Computing network resource scheduling method and device
By performing spatiotemporal sequence prediction and multi-agent resource scheduling in the cloud edge network, the resource scheduling accuracy problem caused by changes in the resource state of terminal equipment in the mobile network is solved, and the full utilization of resources and the reduction of task delay are achieved.
Patent Information
- Application Number
- CN202410970236.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-19
AI Technical Summary
In a mobile network, dynamic changes in the resource state of the terminal device lead to a reduction in the accuracy of traditional resource scheduling methods, and incorrect task unloading can easily lead to resource waste, affecting the normal completion of computing tasks.
By obtaining the clustering results of the end-side nodes, real-time resource information data and real-time task information data of the cloud edge network, the spatiotemporal sequence prediction is carried out, the end-side resource information data is predicted, and the tasks are performed using the dual deep Q network MADDQN algorithm of multiple agents. resource scheduling to achieve the target resource scheduling results.
It improves the accuracy of resource scheduling, realizes full utilization of resources, and reduces the delay in task completion.
Smart Images

Figure CN118524001B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communication technology, and in particular to a computing network resource scheduling method and device. Background Art
[0002] The cloud-edge-end architecture is a three-layer framework consisting of a cloud data center (i.e., cloud), edge nodes, and terminal devices. In a mobile network, the resource information of the cloud and edge nodes is usually known, but the end-side devices (i.e., terminal devices) are in motion, and the state of resources is constantly changing. They may face the problem of difficult task offloading caused by the uncertainty of resource states. Therefore, how to reasonably schedule and offload tasks for mobile end-side devices is a difficult problem. Among them, task offloading refers to transferring some or all computing tasks from terminal devices to more powerful computing resources, such as cloud servers or edge servers. This process aims to reduce the computing burden of devices or terminals and improve the performance of the overall system.
[0003] In order to solve the difficulty of scheduling resources on the mobile side, existing solutions are to model the resource allocation problem as an optimization problem, directly solve the optimization problem through reinforcement learning algorithm or optimization algorithm, and issue the best scheduling strategy. However, in the scenario of mobile network, the temporal and spatial variation characteristics of network topology and computing network resources will affect the accuracy of traditional resource scheduling methods. Incorrect task offloading can easily lead to resource waste and affect the normal completion of computing tasks. Summary of the invention
[0004] The purpose of the present invention is to provide a computing network resource scheduling method and device to improve accuracy and achieve full utilization of resources.
[0005] In a first aspect, an embodiment of the present invention provides a computing network resource scheduling method, which is applied to a cloud-edge-end network composed of a cloud node, an edge node, and an end-side node, wherein the resource state of the end-side node changes dynamically; the computing network resource scheduling method includes:
[0006] Obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes current cloud resource information data, current edge resource information data and actual end-side resource information data at the previous moment;
[0007] According to the end-side node clustering result and the actual end-side resource information data, a spatiotemporal sequence prediction is performed on the resource information of each of the end-side nodes to obtain the predicted end-side resource information data at the current moment;
[0008] The resource scheduling of the task is performed according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result.
[0009] Further, the obtaining of the end-side node clustering result of the cloud-edge-end network includes:
[0010] Acquire communication delay information of the cloud-edge-end network within a preset time range, wherein the communication delay information includes first delay sequence information between the end-side node and the edge node and second delay sequence information between two end-side nodes;
[0011] Each of the end-side nodes is clustered based on the communication delay information to obtain the end-side node clustering result.
[0012] Further, clustering each of the end-side nodes based on the communication delay information to obtain the end-side node clustering result includes:
[0013] Based on the first delay sequence information, taking a node among the end-side nodes whose delay relative to the edge node is unchanged as an initial cluster head;
[0014] Based on the first delay sequence information and the second delay sequence information, a preset number of nodes that meet a first preset condition among the end-side nodes are used as cluster heads; wherein the first preset condition includes that the delay relative to the edge node is unchanged and the delay between the node and the initial cluster head is the largest;
[0015] Based on the second delay sequence information, the unclustered nodes in each of the end-side nodes are added to the corresponding cluster group to obtain the end-side node clustering result; wherein the delay between the node and the cluster head of the corresponding cluster group is the smallest.
[0016] Further, performing spatiotemporal sequence prediction on the resource information of each of the end-side nodes according to the end-side node clustering result and the actual end-side resource information data to obtain the predicted end-side resource information data at the current moment includes:
[0017] Acquire actual cluster resource information data corresponding to each cluster according to the end-side node clustering result and the actual end-side resource information data;
[0018] The actual cluster resource information data corresponding to each of the clusters is input into a preset spatiotemporal prediction model to obtain the predicted end-side resource information data output by the spatiotemporal prediction model; wherein the spatiotemporal prediction model is obtained by performing spatial feature learning and temporal feature learning on the historical end-side resource information data, and the predicted end-side resource information data includes the resource information data of each node in each of the clusters.
[0019] Furthermore, the spatiotemporal prediction model includes a two-layer graph convolutional network, a two-layer long short-term memory network, a random inactivation layer and a fully connected layer connected in sequence.
[0020] Further, performing resource scheduling of tasks according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result includes:
[0021] Determine cluster resource information data according to the predicted end-side resource information data; wherein the cluster resource information data includes cluster resource information of each cluster, and the cluster resource information of the cluster is the sum of resource information data of each node in the cluster;
[0022] Each of the clusters is taken as a cluster node, and the cluster resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data are taken as environmental state inputs. Real-time task unloading is performed in each of the cluster nodes, each of the cloud nodes and each of the edge nodes through the multi-agent dual deep Q network MADDQN algorithm to obtain the target resource scheduling result.
[0023] Furthermore, the actual end-side resource information data includes the available amount of computing resources, the remaining storage capacity and the communication delay information between the node and the cluster head; the real-time task information data includes the data volume of the task, the task's requirement for the type of computing resources, the task's minimum limit on the amount of computing resources and the task's maximum completion deadline.
