Cloud laboratory resource allocation weight optimization method and system based on dynamic load balancing

Abstract

The invention provides a cloud laboratory resource allocation weight optimization method and system based on dynamic load balancing, and relates to the technical field of resource scheduling, and the method comprises the steps: firstly collecting the response time and throughput data of micro-service calling, and the power consumption and temperature data of a server cabinet, and then calculating the heat dissipation efficiency through the power consumption and temperature data, constructing a service dependence topology in combination with response time and throughput, and then performing space-time fusion on the service dependence topology and heat dissipation efficiency to obtain a coupling graph model; then analyzing the coupling graph model through an attention mechanism to obtain a hot spot prediction result, searching a Pareto optimal weight allocation set through multi-objective optimization by taking load balancing and heat dissipation energy efficiency maximization as targets, and then selecting weights according to a strategy, so that dynamic balancing can be realized, the resource utilization rate is further improved, and the performance and the heat dissipation effect are considered.

Description

Technical Field

[0001] This application relates to the technical field of resource scheduling, and in particular to a method and system for optimizing the allocation weight of cloud laboratory resources based on dynamic load balancing. Background Technology

[0002] In scenarios such as university research and enterprise R&D, the efficient operation of cloud laboratories relies on precise resource allocation strategies. Dynamic load balancing weight optimization is the core supporting technology. Furthermore, as the scale of microservice and artificial intelligence experiments expands, the requirements for resource utilization efficiency and operational stability become increasingly urgent, and the application prospects of related optimization methods are broad.

[0003] Currently, most mainstream cloud lab load balancing technologies employ load-aware scheduling and adaptive weight adjustment. Their core is to formulate resource allocation strategies based on performance indicators such as microservice response time and throughput in order to achieve load distribution.

[0004] However, these technologies can easily lead to overheating of equipment in high-load areas. This can reduce computing power by forcibly reducing the frequency, waste heat dissipation resources, and make it difficult to balance service performance and equipment stability. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for optimizing the resource allocation weight of cloud laboratories based on dynamic load balancing, so as to solve the problem of unreasonable resource allocation in cloud laboratories caused by the disconnect between resource allocation and heat dissipation status in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a cloud laboratory resource allocation weight optimization method based on dynamic load balancing, comprising: Collect response time and throughput data of inter-service call links running in the cloud laboratory, and capture power consumption and temperature data of each server rack; Based on the power consumption and temperature data, calculate the heat dissipation efficiency data of the corresponding server rack; Based on the response time and throughput data, a service dependency topology is constructed, and the service dependency topology is spatiotemporally fused with the heat dissipation efficiency data to construct a coupled graph model. An attention mechanism is used to perform evolutionary analysis on the coupled graph model to obtain hotspot prediction results; Based on the hotspot prediction results, with load balancing maximization and heat dissipation efficiency maximization as optimization objectives, a multi-objective optimization method is adopted to perform Pareto optimal search on the resource allocation weights to obtain the weight allocation set. According to the preset resource allocation strategy, the final weight value is selected from the weight allocation set to achieve dynamic load balancing of resources in the cloud laboratory.

[0007] Optionally, the step of using an attention mechanism to perform evolutionary analysis on the coupled graph model to obtain hotspot prediction results includes: The load feature vectors and heat dissipation features of each node in the coupled graph model are concatenated to form a fused feature vector; Based on the graph attention network, the attention score between each pair of adjacent nodes in the coupled graph model is calculated, and the weight distribution between nodes is generated according to the attention scores of all pairs of adjacent nodes. Based on the weight distribution, the fused feature vectors of all neighboring nodes of each node are weighted and aggregated to update the state representation of the node. The propagation process of the node state in the coupled graph model is simulated through multi-round graph convolution operations to obtain the evolution law of each node state. Based on the aforementioned evolution pattern, a graph search algorithm is used to identify the transmission path between changes in service load and fluctuations in heat dissipation status, and an interaction network is constructed based on the transmission path. Using the evolutionary pattern as input and the interaction network as constraints, the state of each node in the coupled graph model within a preset time period is deduced through a pre-trained time-series prediction model to obtain hotspot prediction results.

[0008] Optionally, the step of using the evolutionary pattern as input and the interaction network as constraints to deduce the state of each node in the coupled graph model within a preset future time period through a pre-trained time-series prediction model to obtain hotspot prediction results includes: The evolution pattern is represented as the state time series of each node, and input together with the adjacency matrix of the interaction network into the pre-trained time series prediction model. Based on the node state and interaction network of the previous moment, the time-series prediction model iteratively calculates the node state of the next moment to deduce the continuous change trajectory of the load level and heat dissipation efficiency of each node within a future preset time period. The continuously changing trajectory is scanned according to the preset load threshold and heat dissipation efficiency threshold. During the scanning process, nodes with load greater than the load threshold and heat dissipation efficiency less than the heat dissipation efficiency threshold are marked. Cluster all the nodes in the labeled nodes to form hotspot regions, and output the prediction results containing the hotspot regions and the corresponding transmission paths.

