Edge network cluster division method considering scale constraint

Abstract

The invention relates to an edge network cluster division method considering scale constraints, and belongs to the technical field of edge computing networks. Comprising the following steps: establishing a gravitational relationship between nodes in an edge computing system by taking node comprehensive performance as mass, node similarity as a gravitational constant and node communication efficiency as a distance; constructing a cluster division optimization model taking maximization of the modularity of the edge node network as a target and taking a cluster scale as a constraint according to the gravitation relationship; carrying out cluster division on network nodes by using a Louvain algorithm with cluster scale constraint to form a cluster structure meeting service requirements and scale control requirements; and each node participates in resource collaboration and service scheduling in the cluster according to the cluster structure to realize distributed collaboration processing of calculation tasks. According to the invention, load balancing and efficient resource utilization are realized, the service execution efficiency is improved, and the service fluctuation caused by uneven load is reduced.

Description

Technical Field

[0001] The present invention relates to the field of edge computing network technology, and in particular to a method for partitioning edge network clusters taking into account scale constraints. Background Art

[0002] In recent years, with the widespread adoption of real-time applications such as IoT devices, industrial automation, and intelligent transportation, edge computing has become a key enabling technology for low-latency, high-reliability services. The massive influx of terminal devices has driven the clustered deployment of edge computing nodes. However, the optimal partitioning of node clusters within edge networks has become a key challenge in ensuring service performance and resource utilization. Cluster deployment requires consideration not only of node resource heterogeneity, service type differences, and communication quality, but also of load balancing across the entire network. It is crucial to avoid both the fragmentation of network resources that leads to reduced utilization and the over-centralization that leads to network congestion and increased management complexity. Achieving efficient and balanced cluster partitioning in complex and dynamic environments has become a core challenge in optimizing edge computing network topology.

[0003] However, existing clustering methods ignore the relationship between cluster size and network performance, lacking comprehensive consideration of factors such as scale constraints, communication overhead, and resource distribution. This leads to resource imbalance within the cluster and frequent cross-cluster communication. This, in turn, causes fluctuations in response latency, concentrated node load, and decreased resource utilization during service execution, impacting service quality and stability. To address these issues, there is an urgent need for a clustering method that can proactively adapt to changes in network status while constraining scale in edge environments, thereby achieving efficient and structurally balanced cluster organization in edge networks. Summary of the Invention

[0004] In view of the above analysis, the present invention aims to disclose an edge network cluster partitioning method considering scale constraints, so as to solve the problem that the existing technology is difficult to achieve cluster structure balance and service efficiency under dynamic network structure and resource heterogeneity conditions.

[0005] The present invention discloses a method for partitioning edge network clusters considering scale constraints, comprising: performing resource, availability, and load modeling analysis on all nodes in the edge computing system to obtain a multi-dimensional feature vector including resource characteristics, availability characteristics, and load characteristics;

[0006] Calculate the node's comprehensive performance based on the multi-dimensional feature vector, calculate the node similarity based on the node's service content, and calculate the node's communication efficiency based on the node's response delay;

[0007] The gravitational relationship between nodes is established with the node comprehensive performance as quality, node similarity as gravitational constant, and node communication efficiency as distance;

[0008] Based on the gravitational relationship between nodes, a cluster partitioning optimization model is constructed with the goal of maximizing the modularity of the edge node network and the cluster size as a constraint;

[0009] Use the Louvain algorithm with cluster size constraints to cluster network nodes and form a cluster structure that meets service needs and scale control requirements;

[0010] Each node participates in resource coordination and service scheduling within the cluster according to the cluster structure, realizing distributed collaborative processing of computing tasks.

[0011] Furthermore, it represents the node v in the edge computing system i and v j The gravitational function of the gravitational relationship between them is:

[0012]

[0013] in, and is the node v at time t i and v j Comprehensive performance; β i,j For node v i and v j The service similarity between i,j For node v i and v j The quality of the connection link between nodes; RTT0 is the response delay threshold between nodes. If node v i and v j If the response delay exceeds the threshold, the link is considered failed.