[0024] In a second aspect, an embodiment of the present invention further provides a computing network resource scheduling device, which is applied to a cloud-edge-end network composed of a cloud node, an edge node, and an end-side node, and the resource state of the end-side node changes dynamically; the computing network resource scheduling device includes:
[0025] A data acquisition module, used to obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes current cloud resource information data, current edge resource information data and actual end-side resource information data at the previous moment;
[0026] A spatiotemporal prediction module, configured to perform spatiotemporal sequence prediction on the resource information of each of the end-side nodes according to the end-side node clustering result and the actual end-side resource information data, to obtain the predicted end-side resource information data at the current moment;
[0027] The resource scheduling module is used to perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result.
[0028] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and when the processor executes the computer program, the computing network resource scheduling method described in the first aspect is implemented.
[0029] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the computing network resource scheduling method described in the first aspect is executed.
[0030] The computing network resource scheduling method and device provided by the embodiment of the present invention are applied to a cloud-edge-end network composed of cloud nodes, edge nodes and end-side nodes, and the resource status of the end-side nodes is dynamically changing; when performing computing network resource scheduling, first obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes the current cloud resource information data, the current edge resource information data and the actual end-side resource information data at the previous moment; then, according to the end-side node clustering results and the actual end-side resource information data, perform spatiotemporal sequence prediction on the resource information of each end-side node to obtain the predicted end-side resource information data at the current moment; then, perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain the target resource scheduling result. In this way, through spatiotemporal sequence prediction, it is possible to cope with the situation where the resource status of terminal devices in the mobile network changes in real time and is difficult to obtain, and the predicted data can assist in the decision-making of resource scheduling, improve the accuracy of resource scheduling, and achieve full utilization of resources. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0032] Figure 1 It is a cloud-edge-device collaboration framework;
[0033] Figure 2 A flowchart of a computing network resource scheduling method provided by an embodiment of the present invention;
[0034] Figure 3 A cloud-edge-device collaboration framework based on clustering is provided for an embodiment of the present invention;
[0035] Figure 4 A schematic diagram of a GCN-LSTM algorithm framework provided by an embodiment of the present invention;
[0036] Figure 5 A schematic diagram of the structure of an LSTM module provided in an embodiment of the present invention;
[0037] Figure 6 A schematic diagram of the principle of a computing network resource scheduling method provided by an embodiment of the present invention;
[0038] Figure 7 The experimental results of a computing network resource scheduling method provided by an embodiment of the present invention;
[0039] Figure 8 A schematic diagram for comparing test results of using a spatiotemporal prediction module and not using a spatiotemporal prediction module provided by an embodiment of the present invention;
[0040] Fig. 9 A schematic diagram of the structure of a computing network resource scheduling device provided by an embodiment of the present invention;
[0041] Fig.10 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0042] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0043] 1) Cloud-edge-device collaboration framework
[0044] Cloud computing is a technology that provides computing resources and services through the Internet. Based on virtualization technology, it abstracts and aggregates computing, storage and network resources to form a unified cloud platform. Cloud computing can provide on-demand, elastic and scalable computing capabilities, and users can flexibly use and adjust resources according to their needs.
[0045] Edge computing is a distributed computing model that pushes computing and data processing to the edge of the network, close to data sources and terminal devices, to provide faster, low-latency computing and data processing capabilities. The goal of edge computing is to place computing resources and services as close to data sources as possible, reduce data transmission and processing delays, and improve system response speed and efficiency.
[0046] like Figure 1 As shown in the figure, the cloud-edge-end collaboration framework is a three-layer framework consisting of cloud data centers (i.e., cloud), edge nodes (i.e., edge devices), and terminal devices (i.e., terminals). The cloud and edge nodes are usually located in fixed locations on the network, but a large number of terminal devices may be in a mobile state. These terminals in a mobile state also have certain computing, storage, and network transmission resource capabilities. Effective use of these resources can achieve low-latency processing for services initiated by the terminals and improve user experience.
[0047] 2) Clustering technology
[0048] Clustering is the process of dividing the nodes in the network into different clusters or groups. Each cluster usually consists of a cluster head and multiple cluster members. The cluster head is responsible for coordinating and managing the communication and data transmission within the cluster. Clustering divides the network nodes into several clusters, so that the similarity of the nodes in each cluster is as large as possible, and the differences between clusters are as large as possible. Clustering can make the network dynamically adapt to the dynamic changes of the topology structure and improve the reliability and scalability of the network. Therefore, clustering technology is suitable for use in mobile networks.
[0049] 3) Multivariate Time Series Forecasting
[0050] Multivariate time series (MTS) forecasting refers to the task of forecasting and analyzing data with multiple variables and time dimensions. This kind of forecasting task is common in many fields, such as traffic flow forecasting, weather forecasting, financial market forecasting, etc. By forecasting multivariate time series, it can help understand and predict future trends, developments, and changes. In multivariate time series forecasting, it is necessary to consider the relationship between multiple variables and the changes in the time dimension. The challenge of this forecasting task is how to effectively capture the dependencies between variables and the evolution of time. If the data to be predicted has both time attributes and spatial attributes, it is necessary to learn the time dependency and spatial dependency separately.
[0051] 4) Spatiotemporal Graph Neural Network
[0052] Graph Neural Network is a neural network structure used to process graph data, which can effectively process non-Euclidean spatial data. Unlike traditional deep learning methods, graph data is composed of nodes and edges. Each node can be represented as a vector or feature, and the edge represents the relationship between nodes. Graph Neural Network analyzes and processes graph data by learning the relationship and features between nodes.
