An omnipresent computing power network computing power resource scheduling method

Abstract

This invention discloses a method for scheduling computing resources in ubiquitous computing networks (UCPN). Starting from the topology and resource characteristics of UCPN networks, a unified UCPN hypergraph model is proposed. This model consists of heterogeneous nodes with differentiated resources and hyperedges formed by inter-node dependencies and supply capabilities. Each hyperedge contains a set of heterogeneous nodes with similar resource supply capabilities, rather than a bilateral relationship between two individuals, thus improving node selection and resource mapping speed. Based on the constructed UCPN hypergraph model, a hypergraph convolutional attention network for UCPN computing resource scheduling is proposed. This network achieves end-to-end data transmission path planning while scheduling computing nodes through hypergraph convolution and graph attention mechanisms. Simulation results show that the proposed HGcov-atten outperforms benchmark schemes in terms of task execution time and resource utilization.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, and mainly designs an effective and fast method for scheduling computing resources in ubiquitous computing networks. Background Technology

[0002] Ubiquitous Computing Power Network (UCPN) interconnects heterogeneous nodes at different distances and with varying computing power. By scheduling computing, communication, storage, and control resources close to the data source on demand, it enables on-network computation, communication, and control of massive and diverse computationally intensive tasks, significantly improving processing efficiency and reducing cloud pressure while providing users with faster response times. While UCPN offers many advantages, achieving flexible scheduling of computing resources to meet the quality of service requirements of massive and diverse computationally intensive tasks is extremely challenging. Specifically:

[0003] UCPN communication and computing resources are characterized by their distributed, fragmented, and limited nature, with resources interacting and constraining each other. UCPN nodes are diverse in type and geographically dispersed, with significant differences in the communication, computing, and storage resource capabilities offered by different types of nodes. Furthermore, the multidimensional resources among nodes are independent of each other. However, to complete a specific computing task, these multidimensional resources may exist in a group. During task execution, situations can easily arise where a device has sufficient computing power but suffers from limited transmission resources due to environmental factors. Therefore, to achieve effective collaboration among multidimensional resources, the resource scheduling scheme should address the differences in resource supply capabilities among heterogeneous nodes and the potential mutual influences between resources, thereby improving the effectiveness of resource scheduling and achieving rapid response.

[0004] Currently, research on resource scheduling in complex networks has made considerable progress, mainly categorized into model-based and data-driven methods. Model-based methods require establishing an optimization problem and solving it using techniques such as graph coloring, graph matching, graph partitioning, and game theory. As the scale of networks and data increases, optimization-based or heuristic solutions become computationally intensive and difficult to solve globally. Unlike model-based methods, data-driven methods reduce reliance on physical models. State-of-the-art deep learning algorithms, including GCN, GAT, and GraphSAGE, are naturally adapted to non-Euclidean graph structures and can be used to design (model-free) data-driven network resource scheduling, relying only on the system's historical dataset. Both of these methods essentially model the irregular topology of complex networks as graphs, including vertices and edges between vertex pairs. While traditional graph structures can effectively model pairwise relationships (i.e., first-order relationships) between two nodes, they neglect the differences in resource supply capabilities among heterogeneous nodes in a UCPN and the potential higher-order group relationships between computing power and communication resources. Therefore, traditional graphs cannot fully represent the potential high-order information in UCPN multidimensional resources, thus limiting the flexibility, accuracy, and efficiency of the above methods in UCPN multidimensional resource scheduling.

[0005] In recent years, Hypergraph Neural Networks (HGNNs) with hypergraph structures have been proven to represent high-order interactions between network elements and handle multimodal data. Hypergraphs are generalizations of traditional graphs; a single hyperedge can connect any number of vertices, representing complex multi-dimensional interactions. Furthermore, HGNNs offer flexibility in hyperedge generation, generating hyperedges with different information based on specific application scenarios. These multi-type hyperedges contain more nodes and higher-order interactions, exhibiting nesting, heterogeneity, and multi-layered, multi-dimensional characteristics. Therefore, by aggregating relevant hyperedge features, HGNNs can naturally handle various high-order relationships and possess good scalability. While hypergraphs can effectively model relationships between multiple nodes, methods for unified modeling and management scheduling of multi-dimensional resources are still in their early stages of research. How to consider the high-order relationships between multi-dimensional resource providers and establish a suitable hypergraph structure for UCPN remains a worthwhile area of ​​study. In addition, we need to effectively perform distributed communication and joint scheduling of computing resources in UCPN under multi-heterogeneous node collaboration. All nodes with computing and storage resources will be considered target computing nodes, and computing tasks will be scheduled to multiple target computing nodes along the optimal path, rather than simply being transferred to the server. Therefore, how to dynamically match and schedule the corresponding computing power, communication, and other resources according to the computing needs of complex and diverse tasks and the performance status of the network is also a very challenging problem. Summary of the Invention

[0006] Purpose of the invention: To address the problems of the surge in the number of services in ubiquitous computing networks, the fragmentation, distribution, and limited nature of computing and communication resources, and the low resource utilization caused by mutual influence and constraints among resources. This invention proposes an effective and fast method for scheduling computing resources in ubiquitous computing networks, which improves the utilization rate of computing resources through effective scheduling of computing resources and rapid establishment of end-to-end optimal paths.