[0009] Optionally, based on the hotspot prediction results, and with load balancing and heat dissipation efficiency as optimization objectives, a multi-objective optimization method is used to perform a Pareto optimal search on the resource allocation weights to obtain a weight allocation set, including: Based on the hotspot prediction results, the overload ratio and underheating ratio of each node in the hotspot area are extracted, and combined with the topology of the conduction path, the comprehensive risk assessment value of each node is calculated as a risk weighting factor. Based on the aforementioned risk weighting factor, a first penalty coefficient and a second penalty coefficient are generated for each server rack node; Combining the total resource constraints of the cloud laboratory, the upper limit of the resource capacity of a single server rack, and the heat dissipation constraints of the heat dissipation system, a multi-objective optimization model with resource allocation weight as the decision variable is constructed. The multi-objective optimization model includes a first function that maximizes load balancing and a second function that maximizes heat dissipation efficiency. A Pareto sort-based multi-objective optimization method is adopted to perform Pareto front search within the feasible region defined by the multi-objective optimization model. During the search process, the function values ​​of the first function and the second function are corrected according to the first penalty coefficient and the second penalty coefficient, respectively, to obtain the target value. When the preset convergence condition is met, all weight schemes corresponding to the Pareto optimal solution set of the last iteration are used as the weight allocation set.

[0010] Optionally, based on the hotspot prediction results, the proportion of overload and underheating of each node within the hotspot area is extracted, and combined with the topology of the conduction path, a comprehensive risk assessment value for each node is calculated, including: Based on the hotspot prediction results, the load value and heat dissipation efficiency value of each node in the hotspot area are extracted; The load value and heat dissipation efficiency value are compared with the preset load safety threshold and heat dissipation efficiency safety threshold, respectively, to calculate the overload ratio and underheating ratio of each node. Based on the topology of the transmission path, the upstream influence source and downstream influence range of each node are traced and propagated along the path direction to calculate the path influence coefficient, which reflects the criticality of the node in the interactive network. The overload ratio, insufficient heat dissipation ratio, and path influence coefficient are weighted and fused to obtain the comprehensive risk assessment value for each node.

[0011] Optionally, the step of constructing a service dependency topology based on the response time and throughput data, and spatiotemporally fusing the service dependency topology with the heat dissipation efficiency data to construct a coupled graph model includes: Based on the response time and throughput data, the connection tightness between microservice nodes is determined to obtain the node connection strength; A node relationship network is constructed based on the node connection strength, and a service dependency topology is constructed based on the node relationship network; By combining the deployment records of the cloud laboratory microservices, the service dependency topology and the heat dissipation efficiency data are spatiotemporally fused to obtain a coupled graph model.

[0012] Optionally, the step of combining the deployment records of cloud lab microservices and spatiotemporally fusing the service dependency topology with the heat dissipation efficiency data to obtain a coupled graph model includes: By combining the deployment records of microservices in the cloud laboratory, each microservice in the service dependency topology is matched and associated with its corresponding server rack to obtain a mapping relationship; Based on the heat dissipation efficiency data, heat dissipation characteristics are added to each server rack node. Based on the mapping relationship, all the heat dissipation characteristics are integrated to construct an attribute graph that includes service call relationships and physical heat dissipation status. A correlation analysis between service load and heat dissipation status is performed on the attribute graph to obtain a coupled graph model.

[0013] Secondly, this application provides a cloud laboratory resource allocation weight optimization system based on dynamic load balancing, including: The data acquisition module is used to collect response time and throughput data of the call links between microservices running in the cloud laboratory, and to capture power consumption and temperature data of each server rack. The calculation module is used to calculate the heat dissipation efficiency data of the corresponding server rack based on the power consumption and temperature data. The construction module is used to construct a service dependency topology based on the response time and throughput data, and to perform spatiotemporal fusion of the service dependency topology and the heat dissipation efficiency data to construct a coupled graph model. The analysis module is used to perform evolutionary analysis on the coupled graph model using an attention mechanism to obtain hotspot prediction results; The search module is used to perform Pareto optimal search on the resource allocation weights based on the hotspot prediction results, with the optimization objectives of maximizing load balancing and maximizing heat dissipation efficiency, and to obtain the weight allocation set by adopting a multi-objective optimization method. The selection module is used to select the final weight value from the weight allocation set according to the preset resource allocation strategy, so as to realize dynamic load balancing of resources in the cloud laboratory.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, used to execute the computer program, implements the steps of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing as described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing described in the first aspect above.

