys the matched microservice instances to physical nodes in the physical network, and determines the link routes between the physical nodes to obtain a multi-dimensional resource scheduling scheme, specifically including: Based on the performance requirements of the subtasks and the characteristic indicators of each microservice cluster after clustering, candidate microservice clusters that are suitable for the current subtasks are selected. A pre-trained first agent selects the corresponding microservice instance for each subtask from the candidate microservice cluster, and a pre-trained second agent allocates the matched microservice instance to the corresponding physical node based on cross-layer mapping constraints and physical node resource limitations. Based on the physical node deployment results, resource constraint pruning is performed on network links and link routing is planned. By integrating all decision results of microservice matching, node deployment, and link routing, a multi-dimensional resource scheduling scheme is obtained.

10. A multi-dimensional resource scheduling method for 6G networks oriented towards immersive communication services according to claim 8, characterized in that, The multi-agent deep reinforcement learning model is trained based on a hybrid reward function; wherein, the hybrid reward function includes a latency reward, a QoS satisfaction reward, and a load balancing reward.

Images