[0007] Technical solution: To achieve the above objectives, the technical solution adopted by this invention is as follows:

[0008] A method for scheduling computing resources in a ubiquitous computing network includes the following steps:

[0009] Step S1: Establish a UCPN hypergraph structure consisting of vertices and hyperedges. The vertices are composed of heterogeneous nodes with differentiated resources. Correspondingly, construct hyperedge groups with dependency relationships, connectivity features, k-hop neighbors, and multi-dimensional resource attributes.

[0010] Step S2: Based on the UCPN hypergraph structure obtained in Step S1, design the UCPN computing resource scheduling algorithm, namely the Hypergraph Convolutional Attention Network (HGcov-atten). This hypergraph convolutional attention network consists of two parts: hypergraph convolution and graph attention mechanism. The hypergraph convolution is responsible for generating UCPN heterogeneous node evaluation values, which reflect the computing resource supply capacity of heterogeneous nodes. The Q heterogeneous nodes with larger evaluation values ​​are selected as the target computing node set to achieve dynamic scheduling of computing resources. The graph attention mechanism reflects the importance of heterogeneous nodes through link transmission rate, latency, and reliability characteristics. Based on the importance of heterogeneous nodes, the optimal next hop is gradually found between the source node and the target computing node to establish an end-to-end path.

[0011] Beneficial effects:

[0012] This invention proposes an efficient and fast method for scheduling computing resources in ubiquitous computing networks. Addressing the challenges of surging demand for large-scale distributed high-computing applications, fragmented, distributed, and differentiated computing and communication resources, and low resource utilization due to mutual constraints and influences among resources, this method introduces a hypergraph to classify or group heterogeneous nodes providing computing resources in the ubiquitous computing network. This achieves preliminary screening of complex resources and improves the efficiency of heterogeneous node selection and resource mapping. Simultaneously, a computing resource scheduling algorithm is designed to achieve efficient scheduling of computing resources and rapid establishment of end-to-end optimal paths, improving the speed of multi-dimensional resource collaborative scheduling and the utilization rate of computing resources. Simulation results show that the proposed HGcov-atten outperforms benchmark schemes in terms of task execution time and resource utilization. Attached Figure Description

[0013] Figure 1This is a schematic diagram of the computing resource scheduling framework for ubiquitous computing networks according to the present invention;

[0014] Figure 2 It is a schematic diagram of hyperedges with dependency and connectivity characteristics;

[0015] Figure 3 Here are schematic diagrams of k-hop neighbor and multi-dimensional resource attribute hyperedge group generation; where (a) is a schematic diagram of k-hop neighbor hyperedge group; and (b) is a schematic diagram of multi-dimensional resource attribute hyperedge group.

[0016] Figure 4 This is a schematic diagram of the instantiation of UCPN computing power node distribution;

[0017] Figure 5 This is a diagram illustrating how the execution time of UCPN computing tasks varies with the number of computing tasks.

[0018] Figure 6 This is a diagram illustrating how the utilization rate of UCPN computing resources changes with the number of computing tasks. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of this invention by those skilled in the art in various equivalent forms fall within the scope defined by the appended claims.

[0020] This invention proposes an effective and fast method for scheduling computing resources in ubiquitous computing networks, comprising the following steps:

[0021] Step S1: Establish a new, unified UCPN hypergraph structure, consisting of vertices and hyperedges. Vertices are composed of heterogeneous nodes with differentiated resources. Considering the interdependencies between heterogeneous nodes in UCPN, as well as the differentiated supply capabilities of computing power and communication resources of heterogeneous nodes, hyperedge groups with dependency relationships, connectivity features, k-hop neighbors, and multi-dimensional resource attributes are constructed accordingly.

[0022] Step S2: Based on the UCPN hypergraph structure obtained in Step S1, design the UCPN computing resource scheduling algorithm, namely the Hypergraph Convolutional Attention Network (HGcov-atten). This network consists of two parts: hypergraph convolution and graph attention mechanism. Hypergraph convolution is responsible for generating UCPN heterogeneous node-level evaluation values, which reflect the computing resource supply capacity of heterogeneous nodes. Q heterogeneous nodes with larger evaluation values ​​are selected as the target computing node set to achieve dynamic scheduling of computing resources. The graph attention mechanism reflects the importance of heterogeneous nodes through link transmission rate, latency, and reliability characteristics. Based on the importance of heterogeneous nodes, it progressively searches for the optimal next hop between the source node and the target computing node, thereby establishing an end-to-end path.

[0023] Step S3: Evaluate the performance of the computing resource scheduling method proposed by UCPN. The performance of the proposed HGcov-atten algorithm in terms of task execution time and resource utilization is discussed.