[0014] Furthermore, the node v in the current cycle t i The comprehensive performance is:

[0015]

[0016] Where, is the weight assigned to the jth feature; among them, the first feature is a resource feature vector consisting of sharable computing power, storage capacity, and network bandwidth; the second feature The availability feature vector is composed of the interaction success rate and online probability; the third feature is the load feature vector consisting of occupied storage space and bandwidth.

[0017] Furthermore, the first feature Expressed as:

[0018]

[0019] Among them, point vi The shared computing power in the current cycle is c i , the storage capacity is m i , the network bandwidth is b i ;c max , m max , b max are the maximum values ​​of computing power, storage space and network bandwidth in the entire network respectively; θ1 is the resource adjustment coefficient;

[0020] The second feature Expressed as:

[0021]

[0022] Among them, node v i The success rate of task execution in the current cycle is The online probability is θ2 is the availability adjustment coefficient;

[0023] The third characteristic Expressed as:

[0024]

[0025] Among them, node v i The occupied storage space in the current cycle is The occupied bandwidth is is the service status of the node, Indicates that the node is available, otherwise

[0026] Furthermore, node v i and v j The service similarity is expressed as:

[0027]

[0028] in, and Node v i and v j The content vector mean of v i,n 、v j,n Node v i 、v j Vector value for the nth service type request, where N is the total number of service types.

[0029] Furthermore, the node v i and v j The link quality between

[0030]

[0031] Among them, node v i and v j The response delay is RTT i,j , the shortest hop number is J i,j ;RTT0 is the response delay threshold.

[0032] Furthermore, the network modularity defined based on the node attraction relationship is calculated as follows

[0033]

[0034] Among them, Y is the modularity of the network when the cluster structure is C; W is the total weight of all edges in the network; v i 、v j 、v k are all nodes in the node set V; δ(v i ,v j ) is the cluster indicator function; |C i | is the number of nodes in the i-th cluster, C max and C min are the upper and lower bounds of the cluster size, respectively.

[0035] Furthermore, the Louvain algorithm with cluster size constraint integrates cluster splitting and merging mechanisms based on Louvain, so that clusters exceeding the upper limit of size are split into multiple subclusters, and smaller clusters are merged.

[0036] Furthermore, the Louvain algorithm with cluster size constraint includes:

[0037] Step S501: Initialize the network topology graph and calculate the content vector, multi-dimensional feature vector and node gravity of all edge nodes;

[0038] Step S502: each current node is regarded as a cluster to form an initial cluster structure;

[0039] Step S503: Calculate the modularity gain of each cluster when merged with other clusters. If the cluster size is higher than the lower threshold, merge the clusters with the largest modularity gain greater than 0. Otherwise, the cluster structure remains unchanged. If the cluster size is lower than the lower threshold, there is no need to require the modularity gain to be greater than 0. Merge the clusters with the largest modularity gain.

[0040] Step S504: Iterate step S503 until all cluster structures no longer change;

[0041] Step S505: Traverse all clusters to determine whether the cluster size exceeds the upper threshold. If so, divide the cluster into multiple subclusters using the Louvain algorithm.

[0042] Step S506: Repeat step S505 until the size of all clusters does not exceed the upper threshold;

[0043] Step S507: compress all nodes in each cluster into one node, merge the edge weights and calculate the edge weights of the new network structure;

[0044] Step S508: Repeat steps S502-507 until the network cluster structure no longer changes, and output the EC network cluster division result.

[0045] Furthermore, in step S503, cluster C is calculated. i with C j The combined modularity gain function is:

[0046]

[0047] in, For cluster C i Connect to cluster C j The sum of the link weights of the intermediate nodes; A j To connect to cluster C j The link weight of the middle node; R i To connect to cluster C i The link weight and .

[0048] The present invention can achieve the following beneficial effects:

[0049] The edge network cluster partitioning method considering scale constraints disclosed in the present invention solves the problems of network node heterogeneity, uneven resource distribution and diverse service requests in the edge computing environment by constructing a cluster organization with a reasonable structure and controllable scale, realizes load balancing and efficient resource utilization, improves service execution efficiency, and reduces service fluctuations caused by uneven load. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. Throughout the drawings, the same reference symbols denote the same components.