[0053] A spatiotemporal graph is a variable graph structure used to represent the relationship between different nodes within a certain spatiotemporal range. A spatiotemporal graph can be a directed graph, an undirected graph, a weighted graph, or an unweighted graph. Spatial-temporal graph neural networks (STGNNs) are used to process spatiotemporal graph data. They take into account the changes in nodes and edges in time and space and can be used to predict and analyze spatiotemporal graph data. The development of spatiotemporal graph neural networks provides an effective method for processing spatiotemporal data, which can extract complex spatiotemporal dependencies and achieve various predictive learning tasks. STGNNs have a wide range of applications in transportation, safety, environment, and public health.
[0054] At present, in order to solve the difficulty of scheduling terminal resources in mobility, the existing solutions are to model the resource allocation problem as an optimization problem, directly solve the optimization problem through reinforcement learning algorithm or optimization algorithm, and issue the best scheduling strategy. However, in the scenario of mobility network, the network topology and the spatiotemporal variation characteristics of computing network resources will affect the accuracy of traditional resource scheduling methods. Based on this, the embodiment of the present invention provides a computing network resource scheduling method and device for mobility network, which can improve the accuracy of resource scheduling through spatiotemporal graph learning and realize full utilization of resources.
[0055] To facilitate understanding of this embodiment, a computing network resource scheduling method disclosed in an embodiment of the present invention is first introduced in detail.
[0056] The embodiment of the present invention provides a computing network resource scheduling method, which is applied to a cloud-edge-end network composed of cloud nodes, edge nodes and end-side nodes. The resource status of the end-side nodes is dynamically changing, that is, the end-side nodes are in a mobile state; the computing network resource scheduling method can be executed by an electronic device with data processing capabilities, such as a scheduling device of a cloud-edge-end network. Figure 2 The flowchart of a computing network resource scheduling method shown in FIG. 1 mainly includes the following steps:
[0057] Step S210, obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge network; wherein the real-time resource information data includes the current cloud resource information data, the current edge resource information data and the actual end-side resource information data at the previous moment.
[0058] In this embodiment, each end-side node of the cloud-edge network is in a mobile state, and the clustering process of the end-side nodes can be triggered periodically (such as 0 o'clock every day), manually, or when a large delay event is detected, to obtain the end-side node clustering result. Among them, the large delay event can be: the delay of completing a statistical task. If the delay of the task within the preset time period is greater than the preset delay threshold, it is determined that a large delay event has been detected; the preset time period can be set according to actual needs, and there is no limit here. For example, the preset time period is the last 10 minutes, half an hour or 1 hour, etc.; the delay threshold can be set according to actual needs, and there is no limit here.
[0059] In some possible embodiments, the end-side nodes can be clustered based on the delay, based on which, obtaining the end-side node clustering result of the cloud-edge-end network includes: obtaining the communication delay information of the cloud-edge-end network within a preset time range, the communication delay information including the first delay sequence information between the end-side node and the edge node and the second delay sequence information between the end-side nodes; clustering each end-side node based on the communication delay information to obtain the end-side node clustering result. Among them, the preset time range can be set according to actual needs, for example, the preset time range can be the last 12 hours or 1 day, etc.
[0060] Since the cluster head needs to frequently process the task information uploaded by the terminal in the cluster and communicate with the outside of the cluster, the device with stable link to the edge side and rich communication resources should be selected as the cluster head, and then the clustering operation is implemented based on the communication delay between the terminal device and the cluster head to minimize the overall communication delay. Based on this, in a possible implementation method, the above-mentioned clustering of each end-side node based on the communication delay information to obtain the end-side node clustering result may include: based on the first delay sequence information, a node with unchanged delay relative to the edge node in each end-side node is used as the initial cluster head; based on the first delay sequence information and the second delay sequence information, a preset number of nodes in each end-side node that meet the first preset condition are used as cluster heads; wherein the first preset condition includes that the delay relative to the edge node is unchanged and the delay between the node and the initial cluster head is the largest; based on the second delay sequence information, the unclustered nodes in each end-side node are added to the corresponding cluster group to obtain the end-side node clustering result; wherein the delay between the node and the cluster head of the corresponding cluster group is the smallest.
[0061] The above-mentioned real-time resource information data can be detected and can be divided into three categories: computing resource information, storage resource information and communication resource information, that is, the current cloud resource information data, the current edge resource information data and the actual terminal resource information data at the previous moment all include these three types of data. Among them, the current cloud resource information data includes the resource information data of each cloud node, the current edge resource information data includes the resource information data of each edge node, and the actual terminal resource information data includes the resource information data of each terminal node.
[0062] Among them, computing resource information can include computing capacity (i.e., average computing time) and the available amount of computing resources of the corresponding type (i.e., available computing resources, which is a time variable); storage resource information includes remaining storage, which is also a time variable; communication resource information can include communication delay information, which is also a time variable. Based on this, the actual end-side resource information data can include available computing resources, remaining storage, and communication delay information between the node and the cluster head.
[0063] The above-mentioned real-time task information data may include information that affects resource allocation, such as the data volume of the task, the demand of the task for the type of computing resources, the minimum amount of computing resources for the task, and the maximum completion deadline of the task. The real-time task information data can be obtained by parsing the relevant parameter data of the task, and the relevant parameter data can be automatically generated (such as the data volume of the task) or manually given (such as the demand of the task for the type of computing resources, the minimum amount of computing resources for the task, and the maximum completion deadline of the task).
[0064] Step S220 , performing spatiotemporal sequence prediction on the resource information of each end-side node according to the end-side node clustering result and the actual end-side resource information data, to obtain the predicted end-side resource information data at the current moment.
[0065] After clustering, the network topology on the end side is divided, and the resource information in each cluster group constitutes a time-varying subgraph, and there is no mutual influence between different cluster groups. Based on this, step S220 may include: according to the clustering result of the end-side nodes and the actual end-side resource information data, obtaining the actual cluster resource information data corresponding to each cluster group; inputting the actual cluster resource information data corresponding to each cluster group into the preset spatiotemporal prediction model to obtain the predicted end-side resource information data output by the spatiotemporal prediction model; wherein the spatiotemporal prediction model is obtained by spatial feature learning and temporal feature learning of the historical end-side resource information data, and the predicted end-side resource information data includes the resource information data of each node in each cluster group.