[0024] First, in step S1, UCPN is described as a hypergraph structure. Let ε be the vertex set, ε be the hyperedge set, and H be the incidence matrix. We extracted four types of relationships between vertices in UCPN to generate hyperedge groups, namely, interdependent hyperedge groups ε. ind ε, connectivity feature of hyperedge group con k-hop neighbor superedge group ε hop (The shortest path set P between source and destination nodes is obtained through the shortest path algorithm) shortest ) and multidimensional resource attribute hyperedge group ε res And the four types of hyperedge groups are merged to form a hyperedge ε, specifically:

[0025] S11, Using the hyperedge group ε with dependency relationships ind Because cascading failures between computing nodes can easily lead to the failure of the computing network, this cascading effect typically occurs on nodes located on the same power line. Therefore, we group computing nodes located on the same power line into a superedge. Specifically, for heterogeneous nodes communicating via power lines, the number of power lines connected to a given heterogeneous node is defined as the node's degree, D. Nodes with a degree D ≥ 3 are grouped together to form a dependent power line. For example... Figure 2 As shown, nodes (7, 4, 5, 6), (7, 8, 9, 10), and (7, 11, 12, 13, 14) are located on three different electric field lines (i.e., red, purple, and blue). Assume that the set of electric field lines with dependencies in UCPN is C, and the binary variable α... c,v To indicate whether node v is on power line c, the set of vertices on power line c is defined as follows: For a vertex set, use the hyperedge group ε with dependencies. ind Represented as:

[0026] ε ind ={N ind (c)|c∈C}

[0027] S12, Hyperedge group ε using connectivity feature con Communication between heterogeneous nodes in UCPN is not limited to power lines; it also includes 4G / 5G, Wi-Fi, LoRa, ZigBee, Bluetooth, and more. Figure 2 As shown. ε con The goal is to fully utilize the connectivity between node pairs to reflect whether a vertex has more next-hop options. Assume the UCPN connectivity matrix is ​​CN, representing the number of communicable links between vertices. For example, node 2 has a connectivity of cn2 = 3, and node 7 has a connectivity of cn7 = 6. Assume the set of vertices with connectivity β is defined as... where β∈[β min ,β max ], β min and β max Let and be the maximum values ​​in CN, respectively. Then, the hyperedge group ε is used based on the connectivity feature. con Represented as:

[0028] ε con ={N con (β)|β∈[β min ,β max ]}

[0029] S13, Using k-hop neighbors for the super-edge group ε hop To ensure the latency requirements of computational tasks, the number of routing hops in UCPN is generally limited. hop The goal is to find the relevant vertices of a computing power center point through the k-hop reachable locations in the graph structure. This can reflect the distance traversed to transfer the computing task to a suitable computing power node, such as... Figure 3 As shown in (a), the k-hop neighborhood of vertex v is defined as... in To determine whether vertex v has a k-hop neighbor node u, the superedge group Φ of the k-hop neighbors is used. hop Represented as:

[0030]

[0031] S14. Hyperedge group Φ using multidimensional resource attributes res In UCPN's distributed computing nodes, computational and communication resources exist in groups to collaboratively complete a computational task. This results in non-paired relationships between computing and communication resources, and no usable graph structure. res The goal is to extract the differentiated multi-dimensional resource supply capabilities among UCPN computing power nodes, such as... Figure 3 As shown in (b). Let This represents two resource types: computing and communication. Represents the vertex set that can be provided Based on the maximum resource value, we differentiate computing nodes with different resource capabilities. For example, we divide the computing resources of all computing nodes in UCPN into ω... cp 0 intervals, ζ cp ∈{0,1 / ω cp ,2 / ω cp ,...,(ω cp -1) / ω cp} belongs to the interval A subset of vertices can be represented as Where τ cp =ζ cp ·ω cp +1, ζ cp and ω cp All are adjustable positive parameters, then the hyperedge group ε constructed through the multidimensional resource attributes between vertices is... res Represented as:

[0032]

[0033] Secondly, in step S2, based on the constructed UCPN hypergraph structure, an effective computing resource scheduling algorithm is designed for UCPN to accurately allocate multi-granularity computing tasks to fragmented computing nodes, ensuring efficient utilization of computing resources on fragmented nodes and high-quality execution of diverse intelligent tasks, thereby improving the flexibility and accuracy of on-demand services for heterogeneous distributed computing power and communication resources. The proposed computing resource scheduling algorithm, namely the Hypergraph Convolutional Attention Network (HGcov-atten), needs to comprehensively consider the network state and the distribution of computing and communication resources to establish end-to-end path planning while scheduling computing nodes. Figure 1 As shown, it consists of two parts: the upper graph represents the initial UCPN topology, and the lower graph represents the algorithm architecture. HGcov-atten mainly includes two parts: hypergraph convolution and hypergraph attention mechanism. The purpose of hypergraph convolution is to select a subset of nodes with better computing power resources in UCPN as the target computing node set. The purpose of the graph attention mechanism is to find the optimal end-to-end path from the source node to the target computing power node set. Specifically:

[0034] S21. Hypergraph convolution originates from hypergraph learning theory and regular graph convolution. Initially, the problem of spectral hypergraph partitioning with normalized hypergraph segmentation was a fundamental problem in hypergraph theory. In the study of the hypergraph partitioning problem, the hypergraph Laplacian operator was proposed, providing a bridge to hypergraph convolution. Therefore, the hypergraph Laplacian operator Δ is key to spectral hypergraph convolution. Δ is a square semidefinite matrix with only real eigenvalues, and the smallest eigenvalue of Δ is 0. Based on these properties, through eigenvalue decomposition Δ = ΦΛΦ T We can obtain a non-negative eigenvalue matrix Λ = diag(λ1,...,λ n ) and the orthogonal eigenvectors Φ = diag(Φ1,...,Φ n ), λ n Φ is an eigenvalue. n These are orthogonal eigenvalues. The continuous eigenvector Φ is used as the spectral sequence signal x = (x1,...,x...). n The Fourier basis of ) is defined as follows: The Fourier transform is defined as follows: Inverse Fourier transform is Therefore, the spectral convolution of the hypergraph signal x and the convolution kernel g can be expressed as:

[0035]

[0036] Where ⊙ is the Hadamard product of the two matrices. Φ T g is considered a trainable filter with an eigenvalue, using g θ (Λ) represents the Fourier coefficient. Since the time complexity of calculating the above equation is O(n),... Furthermore, the eigenvalue decomposition of the hypergraph is very time-consuming; therefore, the 1stChebNet efficient computation method is adopted. This involves using Chebyshev polynomial approximation filtering and the product of the hypergraph signal:

[0037]

[0038] in This is the scaled eigenvector matrix, with a range of [-1, 1]. The scaling is performed to satisfy the Chebyshev polynomial T. k K of (x) th The condition for truncating the expansion, λ max Let θ be the spectral radius. k Let T be the coefficients of the Chebyshev polynomial. The recursive definition of a Chebyshev polynomial is T. k (x)=2xT k-1 (x)-T k-2 (x), where T0(x) = 1 and T1(x) = x. According to the parameters proposed by 1srChebNet, due to scale adaptation, K = 1, λ max =2.

[0039] Given a UCPN hypergraph For a vertex Vertex degree is defined as d(v) = ∑ e∈ε w(e)H(v,e), for a hyperedge e∈ε, the degree of the hyperedge is defined as w(e)H(v,e). Then there is a UCPN hypergraph with N vertices. Spatial features can form a Laplace matrix Δ = ID v -1 / 2 HWD e -1 H T D v -1 / 2 , where I is the identity matrix, and D is the identity matrix. v It is the vertex degree matrix, D e This is the hyper-edge degree matrix, W is initialized to the identity matrix. The hypergraph convolution operation is then obtained:

[0040] (x*g) HG ≈θ0x-θ1D v -1 / 2 HWD e -1 H T D v -1 / 2 x

[0041] To constrain the number of parameters and prevent overfitting, the parameter θ is defined as θ = -2θ1 = 2θ0D. v -1 / 2 H -1 W -1 D e -1 (H T ) - 1 D v -1 / 2 Therefore, hypergraph convolution can be further simplified to...

[0042] (x*g) HG ≈θD v -1 / 2 HWD e -1 H T D v -1 / 2 x

[0043] The input to hypergraph convolution is the task feature X. ser1 Vertex resource characteristics X res With the correlation matrix H. Input matrix This is used to describe the attributes and resource requirements of the l-th task, where C0 is the dimension of the task features. Among them, sr l da represents the source node of task l. l The task size is represented in bits, pt. l (In units of cycles / bits) represents the computational intensity of the task, i.e., the computational resources required to process 1 bit of data. Input matrix Used to represent The resource features of each vertex are given, where C1 is the dimension of the vertex resource features. cp n This represents the amount of computational resources that the nth node can provide. The incidence matrix H reveals the relationships between vertices and edges in a hypergraph. One can be used The dimensional correlation matrix H represents:

[0044]

[0045] X ser1 and X res Substituting these values ​​into the hypergraph convolution expression, we obtain the following expression:

[0046] Y (t+1) =σ sigmoid (D v -1 / 2 HWD e -1 H T D v -1 / 2 X (t) Θ (t) )

[0047] in Y is the input signal of the t-th layer. (t+1) Let σ represent the output signal of the t-th layer, and let σ be the activation function of the sigmoid function. sigmoid .

[0048] The output of the last layer of the hypergraph convolution is the evaluation value of the computing power nodes within k hops of the source node that generated the task. in For Y l Dimensions Y represents the maximum number of computing power nodes within k hops of the source node for the l-th computation task. lThe elements in the matrix are arranged in descending order, and this value reflects the multi-dimensional resource supply capacity of UCPN computing power nodes. However, since Y can only be used to select computing power nodes that meet the computing and communication needs of diverse sub-tasks, and ignores the dynamic matching of network performance status during the transmission process from the source node to the destination node, it is very likely that computing power nodes with suitable computing resources but limited transmission processes will be encountered. Therefore, we select the top Q computing power nodes with higher evaluation values ​​as the target computing power node set for the computing tasks, and finally output the matrix. Y' l The elements are arranged in descending order.