[0051] Figure 1 This is a flow chart of a method for dividing edge network clusters considering scale constraints in an embodiment of the present invention;

[0052] Figure 2 Flowchart of the Louvain algorithm with cluster size constraint in an embodiment of the present invention;

[0053] Figure 3 This is a pseudo code example diagram of the Louvain algorithm with cluster size constraints in an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, which constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0055] In one embodiment of the present invention, the edge computing system uses the simulated network constructed by the Routers-RF dataset as an example, which includes 2113 nodes and 6632 connecting edges. A node is set as the cloud center server, with a bandwidth of 10 5 Mbps, computing power is 1, storage capacity is 2 20 GB, initial availability 100%, initial load rate 0%; 32 nodes as backbone nodes, with bandwidth of 4000Mbps, computing power of 0.8, and storage capacity of 2 6 GB, initial availability 75%, initial load rate 0-75%; the remaining 2080 nodes participate in collaborative computing as edge nodes, with bandwidth of 2800-4000Mbps, computing power of 0.25-0.75, and storage capacity of 2 5 ~2 6 GB, initial availability 25-75%, initial load 0-50%. The total system runtime is set to 1 hour. During each simulation cycle, the user service request frequency does not exceed 3 times per second, and each request consists of 3 consecutive computing tasks. Task parameters follow a uniform distribution, with task data size ranging from 2 to 128 MB, transmission bandwidth 1 to 10 Mbps, a lower limit of 0.4 to 0.8 for the required computing power, and an upper limit of 25 to 60 ms for the required response delay.

[0056] This embodiment discloses a method for dividing edge network clusters taking into account scale constraints, such as Figure 1 As shown, the following steps are included:

[0057] Step S101: Perform resource, availability, and load modeling analysis on all nodes in the edge computing system to obtain a multi-dimensional feature vector including resource characteristics, availability characteristics, and load characteristics;

[0058] Step S102: Calculate the node comprehensive performance based on the multi-dimensional feature vector, calculate the node similarity based on the service content of the node, and calculate the node communication efficiency based on the node response delay;

[0059] Step S103: establish an attraction relationship between nodes with node comprehensive performance as quality, node similarity as attraction constant, and node communication efficiency as distance;

[0060] Step S104: Based on the gravitational relationship between nodes, a cluster partition optimization model is constructed with the goal of maximizing the modularity of the edge node network and the cluster size as a constraint;

[0061] Step S105: Use the Louvain algorithm with cluster size constraints to cluster the network nodes to form a cluster structure that meets service needs and scale control requirements;

[0062] Step S106: Each node participates in resource coordination and service scheduling within the cluster according to the cluster structure, thereby realizing distributed collaborative processing of computing tasks.

[0063] Specifically, this embodiment optimizes the node cluster division in the mobile edge computing network. Therefore, in step S101, the characteristics of the nodes accessing the network are modeled and analyzed, including:

[0064] 1) Modeling and analyzing the resource characteristics of network nodes;

[0065] Since edge nodes have resource heterogeneity in terms of computing power, storage space, and network bandwidth, and the physical dimensions and orders of magnitude of various resources are significantly different, in order to ensure that the relative weights of different resource dimensions are consistent when comprehensively expressed, the original resource values ​​need to be normalized. In this embodiment, a normalization strategy combining logarithmic function and maximum absolute value normalization is adopted to uniformly map the computing power, storage capacity, and available bandwidth of the nodes to the same numerical space, thereby ensuring that the resource characteristics are comparable and stable in subsequent modeling. The current cycle t of the system operation node v i The shared computing power is c i , the storage capacity is m i , the network bandwidth is b i , then node v i Resource feature vector It can be expressed as:

[0066]

[0067] Among them, c max , m max , b max are the maximum values ​​of computing power, storage space and network bandwidth in the entire network respectively; θ1 is the resource adjustment coefficient.