[0016] The technical solution of this application has the following beneficial effects: This application collects microservice call data and server rack power consumption and temperature data to comprehensively understand the service operation status and equipment physical status, providing complete data support for subsequent optimization. Then, based on power consumption and temperature data, it calculates heat dissipation efficiency, transforming the equipment's heat dissipation capacity into an intuitive and measurable quantitative indicator that clearly reflects the rack's operational health. Next, it constructs a service dependency topology and integrates heat dissipation efficiency data to form a coupled graph model, breaking down the analysis barriers between services and hardware and enabling a correlation between their states. Subsequently, it uses an attention mechanism to analyze the coupled graph model to obtain hotspot prediction results and pre-determine high-load, poorly cooled risk areas, transforming passive handling into proactive prevention. Then, it uses a dual-objective optimization search to find the Pareto optimal weight allocation set, considering both load balancing and heat dissipation efficiency, to provide diverse optimization solutions adaptable to different needs. Finally, it selects the final weights according to the strategy to achieve dynamic balancing, enabling precise matching of resource allocation to real-time status, thereby improving the overall operational efficiency of the cloud laboratory.

[0017] Furthermore, this application strengthens the relationship between nodes by integrating the dual-dimensional features of load and heat dissipation, and combines them with time-series prediction capabilities to achieve early prediction of hotspot areas. The prediction results include both the location of risk nodes and the associated state propagation path, which can provide a more targeted decision-making basis for subsequent resource allocation optimization. Attached Figure Description

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

[0019] Figure 1 A flowchart illustrating a cloud laboratory resource allocation weight optimization method based on dynamic load balancing, provided for an embodiment of this application; Figure 2 This is a schematic diagram illustrating a specific implementation of step S104 provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a cloud laboratory resource allocation weight optimization system based on dynamic load balancing provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In the daily operation of cloud labs, the rationality of resource allocation directly affects experimental efficiency and equipment safety. However, current mainstream resource allocation optimization methods only focus on the "surface performance" of microservice response speed and data processing volume, and then distribute the workload by adjusting the resource allocation ratio. But these methods ignore the actual heat dissipation of server racks. That is, racks in high-load areas are prone to becoming "hot" due to continuous full-load operation. This may not only affect the experimental progress due to automatic speed reduction caused by overheating, but also cause the cooling system to consume additional resources, ultimately leading to the awkward situation of "using a lot of resources but not being efficient".

[0021] To address the aforementioned issues, this application proposes a cloud laboratory resource allocation weight optimization method based on dynamic load balancing. This method collects microservice operational data and simultaneously records server power consumption and temperature to calculate the heat dissipation capacity of each rack. Subsequently, it integrates service dependencies with heat dissipation data to form a comprehensive model that reflects both "service demand" and "hardware status." Intelligent analysis then predicts potential "high temperature, high load" risk areas. Based on this, the resource allocation ratio is optimized to balance the goals of "even resource utilization" and "preventing equipment overheating," ultimately achieving a precise match between resource allocation and actual operating status. This avoids server shutdowns due to overheating and ensures that every resource is used effectively, improving the overall operational efficiency of the cloud laboratory.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide a cloud laboratory resource allocation weight optimization method based on dynamic load balancing, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101: Collect response time and throughput data of the call links between microservices running in the cloud laboratory, and capture power consumption and temperature data of each server rack.

[0024] Among them, the inter-microservice call chain refers to the communication path between different microservices in the cloud laboratory to request and collaborate with each other to complete experimental tasks; the response time refers to the time from when a microservice sends a request to when it receives a response; and the throughput refers to the amount of data successfully transmitted between microservices per unit time. Power consumption data refers to the real-time or periodic quantitative record of the electrical power consumed by each component inside the server rack during operation. These components include processors, memory, hard drives, fans, etc. Temperature data refers to the real-time temperature measurement values ​​inside the server rack or in key heat dissipation areas, such as heat sinks, air inlets and outlets, and rack thermal aisles.

[0025] S102. Based on the power consumption and temperature data, calculate the heat dissipation efficiency data of the corresponding server rack.

[0026] Among them, heat dissipation efficiency data refers to quantitative information used to measure the heat dissipation capacity of server racks. It combines the energy consumption of the rack with the heat generation status of the equipment, and can intuitively reflect the actual ability of the rack to dissipate the heat generated during operation.