[0049] S22. The main idea behind implementing UCPN routing using the graph attention mechanism is to assign different weights to first-order neighbor nodes based on network performance metrics (i.e., transmission rate, latency, and reliability). The neighbor node with the largest weight is selected as the optimal next hop, which is also the central node of the next graph attention mechanism. The graph attention mechanism is executed multiple times on different central nodes to generate the optimal end-to-end path. In the graph attention mechanism, the input matrix is ​​the target computing power node set Y' and the task communication requirement characteristics. and network performance characteristics Input matrix Used to describe the lth s The QoS requirements of each subtask, where C2 is the dimension of the task features. Where tr l Indicates the transmission rate requirement, dl l Indicates latency requirements, rl l Represents reliability requirements. Input matrix Used to describe The network performance characteristics of each vertex, where C3 is the dimension of the network performance characteristics. tr e(m,n) Let lt represent the transmission rate of link e(m,n) between nodes m and n. e(m,n) pl represents the delay of link e(m,n). e(m,n) This represents the reliability of link e(m,n). The following section will detail the specific process by which the graph attention mechanism achieves optimal next-hop selection.

[0050] First, a shared weight matrix is ​​trained for all computing nodes in UCPN. This serves as a mapping between the input and output features of a node. During computation, the weight matrix is ​​first used to map the features of the first-order neighboring computing power nodes of the source node v0, and the results of these mappings are concatenated. Then, a feedforward neural network is input, mapping the concatenated vector to a single value. Finally, the LeakyReLU activation function is used to obtain the absolute attention coefficient of node n with respect to the source node v0.

[0051]

[0052] Where || represents the concatenation operation. This indicates the importance of node n to the source node v0. It is the set of first-order nearest neighbors of the source node v0 (including v0). Let F' be the weight matrix of the feedforward neural network, and F' be the dimension of the output features.

[0053] To facilitate comparison of coefficients across nodes with different computing power, the softmax function is used to normalize the absolute attention coefficients of all neighboring nodes of the source node, resulting in the relative attention coefficients of node n to the source node v0:

[0054]

[0055] Finally, the relative attention coefficients are used to calculate a linear combination of the corresponding features, serving as the final output feature for each computing node in the neighborhood. The calculated relative attention coefficients are then used to weight and sum the first-order neighbors of the source node, and the result is applied through the activation function σ to obtain the feature representation of the source node v0.

[0056]

[0057] Output of the source node's self-attention layer The probability that a neighboring node containing the source node v0 is the optimal next hop is calculated, and the neighbor with the highest probability is selected as the optimal next hop. Executing the graph attention mechanism once can only find the optimal next hop for the source node v0. Our goal is to find the path from the source node v0 to the destination node v. I The end-to-end path is defined as P{v0,v... I} = [v0, v1, v2, ..., v I It is a vertex sequence. Therefore, the optimal next hop v1 is regarded as the source node of the next self-attention layer, and a graph attention mechanism is computed for the new source node. This process is repeated, and each time the graph attention mechanism is executed, the probability that the first-order neighbor of the source node belongs to the optimal next hop is generated, and the neighbor node with the largest probability value is selected as the source node of the next self-attention layer, until the first-order neighbor of the source node contains the target computing power node v. IThe operation terminates upon reaching the specified time. Therefore, the graph attention mechanism, through deep learning of node information under different network conditions, treats the optimal next hop output of each self-attention layer as an intermediate result, primarily focusing on the optimal end-to-end path P = {P1, P2, ..., P...}. L The formation of}, P l ={P l,1 ,P l,2 ,...,P l,Q}

[0058] S23. To meet the performance requirements of diverse computing tasks, we select Y' and P such that the execution time T of the l-th computing task is... l The minimized number of computing nodes and their optimal end-to-end paths are the final result of UCPN computing resource scheduling, and the execution time T of the computing task is... l It equals the sum of the transmission time of the computation task from the source node to the destination computing node and the computation time on the destination computing node.

[0059] Finally, in step S3, we use netlogo to simulate the UCPN50 node network topology, such as... Figure 4 As shown in the diagram. Since information on the original computing resource distribution and requirements of UCPN is unavailable, we assume that cloud-side computing resources are relatively abundant, and the computing resources of edge-side and terminal-side computing nodes follow a normal distribution, with mean values ​​of 20GHz and 10GHz, respectively. Simultaneously, computing tasks are randomly generated at edge-side and terminal-side computing nodes, and the computing resource requirements of these tasks also follow a normal distribution, with a mean value of 3GHz. Wired communication is used between cloud-side and edge-side computing nodes, while edge-side computing nodes communicate via a mesh network consisting of power line carrier (HPLC) and low-power radio (RF). Communication between terminal-side and edge-side computing nodes is via WiFi. To simplify the experiment, only the RF and HPLC links in UCPN are simulated; the latency and reliability of other link types are randomly generated.