[0066] In some possible embodiments, the above-mentioned spatiotemporal prediction model can be obtained by using the GCN (Graph Convolutional Networks, convolutional neural network)-LSTM (Long Short-Term Memory, long short-term memory network) algorithm, that is, the spatiotemporal prediction model can adopt the GCN-LSTM model, which is mainly composed of two parts: a graph convolution network and a long short-term memory network. Optionally, the spatiotemporal prediction model may include two layers of graph convolution networks, two layers of long short-term memory networks, a random inactivation layer, and a fully connected layer connected in sequence. The actual cluster resource information data corresponding to each cluster is three time series data corresponding to the available amount of computing resources, the remaining storage capacity, and the communication delay information between the node and the cluster head. These three time series data are used as inputs, and the topological structure of the cloud-edge network is captured by using a two-layer graph convolution network to obtain spatial features. The time series with spatial features is then input into the two-layer long short-term memory network, and the dynamic changes of the time series are obtained through information transmission between units to capture the time features; finally, the result (i.e., the prediction end-side resource information data) is obtained through the random inactivation layer (i.e., the dropout layer) and the fully connected layer.
[0067] Step S230, performing resource scheduling of the task according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data, and obtaining a target resource scheduling result.
[0068] Through steps S210 and S220, the resource information data of all nodes in the cloud-edge network can be obtained, and then based on the preset optimization goals and constraints, task offloading can be performed between each node; wherein, the optimization goal is to minimize the total delay of each task, and the constraints may include the task's constraints on the amount of computing resources (that is, the available amount of computing resources of the node is greater than or equal to the minimum amount of computing resources for the task), the task's constraints on the storage resource requirements (that is, the remaining storage capacity of the node is greater than or equal to the data volume of the task) and the time constraints for task completion (the time for task completion is less than or equal to the maximum completion deadline for the task).
[0069] In some possible embodiments, each cluster group corresponding to the end-side node is used as a node for task unloading, and the MADDQN (multi-agent dual deep Q network) algorithm can be used to solve the above optimization problem, achieve convergence, and find a global approximate optimal solution. Based on this, step S230 may include: determining cluster resource information data based on the predicted end-side resource information data; wherein the cluster resource information data includes cluster resource information of each cluster group, and the cluster resource information of the cluster group is the sum of the resource information data of each node in the cluster group; taking each cluster group as a cluster group node, and taking the cluster resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data as the environment state input, and performing real-time task unloading in each cluster group node, each cloud node and each edge node through the multi-agent dual deep Q network MADDQN algorithm to obtain the target resource scheduling result.
[0070] The computing network resource scheduling method provided by the embodiment of the present invention is applied to a cloud-edge-end network composed of cloud nodes, edge nodes and end-side nodes, and the resource status of the end-side nodes is dynamically changing; when performing computing network resource scheduling, first obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes the current cloud resource information data, the current edge resource information data and the actual end-side resource information data at the previous moment; then, according to the end-side node clustering results and the actual end-side resource information data, perform spatiotemporal sequence prediction on the resource information of each end-side node to obtain the predicted end-side resource information data at the current moment; then, perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain the target resource scheduling result. In this way, the spatiotemporal sequence prediction can cope with the situation that the resource status of the terminal device in the mobile network changes in real time and is difficult to obtain, and the predicted data can assist the decision-making of resource scheduling, improve the accuracy of resource scheduling, and realize the full utilization of resources.
[0071] For ease of understanding, refer to Figures 3 to 6 The above-mentioned computing network resource scheduling method is introduced by way of example.
[0072] The embodiment of the present invention proposes a computing network resource scheduling method based on spatiotemporal graph learning for a mobile network, which mainly solves the following problems:
[0073] 1) The cloud-edge-end collaboration framework can combine the resources of the cloud and the edge to provide better services. However, when a large number of end-side nodes are in a mobile state, the network structure will be too complex. In this case, the complexity of resource scheduling and allocation will be greatly increased. The embodiment of the present invention proposes a resource allocation framework for dynamic resource prediction in cloud-edge-end collaboration, and introduces clustering technology to manage the end-side nodes of the mobile network, which can effectively solve the problem of overly complex network structure and facilitate resource allocation operations.
[0074] 2) In a highly dynamic environment such as a mobile network, the time-varying characteristics of the network topology make it difficult to offload tasks. Incorrect task offloading can easily lead to resource waste and affect the normal completion of computing tasks. Most of the current resource allocation schemes cannot accurately predict and characterize the resource information of each node based on the spatiotemporal correlation characteristics between nodes, and directly allocating resources cannot achieve the expected results. In view of the dynamic resources in the mobile network, an embodiment of the present invention proposes a resource allocation method based on spatial-temporal graph learning. The embodiment of the present invention uses a multivariate time series prediction model based on the GCN-LSTM algorithm to perform multivariate spatiotemporal prediction on the dynamic topology composed of cloud edge nodes. After obtaining the node resource information, it is used as the environmental state input for resource allocation decisions, thereby assisting the decision-making of resource allocation and achieving full utilization of resources.
[0075] The specific plan is as follows:
[0076] 1) Cloud-edge-device collaboration framework based on clustering technology
[0077] The numerous resource nodes under the cloud-edge-end collaboration framework can be represented as , the end-side node is u Indexed , the edge node is m Indexed , cloud node is n Indexed . Using discrete time slots To characterize the system dynamics, each time slot matches the update time scale of resource provisioning decisions. The cloud-edge-end collaborative network is modeled as an undirected acyclic graph, where the edges between nodes are links. To represent dynamic network topology. Vertex set represents the set of all available cloud edge nodes at time t. If there is a communication link between them, then there is an edge .