[0027] In this application example, power consumption and temperature data are used as the core basis. First, the standard temperature of the cabinet under normal load is determined as a reference benchmark. Then, the balance between heat generation and dissipation is judged by combining the difference between real-time power consumption and the difference between actual temperature. Finally, quantitative heat dissipation efficiency data is obtained by using common heat dissipation capacity calculation methods, such as correlating power consumption heat generation coefficient with temperature change rate.

[0028] This application transforms basic physical data such as power consumption and temperature of the cabinet into intuitive heat dissipation capacity indicators, which can provide a clear basis for subsequent judgment of the equipment's load-bearing limit and avoid the impact of insufficient heat dissipation on operation.

[0029] S103. Based on the response time and throughput data, construct a service dependency topology, and perform spatiotemporal fusion of the service dependency topology and the heat dissipation efficiency data to construct a coupled graph model.

[0030] Among them, the service dependency topology refers to a network structure that intuitively presents the relationship between microservices in the cloud laboratory. Nodes refer to microservices, and the connections between nodes refer to the dependency and collaboration relationships between services. It is used to clearly show the associated logic of service operation. The coupling graph model is a comprehensive model that integrates the microservice call relationship in the cloud laboratory with the heat dissipation status of the server rack. It can be used to simultaneously reflect the service operation requirements and hardware carrying capacity, thereby providing a unified data carrier for subsequent analysis.

[0031] In one specific implementation, step S103 includes: Step 1031: Based on the response time and throughput data, determine the connection tightness between microservice nodes to obtain the node connection strength.

[0032] Among them, node connection strength is a quantitative indicator that measures the degree of dependence between microservice nodes. Its value directly reflects the degree of mutual influence between two microservices in collaboration.

[0033] In this application example, a comprehensive calculation method based on service interaction quality can be used to determine the connection strength between microservice nodes. For example, the formula for calculating node connection strength is as follows: ,in, Refers to nodes With another node The connection strength, This is the weighting coefficient, for example, a value of 0.5. Refers to nodes With another node Real-time throughput between This refers to the average throughput among all nodes. Refers to nodes With another node Real-time response time This refers to the average response time among all nodes, used to determine the connection strength between nodes.

[0034] For example, a cloud lab deploys microservice nodes A, B, and C. Basic data collected by the service monitoring system shows that the response time between A and B is 0.2 seconds and the throughput is 100 megabytes per second; the response time between A and C is 0.5 seconds and the throughput is 60 megabytes per second. Based on this, using the collected response time and throughput data, the interaction data between all nodes is first statistically analyzed, resulting in an average throughput of 80 megabytes per second and an average response time of 0.35 seconds. Then, the connection strength between A and B is calculated using the node connection strength calculation formula, resulting in a connection strength of 1.5 between A and B, and a connection strength between A and C of 0.725.

[0035] Step 1032: Construct a node relationship network based on the node connection strength, and construct a service dependency topology based on the node relationship network.

[0036] In step 1032, node pairs with strengths higher than a preset threshold are connected with solid lines, and node pairs with strengths lower than the threshold are connected with dashed lines to construct a node relationship network. Then, the dashed line connections are removed, and a service dependency topology is finally formed.

[0037] For example, if the strength threshold is set to 0.3, since 1.5 and 0.725 are both higher than the strength threshold of 0.3, solid lines are used to connect AB and AC to build a node relationship network, and redundant links without interaction are eliminated to form a service dependency topology that includes AB and AC.

[0038] Step 1033: Combining the deployment records of the cloud laboratory microservices, the service dependency topology and the heat dissipation efficiency data are spatiotemporally fused to obtain a coupled graph model.

[0039] Among them, spatiotemporal fusion refers to the real-time correlation and integration of service dependencies and heat dissipation efficiency in the cloud laboratory scenario to ensure that the matching relationship between service status and hardware status conforms to the actual runtime sequence; the coupled graph model is a heterogeneous attribute graph model specifically designed for the analysis of the interaction relationship between microservice system load and rack heat dissipation status, including microservice nodes, server rack nodes, coupling edges, quantitative values ​​of heat dissipation status, edge weight values, and correlation coefficient thresholds.

[0040] Step 1033 may specifically include the following steps: Step a1: Based on the deployment records of the cloud lab microservices, match and associate each microservice in the service dependency topology with its corresponding server rack to obtain the mapping relationship.

[0041] The mapping relationship refers to the one-to-one correspondence between the microservice nodes in the service dependency topology and the server racks that actually carry the microservice, which is used to connect the "service logic layer" and the "hardware physical layer".

[0042] For example, by combining the cloud lab microservice deployment records, service dependency topology A is matched to rack 1, B to rack 2, and C to rack 3 to obtain a clear mapping relationship between microservices and racks.