[0060] The proposed computational resource scheduling algorithm (HGcov-atten) mainly implements two major functions: scheduling of computational nodes and establishing end-to-end paths. We use these two functions as guides to combine different comparative algorithms to implement UCPN computational resource scheduling. Graph Convolutional Networks (GCNs), as efficient spectral networks, take the connection relationships (i.e., adjacency matrices) between UCPN node pairs as input and achieve convolution of irregular data through the spectral representation of the graph. Therefore, HGNN and GCN are compared to implement the computational node scheduling function. Shortest path is often used as a benchmark algorithm for network routing optimization; therefore, graph attention mechanisms are compared with shortest path algorithms to establish end-to-end paths. The comparison methods are summarized as Hypergraph Convolutional Shortest Path Algorithm (HGcov-shortest), Graph Convolutional Attention Network (Gcov-atten), and Graph Convolutional Shortest Path Algorithm (Gcov-shortest).

[0061] Figure 5 The proposed algorithm is compared with the final output of three benchmark algorithms, namely HGcov-shortest, Gcov-atten, and Gcov-shortest, i.e., the task execution time T. l A comparison was made, with the number of computational tasks set to 10, 20, 30, and 40 respectively. From Figure 5 In the results, we can observe that the execution time of both the proposed algorithm and the benchmark algorithm increases with the number of computational tasks. However, regardless of the number of tasks, the execution time of our proposed algorithm is lower than that of the benchmark scheme, demonstrating the advantages of our proposed scheme. The reason for this is that when allocating computing power nodes, we use a hypergraph to reconstruct hyperedge groups containing various information for UCPN. Therefore, hypergraph convolution can group computing power nodes with similar computing power and communication conditions into the same hyperedge based on the node's computing power and communication conditions, and only select computing power nodes with k hops or less, thereby minimizing the time spent on mining potential data interaction relationships and traversing all computing power nodes in UCPN during algorithm training. Furthermore, when using the graph attention mechanism to establish end-to-end paths for computational tasks, we consider three performance indicators: transmission rate, latency, and reliability, which to some extent avoids the congestion phenomenon that may occur in shortest path algorithms.

[0062] Figure 6This paper demonstrates the change in network computing resource utilization as the number of computing tasks increases when using the HGNN and GCN algorithms to allocate computing power nodes in UCPN. It can be seen that when the number of computing tasks is 10, the network resource utilization is low because the number of computing tasks is small, and computing resources are relatively abundant. As the number of computing tasks increases, the computing resource utilization of both algorithms also increases. The HGNN algorithm achieves higher network computing resource utilization because it fully utilizes the multi-dimensional resource attribute relationships between heterogeneous nodes in UCPN, finding computing power nodes that better match the computing power requirements of the computing tasks, thereby improving the utilization of computing power resources.

Claims

1. A method for scheduling computing resources in a ubiquitous computing network, characterized in that... Includes the following steps: Step S1: Establish a UCPN hypergraph structure consisting of vertices and hyperedges. The vertices are composed of heterogeneous nodes with differentiated resources. Correspondingly, construct hyperedge groups with dependency relationships, connectivity features, k-hop neighbors, and multi-dimensional resource attributes. Step S2: Based on the UCPN hypergraph structure obtained in Step S1, design the UCPN computing resource scheduling algorithm, namely the Hypergraph Convolutional Attention Network (HGcov-atten). This hypergraph convolutional attention network consists of two parts: hypergraph convolution and graph attention mechanism. The hypergraph convolution is responsible for generating UCPN heterogeneous node evaluation values, which reflect the computing resource supply capacity of heterogeneous nodes. The Q heterogeneous nodes with larger evaluation values ​​are selected as the target computing node set to achieve dynamic scheduling of computing resources. The graph attention mechanism reflects the importance of heterogeneous nodes through link transmission rate, latency, and reliability characteristics. Based on the importance of heterogeneous nodes, the optimal next hop is gradually found between the source node and the target computing node to establish an end-to-end path.