[0078] In order to effectively manage terminal devices, a clustering-based cloud-edge-end collaboration framework is adopted, such as Figure 3 As shown in the figure, the cloud and edge are located in fixed positions, and their information is known to the terminal devices. The terminal devices are divided into several clusters through clustering. The cluster head of each cluster acts as a resource management center of the cluster, responsible for obtaining the resource status within the cluster and then making decisions on task offloading. In addition, the cluster head is also responsible for communication between clusters and within the cluster to the cloud.
[0079] The embodiment of the present invention adopts a clustering algorithm based on delay. Since the cluster head needs to frequently process the task information uploaded by the terminals in the cluster and communicate with the outside of the cluster, the device with stable link to the edge side and rich communication resources should be selected as the cluster head, and then the clustering operation is implemented based on the communication delay between the terminal device and the cluster head to minimize the overall communication delay. The clustering process is shown in Algorithm 1.
[0080]
[0081] Assuming that the uplink and downlink are symmetrical, the link delay between two nodes is considered to be the round-trip delay. The link delay between . It means that the delay of cluster head relative to edge node is approximately unchanged. The cluster after division is expressed as , the cluster head is represented by After clustering, the communication delay information between cluster head and edge node and between cluster head and cluster head is relatively known. The unknown information is Communication delay between non-cluster head terminal devices and cluster head in the cluster ,in .
[0082] The process of the delay-based clustering algorithm is mainly: first take a device with a constant delay relative to an edge node (can be a random edge node) as the initial cluster head, then take the device that satisfies the constant delay relative to the edge node and has the largest delay with the initial cluster head as the cluster head, so that all cluster heads can be obtained. According to the delay from the device to the cluster head, the node with the smallest delay with the cluster head is added to the cluster group. In this way, the final clustering result is obtained.
[0083] 2) Node resource model
[0084] node The computing resources are , expressed as two parameters and .parameter Indicates computing power (average amount of computing time). Indicates the available amount of computing resources of the corresponding type, which is a time variable. Storage resource information Represented as the remaining storage, where is the total storage resource of the node, It is the storage resource currently occupied by the node. The link delay between .
[0085] Computing resource analysis: First, assume that the computing resources in the cloud are idealized and can always meet the needs of the task. The computing resources in the cloud and edge are known information that can be directly obtained by the end user. However, the available computing resources in the terminal device are Unknown to each other.
[0086] The computing power of heterogeneous computing resources can be expressed as
[0087] (1)
[0088] in Corresponding to the computing power of CPU (Central Processing Unit), GPU (graphics processing unit), and NPU (Neural Processing Unit), The computational requirements of the corresponding task.
[0089] Communication Resource Analysis: Setting Nodes The communication bandwidth between , the link transmission power is , represents the distance between nodes, Expressed as the channel loss factor, the channel gain is , the channel noise is The data transfer rate is
[0090] (2)
[0091] 3) Task Model
[0092] Consider the task initiated by the end node u They are all indivisible minimum tasks and are independent of each other. They are represented by four parameters .in The amount of data representing the task corresponds to the minimum limit for storage resources. is the task's requirement for computing resource type, expressed as . Indicates the minimum amount of computing resources required by a task. It is represented as the maximum completion deadline of the task. The task is offloaded from node u to node i.
[0093] The task's constraints on computing resources are:
[0094] (3)
[0095] The task's requirements for storage resources are:
[0096] (4)
[0097] pass The computing power of the corresponding node can be obtained , the calculation completion time of the task is:
[0098] (5)
[0099] If the computing task is offloaded locally, there is no transmission delay, that is, If the computing task is not offloaded locally and needs to be transmitted, there will be communication delay. :
[0100] (6)
[0101] If there is a task being processed on node i at the current moment, there is a waiting delay , Represented as a cache:
[0102] (7)
[0103] The final time constraint for task completion is:
[0104] (8)
[0105] The optimization goals of task offloading are:
[0106] (9)
[0107] 4) Resource prediction algorithm
[0108] After clustering, the network topology on the end side is divided, and the resource information in each cluster group constitutes a time-varying subgraph, and different cluster groups do not affect each other. The embodiment of the present invention uses the GCN-LSTM algorithm to predict the unknown information in each subgraph separately, where the unknown information of each node is In this way, the resource status of the terminal device can be obtained, which can then assist in resource allocation. The main framework of the GCN-LSTM algorithm is shown in the figure below.
[0109] The mobility network of each cluster is defined as an undirected graph , V is the vertex set, E is the edge set, Represented as a node and The edges between nodes Neighborhood is The adjacency matrix is represented as .if but ,if but .
[0110] The GCN-LSTM model consists of two parts: a graph convolutional network and a long short-term memory network. Figure 4 As shown, firstly, the unknown variables in the end-side node resource information are The three time series data are used as input X, and the topological structure of the cloud-edge network is captured using a two-layer graph convolutional network to obtain spatial features. Secondly, the obtained time series with spatial features is input into a two-layer long short-term memory sequence model, and the dynamic changes of the time series are obtained through information transmission between units to capture the temporal features; finally, the result is obtained through random inactivation (dropout layer) and fully connected layer.
[0111] Spatial feature learning: The GCN model constructs a filter in the Fourier domain, which acts on the nodes of the graph and their first-order neighbors to capture the spatial features between the nodes, and then constructs the GCN model by stacking multiple convolutional layers. The GCN model can obtain the topological relationship between the central node and the surrounding nodes, encode the topological structure of the network and the attributes of the links, and obtain the spatial dependency relationship. In summary, the GCN model can be used to learn spatial features from resource information data. The 2-layer GCN model can be expressed as:
[0112] (10)
[0113] In the formula, X represents the feature matrix, A represents the adjacency matrix, represents the preprocessing step, is the self-connected structure matrix, where is the identity matrix. is the degree matrix, . and denote the weight matrices of the first and second layers respectively, , Represents the activation function.