[0043] Step a2: Combining the heat dissipation efficiency data, add heat dissipation features to each server rack node. Based on the mapping relationship, integrate all the heat dissipation features to construct an attribute graph that includes service call relationships and physical heat dissipation status.

[0044] Among them, the attribute graph refers to a network structure that contains node feature information. In the cloud laboratory scenario, the nodes in this graph contain information that represents the calling relationship of microservices, and also include the heat dissipation characteristics of the corresponding cabinet, which can realize the co-carrying of service and hardware information in the same graph; the heat dissipation efficiency data of the aforementioned cabinet can be used as heat dissipation characteristics.

[0045] For example, based on this mapping relationship, the heat dissipation efficiency data of racks 1 to 3, 0.8, 0.6 and 0.7, are attached as features to the corresponding microservice nodes and integrated to construct an attribute graph that includes service call relationships and heat dissipation status.

[0046] Step a3: Perform a correlation analysis between service load and heat dissipation status on the attribute graph to obtain a coupled graph model.

[0047] In step a3, based on the constructed attribute graph, an appropriate correlation analysis method is first selected according to the actual data characteristics of the cloud laboratory. Specifically, if the microservice call frequency and rack heat dissipation efficiency data show a linear relationship, Pearson correlation coefficient analysis is used; if the data does not follow a normal distribution or has a non-linear relationship, Spearman rank correlation analysis is used; if it is in a stage with small data volume and incomplete information, such as microservice cold start, grey relational analysis is used. Then, for each combination of microservice node and corresponding rack node in the attribute graph, the selected correlation analysis method is used to calculate the correlation coefficient between the two, and this coefficient is used to measure the strength of the correlation between microservice call frequency and rack heat dissipation efficiency. Next, a coefficient threshold is set, and the calculated correlation coefficient is compared with the coefficient threshold to remove abnormal correlation data with correlation coefficients lower than the coefficient threshold. These abnormal correlation data are noise that cannot reflect the true relationship between load and heat dissipation. For valid correlation data with correlation coefficients higher than the coefficient threshold, the absolute value of the correlation coefficient is used as the weight of the coupling edge between the microservice node and the rack node in the attribute graph. The weight directly reflects the tightness of the relationship between the two. Finally, the attribute graph after noise removal and weight assignment is a coupled graph model that can accurately reflect the interaction between microservice load and rack heat dissipation status.

[0048] For example, based on historical operational data from the cloud lab and empirical patterns in the correlation between load and heat dissipation, a threshold value was set for the preset coefficient in this scenario. Then, assuming the Pearson correlation coefficient analysis method was used to calculate the correlation coefficients between microservice A and its corresponding rack 1 as 0.85, microservice B and its corresponding rack 2 as 0.72, and microservice C and its corresponding rack 3 as 0.68, respectively, the three correlation coefficients were compared with the preset threshold of 0.3. Since 0.85 > 0.3, 0.72 > 0.3, and 0.68 > 0.3, it was determined that A, B, and C all have a valid correlation with the heat dissipation efficiency of their respective racks, and there is no noise data that needs to be removed. This ultimately forms a coupled graph model containing the heat dissipation efficiency of rack 1 (node ​​A) at 0.8, rack 2 (node ​​B) at 0.6, rack 3 (node ​​C) at 0.7, and the connection relationships between the nodes.

[0049] This application constructs a coupled graph model to link and integrate the previously isolated service operation status with the hardware heat dissipation status, providing a unified and reliable analysis platform for subsequent accurate analysis of the matching relationship between service load and hardware capacity.

[0050] S104. An attention mechanism is used to perform evolutionary analysis on the coupled graph model to obtain hotspot prediction results.

[0051] In one specific implementation, such as Figure 2 As shown, step S104 includes: Step 1041: Concatenate the load feature vectors and heat dissipation features of each node in the coupled graph model to form a fused feature vector.

[0052] Among them, the load feature vector refers to the data vector used to quantitatively characterize the operating load status of microservice nodes. The load feature vector includes core operating indicators such as response time and throughput, which are used to intuitively reflect the resource demand of the service. The fusion feature vector refers to the comprehensive vector formed by superimposing the load feature vector of the microservice node with the heat dissipation characteristics of the corresponding cabinet according to the dimensions, which is used to simultaneously carry information on the service operating status and hardware heat dissipation status.

[0053] For example, the coupled graph model of the cloud lab includes microservice nodes A, B, and C, and cabinets 1, 2, and 3 corresponding to each node. Then, the load feature vector of node A is extracted as [0.8, 0.9], and the heat dissipation feature of cabinet 1 is 0.8. These two vectors are combined in order to obtain the fused feature vector of node A as [0.8, 0.9, 0.8]. Then, using the same method, the fused feature vector of node B is obtained as [0.7, 0.8, 0.6], and the fused feature vector of node C is [0.6, 0.7, 0.7].