2. The ubiquitous computing network computing resource scheduling method according to claim 1, characterized in that... In step S1, the UCPN hypergraph structure Let ε be the vertex set, ε be the hyperedge set, and H be the incidence matrix; extract four types of relationships between vertices of UCPN to generate hyperedge groups, namely, interdependent hyperedge groups ε. ind ε, connectivity feature of hyperedge group con k-hop neighbor superedge group ε hop and multidimensional resource attribute hyperedge group ε res The four types of hyperedge groups are merged to form a hyperedge ε, where the k-hop neighbor hyperedge group ε is the most common. hop To obtain the set of shortest paths P between source and destination node pairs using the shortest path algorithm. shortest Specifically: S11, Using hyperedge groups ε with interdependent relationships ind Nodes located on the same power line are assigned to a superedge. For heterogeneous nodes communicating via power lines, the degree D of a node is defined as the number of power lines connected to it. Nodes with a degree D ≥ 3 are grouped together to form a power line with dependencies. Let C be the set of power lines with dependencies in UCPN, and let α be a binary variable. c,v To indicate whether node v is on power line c, the set of vertices on power line c is defined as follows: For a vertex set, use the hyperedge group ε with dependencies. ind Represented as: e ind ={N ind (c)|c∈C} S12, Hyperedge group ε using connectivity feature con Let CN be the connectivity matrix of UCPN, representing the number of communicable links between vertices; assume the set of vertices with connectivity β is defined as... where β∈[β min ,β max ], β min and β max Let ε be the minimum and maximum values ​​in CN, then the hyperedge group ε is determined using the connectivity feature. con Represented as: e con ={N con (β)|β∈[β min ,b max ]} S13, Using k-hop neighbors for the super-edge group ε hop , ε hop The goal is to find the relevant vertices of a computing power center point through the k-hop reachable locations in the graph structure, reflecting the distance traversed to transfer the computing task to a suitable computing power node; the k-hop neighborhood of vertex v is defined as... in If vertex v has a k-hop neighbor node u, then the superedge group ε of the k-hop neighbors is used. hop Represented as: S14. Hyperedge group ε using multidimensional resource attributes res , ε res The goal is to extract the differentiated multi-dimensional resource supply capabilities among UCPN computing power nodes, enabling... This represents two resource types: computing and communication. Represents the vertex set that can be provided The maximum value of each resource class distinguishes computing power nodes with different resource capabilities; the computing resources of all computing power nodes in UCPN are divided into ω. cp 0 intervals, ζ cp ∈{0,1 / ω cp ,2 / ω cp ,...,(ω cp -1) / ω cp } belongs to the interval The vertex subset is represented as Where τ cp =ζ cp ·ω cp +1, ζ cp and ω cp If all parameters are adjustable positive parameters, then the hyperedge group ε of multidimensional resource attributes is used. res Represented as:

3. The ubiquitous computing power network computing power resource scheduling method according to claim 2, characterized in that... In step S2, the hypergraph convolution of the hypergraph convolutional attention network is designed, and the specific process is as follows: Hypergraph Convolution: The Laplacian operator Δ is key to spectral hypergraph convolution. Δ is a square semidefinite matrix with only real eigenvalues, and its smallest eigenvalue is 0. This is achieved through eigenvalue decomposition: Δ = ΦΛΦ T This yields a non-negative eigenvalue matrix Λ = diag(λ1,...,λ). n ) and the orthogonal eigenvectors Φ = diag(Φ1,...,Φ n ), λ n Φ is an eigenvalue. n Orthogonal eigenvalues ​​are used; continuous eigenvectors Φ are applied as the spectral sequence signal x = (x1,...,x...). n The Fourier basis of ) is defined as follows: The Fourier transform is defined as follows: Inverse Fourier transform is The spectral convolution and convolution kernel g of the hypergraph signal x are represented as follows: (x*g) HG =Φ((Φ T g)⊙(Φ T x))=Φg θ (L)F T x Where ⊙ is the Hadamard product of the two matrices; Φ T g is a trainable filter for eigenvalues, using g θ (Λ) represents the time complexity of calculating the above equation. The 1stChebNet efficient computation method is adopted, utilizing the Chebyshev polynomial approximation filter and the product of the hypergraph signal: in This is the scaled eigenvector matrix, with a range of [-1, 1]. The scaling is performed to satisfy the Chebyshev polynomial T. k K of (x) th The condition for truncating the expansion, λ max Let θ be the spectral radius. k Let T be the coefficients of the Chebyshev polynomial; the recursive definition of the Chebyshev polynomial is T. k (x)=2xT k-1 (x)-T k-2 (x), where T0(x) = 1 and T1(x) = x; Given a UCPN hypergraph For a vertex Vertex degree is defined as d(v) = ∑ e∈ε w(e)H(v,e), for a hyperedge e∈ε, the degree of the hyperedge is defined as w(e)H(v,e). A UCPN hypergraph with N vertices Spatial features can form a Laplace matrix Δ = ID v -1 / 2 HWD e -1 H T D v -1 / 2 , where I is the identity matrix, and D is the identity matrix. v It is the vertex degree matrix, D e The hyper-edge degree matrix W is initialized to the identity matrix, resulting in hypergraph convolution operations: (x*g) HG ≈θ0x-θ1D v -1 / 2 HWD e -1 H T Dx -1 / 2 x To constrain the number of parameters and prevent overfitting, the parameter θ is defined as θ = -2θ1 = 2θ0D. v -1 / 2 H -1 W -1 D e -1 (H T ) -1 D v -1 / 2 Hypergraph convolution is further simplified to (x*g) HG ≈θD v -1 / 2 HWD e -1 H T D v -1 / 2 x The input to hypergraph convolution is the task feature X. ser1 Vertex resource characteristics X res With the correlation matrix H, the input matrix Used to describe the attributes and resource requirements of the l-th task, where C0 is the dimension of the task features; Among them, sr l da represents the source node of task l. l Indicates task size, pt l This represents the computational intensity of the task, i.e., the computational resources required to process 1 bit of data; the input matrix... Used to represent The resource features of each vertex, where C1 is the dimension of the vertex resource features; cp n The hypergraph represents the amount of computing resources that the nth node can provide; the incidence matrix H reveals the relationship between vertices and edges. Use one The dimensional correlation matrix H represents: X ser1 and X res Substituting these values ​​into the hypergraph convolution expression, we obtain the following expression: Y (t+1) =σ sigmoid (D v -1 / 2 HWD e -1 H T D v -1 / 2X(t)Θ(t) ) in Y is the input signal of the t-th layer. (t+1) Let σ represent the output signal of the t-th layer, and let σ be the activation function of the sigmoid function. sigmoid ; The output of the last layer of the hypergraph convolution is the evaluation value of the computing power nodes within k hops of the source node that generated the task. in For Y l Dimensions Y represents the maximum number of computing power nodes within k hops of the source node for the l-th computation task. l The elements are arranged in descending order to reflect the multidimensional resource supply capacity of UCPN computing nodes; The top Q computing nodes with the highest evaluation values ​​are selected as the target computing node set for the computing task, and the final output matrix is... Y' l The elements are arranged in descending order.