[0114] Temporal feature learning: Obtaining temporal dependency is another important issue for effective prediction. LSTM (Long Short-Term Time Series Model) can use a gating mechanism to memorize as much long-term information as possible to solve defects such as gradient disappearance in long-term prediction. The temporal feature learning module uses a two-layer LSTM (Long Short-Term Time Series Model) to obtain temporal dependency. The structure of the LSTM module is as follows Figure 5 shown.
[0115] LSTM has three inputs: the input value of the network at the current moment , the output value at the previous moment , and the unit state at the previous moment ; LSTM has two outputs: the current moment LSTM output value and the current cell state LSTM uses two gates to control the content of the cell state d. One is the forget gate, which determines how much of the cell state at the previous moment is retained at the current moment. ; The other is the input gate, which determines the input of the network at the current moment How much is saved to the cell state LSTM uses an output gate to control the cell state. How many outputs are there to the current output value of the LSTM .
[0116] Calculation of forget gate:
[0117] (11)
[0118] in is the weight matrix of the forget gate, It means that two vectors are connected to form a longer vector. is the bias term of the forget gate.
[0119] Calculation of input gate:
[0120] (12)
[0121] in is the weight matrix of the input gate, is the bias term of the input gate.
[0122] Current input unit status , calculated based on the previous output and this input:
[0123] (13)
[0124] in Represented as the weight matrix of the memory gate, It is the bias term of the memory gate. The memory gate is used to control whether the current data is input into the unit.
[0125] Output unit status at the current moment
[0126] (14)
[0127] symbol Represents element-wise multiplication.
[0128] The output gate is:
[0129] (15)
[0130] in is the weight matrix of the output gate, is the bias term of the output gate.
[0131] The final output of LSTM is determined by the output gate and the cell state:
[0132] (16)
[0133] 5) Resource Scheduling Process
[0134] After prediction, each determined constraint condition (3) (4) (8) can be obtained, and each node has a corresponding resource value. Therefore, the feasible domain for task offloading is discrete, and the optimization problem (9) can be transformed into an integer linear programming problem, which is an NP (non-deterministic polynomial) difficult problem. The embodiment of the present invention uses the MADDQN (multi-agent dual deep Q network) algorithm to solve this problem, which can achieve convergence and find the global approximate optimal solution. The mobile network computing network resource scheduling process is as follows Figure 6 shown.
[0135] After the cloud-edge network is clustered on the end side, the entire cluster is taken as a node, and then the entire network is modeled to obtain a network topology with a graph structure. Then, the resource information of the end-side network is predicted in a spatiotemporal sequence to obtain real-time prediction results. The prediction results and the known information of the task requirements and the resource status of the cloud-edge end are used as the environmental state input in the task offloading module, and the MADDQN algorithm is used to make real-time offloading decisions. The task offloading process is carried out in the cloud-edge network after clustering is completed. After the task is scheduled to the end-side cluster, the structure of the cluster will not be changed. The resource information of the end-side node obtained after resource scheduling will be fed back to the spatiotemporal prediction module as the data input of the spatiotemporal prediction module.
[0136] In the decision-making process of the MADDQN algorithm, 1) State update: Each step of the state update not only includes the resource information data of the cloud and the edge, but also the end-side resource information output in real time by the spatiotemporal prediction module. 2) Action decision: Use an array variable to represent the task decision for a certain node, where each element represents each task, and the 0 / 1 variable indicates whether it is unloaded at the node. For example, the action decision of a node is {1,0,0,0,1,0}, which means that tasks 1 and 5 are unloaded at the node, and other tasks are not unloaded at the node. 3) Reward setting: The reward function is the overall delay of task completion, which indicates the corresponding benefit that can be obtained after taking the action strategy. The cumulative benefit is represented by the Q-value function. The larger the reward, the more the cumulative benefit will increase, reflecting that it is more beneficial to execute the action.
[0137] like Figure 6 As shown in the MADDQN task unloading module in Figure 1, the reward and the corresponding action will be input into the replay memory unit. On the one hand, the memory replay unit will output the above reward value to the loss function, and update the current value network by minimizing the calculation process of the loss function. On the other hand, it will output the reward value to the target value network, and guide the update process of the current value network through the target value network. The updated current value network can output the optimal strategy. The ultimate goal is to find the unloading location for each task while maximizing the cumulative reward and minimizing the delay in task completion.
[0138] The computing network resource scheduling method provided by the embodiment of the present invention has the following beneficial effects:
[0139] 1) The computing network resource scheduling method based on spatiotemporal graph learning for mobile networks proposed in the embodiment of the present invention can effectively manage the resources of mobile terminal devices through clustering technology. The traditional scheduling method for the edge or cloud and the terminal side is to consider the appropriate unloading location based on the attributes of the task. In this unloading process, a large number of available terminal devices are not taken into consideration. The cloud-edge-terminal collaborative framework based on clustering can achieve effective utilization of terminal resources through cluster heads, and can offload tasks to other terminal devices through cluster heads.
[0140] 2) Multivariate spatiotemporal series prediction can cope with the situation where the resource status of terminal devices in the mobile network changes in real time and is difficult to obtain. The prediction algorithm can obtain the resource status for a period of time in the future, and then assist in resource scheduling decisions, effectively reducing the delay in task completion.
[0141] In order to verify the effect of the above-mentioned computing network resource scheduling method, the SUMO simulation software is used to simulate the generated data set for verification, simulating the resource allocation under the condition of mobile end-side equipment. The spatiotemporal resource prediction is combined with deep reinforcement learning to offload tasks from the end-side. The experimental test data is 24 hours of real-time network resource information. There are a total of 13 nodes, represented as edge, cloud or cluster, and a total of 1439 rounds of experiments are conducted. The training data of the experiment is 24 hours of network resource information, which is used to train the model. The evaluation indicator of the task offloading result is the average task delay.