[0068] 2) Modeling and analyzing the availability characteristics of network nodes;

[0069] The availability characteristics of edge nodes are used to measure the stability of their continuous service provision during the system operation cycle. They are mainly affected by the success rate of task execution and the length of time the node is online. Considering the temporal nature of node behavior, relying solely on a single observation cycle may lead to large fluctuations in availability evaluation results. By combining the node's historical behavior with its current status, a weighted fusion availability feature representation is constructed. i In the current cycle t of the system operation, the task execution success rate is The online probability is Then its availability feature vector It can be expressed as:

[0070]

[0071] Where θ2 is the availability adjustment coefficient.

[0072] 3) Model and analyze the load characteristics of network nodes;

[0073] The load characteristics of edge nodes are used to characterize their resource usage. Due to the Matthew effect of "service optimization" in the process of service request, some nodes may become task clustering hotspots, forming a hot zone effect, leading to resource overload, which in turn causes node failure, network topology instability, and system robustness degradation. In order to reduce the risk of node overload and limit extreme resource inclination, a load characteristic modeling method for resource occupancy is proposed. In the current cycle t of system operation, node v i The storage space occupied is The occupied bandwidth is Then the load characteristic vector It can be expressed as:

[0074]

[0075] in, The service status of the node is determined based on the node storage and bandwidth usage thresholds to determine whether the node is currently available. Indicates that the node is available, otherwise

[0076] After the node feature model is established in step S101, in step S102, the node comprehensive performance is calculated based on the multi-dimensional feature vector, the node similarity is calculated based on the node service content, and the node communication efficiency is calculated based on the node response delay, including:

[0077] 1) Modeling and analyzing the comprehensive performance of network nodes

[0078] Specifically, the nodes are comprehensively evaluated based on multi-dimensional features, and the entropy weight theory is used to assign weights to the comprehensive performance evaluation of the nodes to reflect the amount of information of different features. The weighted sum of each feature is calculated to calculate the node v in the observation period t i The comprehensive performance is as follows

[0079]

[0080] in, is the weight assigned to the j-th feature.

[0081] 2) Modeling and analyzing the similarity between network nodes

[0082] In addition, the service similarity of the node content vector analysis should be considered to aggregate nodes with the same service request. The linear correlation between the service contents of two nodes is measured by the Pearson correlation coefficient. i,n 、v j,n Node v i 、v j The vector value of the request for the nth service type (calculated based on the number of requests for the service type by the node), node v i and v j The service similarity is expressed as:

[0083]

[0084] in, and Node v i and v j The content vector mean of .

[0085] 3) Modeling and analyzing the communication efficiency between network nodes

[0086] In addition, it is necessary to analyze the impact of the quality of the communication link between nodes on the collaborative relationship. i and v j The response delay is RTT i,j , the shortest hop number is J i,j , then the link quality between nodes is expressed as

[0087]

[0088] Among them, RTT0 is the response delay threshold. If the node v i and v j If the response delay exceeds the threshold, the link is considered failed.

[0089] Since the collaborative relationship between nodes in the edge network is affected by factors such as node resource status, service request similarity, and link quality, and the collaborative relationship has structural heterogeneity and local correlation, the collaborative relationship between network nodes is modeled as a weighted undirected graph structure based on a gravity graph in step S103 to quantify the collaborative attraction of different nodes to each other and provide basic data support for subsequent cluster division. Specifically, the node comprehensive performance is used as the quality, the node similarity is used as the gravity constant, and the node communication efficiency is used as the distance to construct the node v i and v j The gravitational function of the gravitational relationship between them is expressed as follows:

[0090]

[0091] Specifically, in step S104, based on the attraction relationship between nodes, with the goal of maximizing the modularity of the edge node network, and combined with the upper and lower bounds of the cluster size, a cluster partitioning optimization model is constructed. The modularity is used to measure the difference between the actual edge weight and the theoretical expected weight in the cluster structure. The larger the value, the tighter the connection between the nodes in the cluster and the better the partitioning effect. The network modularity defined based on the node attraction relationship is calculated as follows

[0092]

[0093] sqF min ≤|C i |≤C max ,C i ∈C

[0094] Among them, Y is the modularity of the network when the cluster structure is C; v i 、v j 、v k are all nodes in the node set V; δ(v i ,v j ) is the cluster indicator function; |C i | is the number of nodes in the i-th cluster, C max and C min are the upper and lower bounds of the cluster size, respectively.