4. The ubiquitous computing power network computing power resource scheduling method according to claim 3, characterized in that... In step S2, the graph attention mechanism of the hypergraph convolutional attention network is designed, and the specific process is as follows: In the graph attention mechanism, the input matrix is ​​the target computing power node set Y' and the task communication requirement characteristics. and network performance characteristics Input matrix This is used to describe the QoS requirements of the l-th subtask, where C2 is the dimension of the task features. Where tr l Indicates the transmission rate requirement, dl l Indicates latency requirements, rl l Represents reliability requirements; input matrix C3 is used to describe the network performance characteristics of G links, where C3 is the dimension of the network performance characteristics. Mr e(m,n) Let lt represent the transmission rate of link e(m,n) between nodes m and n. e(m,n) pl represents the delay of link e(m,n). e(m,n) This indicates the reliability of link e(m,n); The specific process by which the graph attention mechanism achieves optimal next-hop selection is as follows: First, a shared weight matrix is ​​trained for all computing nodes in UCPN. This serves as a mapping between the input and output features of a node. During computation, the weight matrix is ​​first used to map the features of the first-order neighboring computing power nodes of the source node v0, and the results of these mappings are concatenated. Then, a feedforward neural network is input to map the concatenated vector to a single value. Finally, the LeakyReLU activation function is used to obtain the absolute attention coefficient of node n to the source node v0. Where || represents the concatenation operation. This indicates the importance of node n to the source node v0. It is the set of first-order nearest neighbors of the source node v0, which includes v0; Let F' be the weight matrix of the feedforward neural network, and F' be the dimension of the output features. To facilitate comparison of coefficients across nodes with different computing power, the softmax function is used to normalize the absolute attention coefficients of all neighboring nodes of the source node, resulting in the relative attention coefficients of node n to the source node v0: Finally, the relative attention coefficient is used to calculate a linear combination of the corresponding features, which serves as the final output feature for each computing node in the neighborhood. The first-order neighbors of the source node are weighted and summed using the calculated relative attention coefficients, and then the feature representation of the source node v0 is obtained through the activation function σ. Output of the source node's self-attention layer The probability that a neighboring node containing the source node v0 belongs to the optimal next hop is used to select the neighboring node with the highest probability value as the optimal next hop. Source node v0 to destination node v I The end-to-end path is defined as P{v0,v... I } = [v0, v1, v2, ..., v I As a vertex sequence, the optimal next hop v1 is considered as the source node of the next self-attention layer, and a graph attention mechanism is computed for the new source node. This process is repeated, and each time the graph attention mechanism is executed, the probability that the first-order neighbor node of the source node belongs to the optimal next hop is generated, and the neighbor node with the highest probability value is selected as the source node of the next self-attention layer, until the first-order neighbor node of the source node contains the target computing power node v. I Operation terminated at that time; Graph attention mechanisms learn node information under different network conditions through deep learning, treating the optimal next hop output of each self-attention layer as an intermediate result to form the optimal end-to-end path P = {P1, P2, ..., P...}. L }, where P l ={P l,1 ,P l,2 ,...,P l,Q } 5. The ubiquitous computing network computing resource scheduling method according to claim 4, characterized in that... Choose the optimal end-to-end path P from the output matrix Y' to ensure that the execution time T of the l-th computation task is minimized. l The minimized number of computing nodes and their optimal end-to-end paths are the final result of UCPN computing resource scheduling, and the execution time T of the computing task is... l It equals the sum of the transmission time of the computation task from the source node to the destination computing node and the computation time on the destination computing node.

Images