[0054] Step 1042: Based on the graph attention network, calculate the attention score between each pair of adjacent nodes in the coupled graph model, and generate the weight distribution between nodes according to the attention scores of all adjacent node pairs.

[0055] Among them, graph attention network refers to a network model based on attention mechanism that can focus on important relationships between nodes. It can highlight key connections by calculating attention scores between nodes to improve the accuracy of association analysis. Attention score is a quantitative value that measures the importance of the relationship between adjacent nodes in the graph. The higher the score, the more significant the influence of the node on the state of the target node. It can be used to quantify the dynamic correlation between load characteristics and heat dissipation characteristics between corresponding nodes. Weight distribution is a probability distribution formed based on the attention scores of all adjacent node pairs, which is used to clarify the contribution ratio of each adjacent node in state propagation.

[0056] In step 1042, a graph attention network is used to take the fused feature vector as input. First, the original attention score of each pair of adjacent nodes is calculated. Specifically, a learnable weight matrix is ​​used to perform a LeakyReLU activation on the result of the combination of the fused feature vectors of two adjacent nodes. This activation function makes the calculation result more in line with the actual needs, and the value obtained after processing is the original attention score of the two nodes.

[0057] After the original attention score is calculated, normalization is performed. Specifically, the original attention scores of a node and its neighboring nodes are converted into exponential form. This exponential value is then divided by the sum of the exponential values ​​of the original attention scores of all the neighboring nodes of that node. The resulting ratio is the normalized attention score. The normalized scores of all neighboring node pairs together constitute the weight distribution between nodes.

[0058] For example, using a graph attention network, a 3x2 learnable weight matrix is ​​set, and after training, the values ​​of this matrix are determined as follows: the first row has 0.2 and 0.3, the second row has 0.1 and 0.4, and the third row has 0.5 and 0.2. Then, when calculating the original attention scores of nodes A and B, the fused feature vectors of nodes A and B are first combined into a 2x3 sequence, and then the values ​​in the weight matrix are calculated corresponding to this sequence. That is, 0.8 multiplied by 0.2 results in 0.16, 0.9 multiplied by 0.1 results in 0.09, and 0.8 multiplied by 0.5 results in 0.4. These three results are added together to get 0.65. Similarly, 0.7 multiplied by 0.3 results in 0.21, 0.8 multiplied by 0.4 results in 0.32, and 0.6 multiplied by 0.2 results in 0.12. These three results are added together to get 0.65.

[0059] Then, LeakyReLU activation was applied to 0.65 to obtain the original attention score of nodes A and B as 0.65, and the original attention score of nodes A and C was calculated as 0.62 using the same method. Next, normalization is performed. The neighboring nodes of node A are B and C, and the exponent values ​​of their original scores are calculated as e^0.65 (approximately 2.138) and e^0.62 (approximately 1.858), respectively. These are then added together to obtain a total of 4.0. Subsequently, the normalized attention score between nodes A and B is calculated as 2.138 divided by 4.0, which is approximately 0.535, and the score between nodes A and C is 1.858 divided by 4.0, which is approximately 0.465. These two values ​​constitute the weight distribution of node A, and the weight distribution of other nodes is calculated in the same way.

[0060] Step 1043: Based on the weight distribution, perform weighted aggregation on the fused feature vectors of all neighboring nodes of each node to update the node's state representation, and simulate the propagation process of the node's state in the coupled graph model through multi-round graph convolution operations to obtain the evolution law of each node's state.

[0061] Among them, state representation refers to the vectorized description of the comprehensive state of a microservice node, which integrates its own characteristics and the influence of neighboring nodes, and is dynamically updated during the propagation process; graph convolution operation refers to the core operation used for node feature propagation and aggregation in the graph model, which can achieve iterative updates of node state through the fusion of neighborhood features; evolution law refers to the trend characteristics of the load state of microservice nodes and the heat dissipation state of corresponding racks over time, which is used to reflect the dynamic correlation pattern between the two.

[0062] In step 1043, based on the weight distribution, the fused feature vectors of all neighboring nodes of each node are calculated and aggregated according to their respective weight ratios, and the state representation of the node itself is updated with this aggregation result. Then, multiple rounds of graph convolution operations are performed, in which the state of all nodes is updated again according to the current weight distribution in each round, and the transmission process of the node state in the entire coupled graph model is simulated, and finally the evolution law of the state of each node over time is summarized.