[0095] The total weight W of all edges in the network is:

[0096]

[0097] Cluster indicator function δ(v i ,v j )=1 means node v i and v j belong to the same cluster, otherwise δ(v i ,v j )=0.

[0098] Specifically, in step S105, the Louvain algorithm with cluster size constraints is used to cluster the network nodes to form a cluster structure that meets service needs and scale control requirements;

[0099] The Louvain algorithm with cluster size constraint integrates cluster splitting and merging mechanisms based on Louvain, so that clusters exceeding the upper limit of size are split into multiple subclusters and smaller clusters are merged.

[0100] The Louvain algorithm process with cluster size constraints is as follows Figure 2 As shown, the steps are as follows:

[0101] Step S501: Initialize the network topology graph and calculate the content vector, multi-dimensional feature vector and node gravity of all edge nodes;

[0102] The content vector is {v i,1 ,v i,2 ,…,v i,N}; Multidimensional feature vector Node gravity F(v i ,v j );

[0103] Step S502: Treat each current node as a cluster to form an initial cluster structure C, C i ∈C represents the i-th cluster;

[0104] Step S503: Calculate the modularity gain of each cluster when merged with other clusters. If the cluster size is higher than the lower threshold, merge the clusters with the largest modularity gain greater than 0. Otherwise, do not change the cluster structure. If the cluster size is lower than the lower threshold, there is no need to require the modularity gain to be greater than 0. Merge the clusters with the largest modularity gain.

[0105] Step S504: Iterate step S503 until all cluster structures no longer change;

[0106] Step S505: Traverse all clusters to determine whether the cluster size exceeds the upper threshold. If so, divide the cluster into multiple subclusters using the Louvain algorithm.

[0107] Step S506: Repeat step S505 until the size of all clusters does not exceed the upper threshold;

[0108] Step S507: compress all nodes in each cluster into one node, merge the edge weights and calculate the edge weights of the new network structure;

[0109] Step S508: Repeat steps S502-507 until the network cluster structure no longer changes, and output the EC network cluster division result.

[0110] Specifically, in step S503, the cluster C is calculated i with C j The combined modularity gain function is:

[0111]

[0112] in, For cluster C i Connect to cluster C j The sum of the link weights of the midpoint nodes; To connect to cluster C j The sum of the link weights of the midpoint nodes; To connect to cluster C i The link weight and .

[0113] Specifically, in step S106, each edge node completes cluster affiliation selection based on its own resource status, service requests, and gravitational relationships with other nodes.

[0114] like Figure 3 This embodiment also provides a pseudocode diagram of the Louvain algorithm with cluster size constraints. This algorithm uses a greedy strategy to merge nodes with the highest modularity gain to form clusters. Clusters exceeding the upper limit of size are divided into multiple subclusters using the Louvain algorithm. Clusters below the lower limit of size can be merged without increasing modularity, ensuring that the cluster size constraint is met. The algorithm has a clear structure and is suitable for cluster partitioning in large-scale heterogeneous edge node environments.

[0115] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for clustering edge networks considering scale constraints, characterized in that: include: Perform resource, availability, and load modeling analysis on all nodes in the edge computing system to obtain a multi-dimensional feature vector including resource characteristics, availability characteristics, and load characteristics; Calculate the node's comprehensive performance based on the multi-dimensional feature vector, calculate the node similarity based on the node's service content, and calculate the node's communication efficiency based on the node's response delay; The gravitational relationship between nodes is established with the node comprehensive performance as quality, node similarity as gravitational constant, and node communication efficiency as distance; Based on the gravitational relationship between nodes, a cluster partitioning optimization model is constructed with the goal of maximizing the modularity of the edge node network and the cluster size as a constraint; Use the Louvain algorithm with cluster size constraints to cluster network nodes and form a cluster structure that meets service needs and scale control requirements; Each node participates in resource coordination and service scheduling within the cluster according to the cluster structure, realizing distributed collaborative processing of computing tasks.