[0142] Figure 7 The average delay of tasks in each round of testing when using the system to allocate computing network resources. It can be seen that the average delay gradually converges to less than 1 second, proving that the method provided in the embodiment of the present invention can effectively reduce task delay.
[0143] Figure 8 A schematic diagram of comparing test results of using a spatiotemporal prediction module and not using a spatiotemporal prediction module is provided in an embodiment of the present invention, such as Figure 8 As shown in the figure, the use of the spatiotemporal prediction module can effectively reduce the task delay, indicating that the performance of the scheduling system is improved after adding the prediction.
[0144] The key technical points of the embodiments of the present invention are as follows:
[0145] 1. A new cloud-edge-end computing network architecture based on clustering technology proposed in an embodiment of the present invention. Mainly for the situation where the terminal device is in motion, the terminal device is divided into different clusters using latency-based clustering. The cluster head manages and maintains the end-side resources, which can effectively simplify the network structure and better deal with the resource management difficulties caused by the mobility of the terminal devices.
[0146] 2. The embodiment of the present invention proposes a predictive scheduling method for dynamic resources in a mobile network. A prediction model based on a spatiotemporal graph prediction algorithm is used to predict the dynamic topology of cloud edge nodes. After obtaining node resource information, each node is dynamically scheduled in real time. This assists in resource allocation decisions, achieves rapid response to tasks and full utilization of resources.
[0147] Corresponding to the above-mentioned computing network resource scheduling method, an embodiment of the present invention further provides a computing network resource scheduling device, which is applied to a cloud-edge-end network composed of cloud nodes, edge nodes and end-side nodes. The resource status of the end-side nodes changes dynamically, that is, the end-side nodes are in a mobile state. Fig. 9 The schematic diagram of the structure of a computing network resource scheduling device is shown, and the computing network resource scheduling device includes:
[0148] The data acquisition module 901 is used to obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes the current cloud resource information data, the current edge resource information data and the actual end-side resource information data at the previous moment;
[0149] The spatiotemporal prediction module 902 is used to perform spatiotemporal sequence prediction on the resource information of each of the end-side nodes according to the end-side node clustering result and the actual end-side resource information data, so as to obtain the predicted end-side resource information data at the current moment;
[0150] The resource scheduling module 903 is used to perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result.
[0151] The computing network resource scheduling device provided by the embodiment of the present invention is applied to a cloud-edge-end network composed of cloud nodes, edge nodes and end-side nodes, and the resource status of the end-side nodes is dynamically changing; when performing computing network resource scheduling, first obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes the current cloud resource information data, the current edge resource information data and the actual end-side resource information data at the previous moment; then, according to the end-side node clustering results and the actual end-side resource information data, perform spatiotemporal sequence prediction on the resource information of each end-side node to obtain the predicted end-side resource information data at the current moment; and then perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain the target resource scheduling result. In this way, the spatiotemporal sequence prediction can cope with the situation that the resource status of the terminal device in the mobile network changes in real time and is difficult to obtain, and the predicted data can assist the decision-making of resource scheduling, improve the accuracy of resource scheduling, and realize the full utilization of resources.
[0152] Furthermore, the above-mentioned data acquisition module 901 is also used to: obtain the communication delay information of the cloud-edge network within a preset time range, the communication delay information including the first delay sequence information between the end-side node and the edge node and the second delay sequence information between the end-side nodes; cluster each of the end-side nodes based on the communication delay information to obtain the end-side node clustering result.
[0153] Furthermore, the above-mentioned data acquisition module 901 is also used to: based on the first delay sequence information, take a node in each of the end-side nodes whose delay relative to the edge node is unchanged as an initial cluster head; based on the first delay sequence information and the second delay sequence information, take a preset number of nodes in each of the end-side nodes that meet the first preset condition as cluster heads; wherein the first preset condition includes that the delay relative to the edge node is unchanged and the delay between the node and the initial cluster head is the largest; based on the second delay sequence information, add the unclustered nodes in each of the end-side nodes to the corresponding cluster group to obtain the end-side node clustering result; wherein the delay between the node and the cluster head of the corresponding cluster group is the smallest.
[0154] Furthermore, the above-mentioned spatiotemporal prediction module 902 is also used to: obtain actual cluster resource information data corresponding to each cluster group according to the end-side node clustering results and the actual end-side resource information data; input the actual cluster resource information data corresponding to each of the cluster groups into a preset spatiotemporal prediction model to obtain the predicted end-side resource information data output by the spatiotemporal prediction model; wherein the spatiotemporal prediction model is obtained by performing spatial feature learning and temporal feature learning on historical end-side resource information data, and the predicted end-side resource information data includes the resource information data of each node in each of the cluster groups.
[0155] Furthermore, the above-mentioned spatiotemporal prediction model includes two layers of graph convolutional networks, two layers of long short-term memory networks, a random inactivation layer and a fully connected layer connected in sequence.
[0156] Furthermore, the resource scheduling module 903 is also used to: determine cluster resource information data based on the predicted end-side resource information data; wherein the cluster resource information data includes cluster resource information of each cluster, and the cluster resource information of the cluster is the sum of resource information data of each node in the cluster; take each cluster as a cluster node, and take the cluster resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data as environment state inputs, and perform real-time task unloading in each cluster node, each cloud node and each edge node through the multi-agent dual deep Q network MADDQN algorithm to obtain the target resource scheduling result.
[0157] Furthermore, the above-mentioned actual end-side resource information data includes the available amount of computing resources, the remaining storage capacity and the communication delay information between the node and the cluster head; the real-time task information data includes the data volume of the task, the task's requirement for the type of computing resources, the task's minimum limit on the amount of computing resources and the maximum completion deadline of the task.
[0158] The computing network resource scheduling device provided in this embodiment has the same implementation principle and technical effects as those of the aforementioned computing network resource scheduling method embodiment. For the sake of brief description, for matters not mentioned in the computing network resource scheduling device embodiment, reference may be made to the corresponding contents in the aforementioned computing network resource scheduling method embodiment.