[0063] For example, to calculate the updated state of node A based on weight distribution, the fused feature vectors of adjacent nodes B and C are first weighted and multiplied. Specifically, the vector values ​​of node B are multiplied by 0.535 to obtain [0.3745, 0.428, 0.321], and the vector values ​​of node C are multiplied by 0.465 to obtain [0.279, 0.3255, 0.3255]. Then, these two sets of multiplied values ​​are added together to obtain the aggregated result [0.6535, 0.7535, 0.6465]. Subsequently, the aggregated result is added to the fused feature vector of node A itself with half the weight to obtain the updated state of node A, represented as [0.7268, 0.8268, 0.7233]. After performing two rounds of such graph convolution operations, the evolution pattern is extracted. For example, the load state of node A gradually increases from 0.8 to 0.95 within 1 hour, while the heat dissipation efficiency of rack 1 gradually decreases from 0.8 to 0.68.

[0064] Step 1044: Based on the evolution pattern, use a graph search algorithm to identify the transmission path between service load changes and heat dissipation fluctuations, and construct an interaction network based on the transmission path.

[0065] Among them, graph search algorithm refers to the algorithm used to traverse the graph model and mine the path relationship between nodes. It can be a depth-first search algorithm, used to accurately locate the transmission path of the mutual influence between service load and heat dissipation status; transmission path refers to the transmission path of the fluctuation of heat dissipation status caused by changes in service load or vice versa, which can clearly show the specific link of the interaction between the two; interaction network refers to the network structure built based on transmission path, used to intuitively present the interaction relationship between service load and heat dissipation status.

[0066] In step 1044, a depth-first search algorithm is used to traverse the coupled graph model and, in combination with the evolution law, to find the specific transmission paths that cause fluctuations in heat dissipation caused by changes in service load or affect service load due to heat dissipation problems. Then, an interactive network that can intuitively show the interaction between the two is constructed using these transmission paths as a framework.

[0067] For example, by using a depth-first search algorithm to traverse the coupled graph model and combining the evolution law, it was found that the increase in load on node A would lead to a decrease in the heat dissipation efficiency of cabinet 1, which in turn would affect the heat dissipation status of the adjacent node B, thus forming a conduction path from A to B; similarly, the conduction path from A to C was identified, and an interactive network was constructed based on these two paths.

[0068] Step 1045: Using the evolution law as input and the interaction network as constraints, the state of each node in the coupled graph model within a preset time period is deduced through a pre-trained time-series prediction model to obtain the hotspot prediction result.

[0069] Among them, the time series prediction model refers to the model that learns the patterns of historical time series data to predict future trends. In the cloud laboratory scenario, this model is used to predict future changes in node load and heat dissipation. It can be a graph recurrent neural network or a spatiotemporal graph convolutional network.

[0070] Step 1045 may specifically include the following steps: Step b1: The evolution law is represented as the state time series of each node, and input together with the adjacency matrix of the interaction network into the pre-trained time series prediction model.

[0071] Among them, the state time series refers to the data sequence formed by arranging the state representations of nodes at different times in chronological order, which can be used to capture the temporal change characteristics of node states; the adjacency matrix is ​​a matrix used to characterize the connection relationship between nodes in the interaction network, and the matrix element values ​​contained therein are used to mark whether there is a correlation between nodes and the strength of the correlation; the pre-trained temporal prediction model is trained using the node state time series over a past period, the adjacency matrix of the interaction network, and the corresponding future node state labeled samples.

[0072] For example, the evolution patterns of each node are transformed into a state time series. The specific transformation process is as follows: extract the core feature values ​​such as load level and heat dissipation efficiency of the start and end time nodes within a 1-hour range from the evolution patterns of each node. Then, for the 6 time nodes divided according to 10-minute intervals, determine the feature values ​​of the middle 4 time points. That is, based on the difference between the values ​​of the start and end nodes, the value change is evenly distributed according to the 10-minute time interval to obtain 6 complete state representations of time nodes containing core features such as load level and heat dissipation efficiency. Then, arrange the state representations of these 6 time nodes in chronological order to form a state time series of 6 data points per hour, with one data point every 10 minutes. At the same time, extract the adjacency matrix of the interaction network. The position values ​​of A and B, A and C in the matrix are 1, which means that there is a connection. Other unconnected positions are 0. Then, input these two data together into a pre-trained time series prediction model. This model is trained using historical state data from the cloud lab over the past 3 months and corresponding future state labeled samples.

[0073] Step b2: Based on the node state and interaction network of the previous moment, the time series prediction model iteratively calculates the node state of the next moment to deduce the continuous change trajectory of the load level and heat dissipation efficiency of each node within a preset time period in the future.