2. The edge network cluster partitioning method considering scale constraints according to claim 1 is characterized in that: Represents node v in the edge computing system i and v j The gravitational function of the gravitational relationship between them is: in, and is the node v at time t i and v j Comprehensive performance; β i,j For node v i and v j The service similarity between i,j For node v i and v j The quality of the connection link between nodes; RTT0 is the response delay threshold between nodes. If node v i and v j If the response delay exceeds the threshold, the link is considered failed.

3. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: Node v in the current cycle t i The comprehensive performance is: Where, is the weight assigned to the jth feature; among them, the first feature is a resource feature vector consisting of sharable computing power, storage capacity, and network bandwidth; the second feature The availability feature vector is composed of the interaction success rate and online probability; the third feature is the load feature vector consisting of occupied storage space and bandwidth.

4. The edge network cluster partitioning method considering scale constraints according to claim 3 is characterized in that: The first feature Expressed as: Among them, point v i The shared computing power in the current cycle is c i , the storage capacity is m i , the network bandwidth is b i ;c max , m max , b max are the maximum values ​​of computing power, storage space and network bandwidth in the entire network respectively; θ1 is the resource adjustment coefficient; The second feature Expressed as: Among them, node v i The success rate of task execution in the current cycle is The online probability is θ2 is the availability adjustment coefficient; The third characteristic Expressed as: Among them, node v i The occupied storage space in the current cycle is The occupied bandwidth is is the service status of the node, Indicates that the node is available, otherwise 5. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: Node v i and v j The service similarity is expressed as: in, and Node v i and v j The content vector mean of v i,n 、v j,n Node v i 、v j Vector value for the nth service type request, where N is the total number of service types.

6. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: Then node v i and v j The link quality between them is expressed as: Among them, node v i and v j The response delay is RTT i,j , the shortest hop number is J i,j ;RTT0 is the response delay threshold.

7. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: The network modularity defined based on the node attraction relationship is calculated as: Where Y is the modularity of the network when the cluster structure is C; W is the total weight of all edges in the network; v i 、v j 、v k are all nodes in the node set V; δ(v i ,v j ) is the cluster indicator function; |C i | is the number of nodes in the i-th cluster, C max and C min are the upper and lower bounds of the cluster size, respectively.

8. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: The Louvain algorithm with cluster size constraint integrates cluster splitting and merging mechanisms based on Louvain, so that clusters exceeding the upper limit of size are split into multiple subclusters and smaller clusters are merged.

9. The edge network cluster partitioning method considering scale constraints according to claim 8, characterized in that: The Louvain algorithm with cluster size constraint includes: Step S501: Initialize the network topology graph and calculate the content vector, multi-dimensional feature vector and node gravity of all edge nodes; Step S502: each current node is regarded as a cluster to form an initial cluster structure; Step S503: Calculate the modularity gain of each cluster when merged with other clusters. If the cluster size is higher than the lower threshold, merge the clusters with the largest modularity gain greater than 0. Otherwise, the cluster structure remains unchanged. If the cluster size is lower than the lower threshold, there is no need to require the modularity gain to be greater than 0. Merge the clusters with the largest modularity gain. Step S504: Iterate step S503 until all cluster structures no longer change; Step S505: Traverse all clusters to determine whether the cluster size exceeds the upper threshold. If so, divide the cluster into multiple subclusters using the Louvain algorithm. Step S506: Repeat step S505 until the size of all clusters does not exceed the upper threshold; Step S507: compress all nodes in each cluster into one node, merge the edge weights and calculate the edge weights of the new network structure; Step S508: Repeat steps S502-507 until the network cluster structure no longer changes, and output the EC network cluster division result.

10. The edge network cluster partitioning method considering scale constraints according to claim 2, characterized in that: In step S503, cluster C is calculated i with C j The combined modularity gain function is: in, For cluster C i Connect to cluster C j The sum of the link weights of the intermediate nodes; A j To connect to cluster C j The link weight of the middle node; R i To connect to cluster C i The link weight and .

Images