[0159] like Fig.10 As shown, an electronic device 1000 provided by an embodiment of the present invention includes: a processor 1001, a memory 1002 and a bus, the memory 1002 stores a computer program that can be run on the processor 1001, when the electronic device 1000 is running, the processor 1001 and the memory 1002 communicate through the bus, and the processor 1001 executes the computer program to implement the above-mentioned network resource scheduling method.
[0160] Specifically, the memory 1002 and the processor 1001 can be general-purpose memories and processors, which are not specifically limited here.
[0161] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the computing network resource scheduling method described in the above method embodiment is executed. The computer-readable storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a RAM, a magnetic disk or an optical disk, and other media that can store program codes.
[0162] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.
[0163] In all examples shown and described herein, any specific values should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.
[0164] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.
[0165] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.
[0166] 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 on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0167] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A computing network resource scheduling method, characterized in that: Applied to a cloud-edge-end network consisting of cloud nodes, edge nodes and end-side nodes, the resource status of the end-side nodes changes dynamically; the computing network resource scheduling method includes: Obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes current cloud resource information data, current edge resource information data and actual end-side resource information data at the previous moment; According to the end-side node clustering result and the actual end-side resource information data, a spatiotemporal sequence prediction is performed on the resource information of each of the end-side nodes to obtain the predicted end-side resource information data at the current moment; Performing resource scheduling of tasks according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data, and the real-time task information data to obtain a target resource scheduling result; The obtaining of the end-side node clustering result of the cloud-edge-end network includes: Acquire communication delay information of the cloud-edge-end network within a preset time range, wherein the communication delay information includes first delay sequence information between the end-side node and the edge node and second delay sequence information between two end-side nodes; Clustering each of the end-side nodes based on the communication delay information to obtain the end-side node clustering result; The performing spatiotemporal sequence prediction on the resource information of each of the end-side nodes according to the end-side node clustering result and the actual end-side resource information data to obtain the predicted end-side resource information data at the current moment includes: Acquire actual cluster resource information data corresponding to each cluster according to the end-side node clustering result and the actual end-side resource information data; Input the actual cluster resource information data corresponding to each of the clusters into a preset spatiotemporal prediction model to obtain the predicted end-side resource information data output by the spatiotemporal prediction model; wherein the spatiotemporal prediction model is obtained by performing spatial feature learning and temporal feature learning on the historical end-side resource information data, and the predicted end-side resource information data includes the resource information data of each node in each of the clusters; The performing resource scheduling of tasks according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result includes: Determine cluster resource information data according to the predicted end-side resource information data; wherein the cluster resource information data includes cluster resource information of each cluster, and the cluster resource information of the cluster is the sum of resource information data of each node in the cluster; Each of the clusters is taken as a cluster node, and the cluster resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data are taken as environmental state inputs. Real-time task unloading is performed in each of the cluster nodes, each of the cloud nodes and each of the edge nodes through the multi-agent dual deep Q network MADDQN algorithm to obtain the target resource scheduling result.
2. The method for scheduling computing network resources according to claim 1, characterized in that: The clustering of each of the end-side nodes based on the communication delay information to obtain the end-side node clustering result includes: Based on the first delay sequence information, taking a node among the end-side nodes whose delay relative to the edge node is unchanged as an initial cluster head; Based on the first delay sequence information and the second delay sequence information, a preset number of nodes that meet a first preset condition among the end-side nodes are used as cluster heads; wherein the first preset condition includes that the delay relative to the edge node is unchanged and the delay between the node and the initial cluster head is the largest; Based on the second delay sequence information, the unclustered nodes in each of the end-side nodes are added to the corresponding cluster group to obtain the end-side node clustering result; wherein the delay between the node and the cluster head of the corresponding cluster group is the smallest.
3. The method for scheduling computing network resources according to claim 1, characterized in that: The spatiotemporal prediction model includes a two-layer graph convolutional network, a two-layer long short-term memory network, a random inactivation layer and a fully connected layer connected in sequence.
4. The method for scheduling computing network resources according to any one of claims 1 to 3, characterized in that: The actual end-side resource information data includes the available computing resources, the remaining storage capacity and the communication delay information between the node and the cluster head; the real-time task information data includes the data volume of the task, the task's requirement for the computing resource type, the task's minimum limit on the amount of computing resources and the task's maximum completion deadline.
5. A computing network resource scheduling device using the computing network resource scheduling method according to any one of claims 1 to 4, characterized in that: Applied to a cloud-edge-end network consisting of cloud nodes, edge nodes and end-side nodes, the resource status of the end-side nodes is dynamically changing; the computing network resource scheduling device includes: A data acquisition module, used to obtain the end-side node clustering results, real-time resource information data and real-time task information data of the cloud-edge-end network; wherein the real-time resource information data includes current cloud resource information data, current edge resource information data and actual end-side resource information data at the previous moment; A spatiotemporal prediction module, configured to perform spatiotemporal sequence prediction on the resource information of each of the end-side nodes according to the end-side node clustering result and the actual end-side resource information data, to obtain the predicted end-side resource information data at the current moment; The resource scheduling module is used to perform task resource scheduling according to the predicted end-side resource information data, the current cloud resource information data, the current edge resource information data and the real-time task information data to obtain a target resource scheduling result.
6. An electronic device comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the computer program, the computing network resource scheduling method described in any one of claims 1 to 4 is implemented.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the computing network resource scheduling method described in any one of claims 1 to 4 is executed.
Citation Information
Patent Citations
Centralized route computing method in software-defined sensing network
CN107969008A
Dynamic adjustment method and system for cloud-side-end communication architecture
CN112867088A
System and Method for Enabling Service Lifecycle Based Policy, Licensing, and Charging in a Network Function Virtualization Ecosystem
US20170104609A1
Method and system for dynamically adjusting communication architecture
WO2022142277A1