[0074] Among them, the continuous change trajectory refers to the dynamic change curve of the load level and heat dissipation efficiency of the node in the future over a period of time, which is output by the time series prediction model and is used to show the continuous evolution process of the state.

[0075] For example, the time series prediction model extrapolates the state of the previous moment to obtain the continuous change trajectory of the next moment in the next hour: the load of node A increases from 0.95 to 1.02, and the heat dissipation efficiency decreases from 0.68 to 0.65; the load of node B increases from 0.7 to 0.88, and the heat dissipation efficiency remains at around 0.6 with slight fluctuations.

[0076] Step b3: Scan the continuously changing trajectory according to the preset load threshold and heat dissipation efficiency threshold, and mark the nodes where the load is greater than the load threshold and the heat dissipation efficiency is less than the heat dissipation efficiency threshold during the scanning process.

[0077] Among them, the load threshold refers to the maximum load threshold that a microservice node can bear, which is used to determine whether the service is overloaded; the heat dissipation efficiency threshold refers to the minimum efficiency threshold for normal heat dissipation of the server rack, which is used to determine whether the hardware is underheating.

[0078] For example, the preset load threshold is 0.9 and the heat dissipation efficiency threshold is 0.7. After scanning, it is found that the load of node A exceeds 0.9 and the heat dissipation efficiency is below 0.7 throughout the entire process, thus meeting the marking condition. Similarly, the load of node B reaches a maximum of 0.88, which does not exceed the threshold, so it is not marked.

[0079] Step b4: Cluster the nodes among all the marked nodes to form hotspot regions, and output the prediction results containing the hotspot regions and the corresponding propagation paths.

[0080] Hotspot areas refer to node clusters where the load level exceeds the load threshold and the heat dissipation efficiency is lower than the heat dissipation efficiency threshold. These are risk areas that require key optimization.

[0081] In step b4, a density clustering algorithm is used to group the marked abnormal nodes that are close to each other and closely related into one category to form hotspot areas. Finally, the hotspot prediction results, which include the location of the hotspot areas and the corresponding influence propagation paths, are output.

[0082] For example, the marked node A is clustered into a separate region, and combined with the previously identified propagation paths from A to B and from A to C, the final hotspot prediction result is: the hotspot region is cabinet 1 corresponding to node A, and the propagation path that may affect B and C.

[0083] This application, by accurately predicting in advance hotspot areas in cloud laboratories where overload and insufficient heat dissipation coexist, and the corresponding impact transmission paths, can provide targeted decision-making basis for subsequent resource allocation weight optimization, helping to avoid operational risks caused by the expansion of hotspot areas.

[0084] S105. Based on the hotspot prediction results, with the optimization objectives of maximizing load balancing and maximizing heat dissipation efficiency, a multi-objective optimization method is adopted to perform Pareto optimal search on the resource allocation weights to obtain the weight allocation set.

[0085] In one specific implementation, step S105 includes: Step 1051: Based on the hotspot prediction results, extract the overload ratio and underheating ratio of each node in the hotspot area, and calculate the comprehensive risk assessment value of each node in combination with the topology of the conduction path, as a risk weighting factor.

[0086] The overload ratio refers to the proportion of the actual load value of microservice nodes in the hotspot area that exceeds the preset load safety threshold, which is used to quantitatively characterize the degree of load overload risk of the nodes; the underload ratio refers to the proportion of the actual heat dissipation efficiency value of server racks in the hotspot area that is lower than the preset heat dissipation efficiency safety threshold, which is used to intuitively reflect the risk of insufficient heat dissipation capacity of the racks.

[0087] Step 1051 may specifically include the following steps: Step c1: Based on the hotspot prediction results, extract the load value and heat dissipation efficiency value of each node in the hotspot area.

[0088] For example, the hotspot area includes microservice node A and node B, corresponding to rack 1 and rack 2 respectively. The preset load safety threshold is 0.9 and the heat dissipation efficiency safety threshold is 0.7. Then, the actual load value of node A is extracted from the hotspot prediction results as 1.02 and the heat dissipation efficiency value is 0.65.

[0089] Step c2: Compare the load value and heat dissipation efficiency value with the preset load safety threshold and heat dissipation efficiency safety threshold respectively, and calculate the load over-limit ratio and heat dissipation under-limit ratio of each node.

[0090] In step c2, the overload ratio is equal to the ratio of the difference between the actual load value and the load safety threshold to the load safety threshold; similarly, the underheating ratio is equal to the ratio of the difference between the heat dissipation efficiency safety threshold and the actual heat dissipation efficiency value to the heat dissipation efficiency safety threshold. Both ratios are non-negative, and the larger the value, the higher the corresponding risk.

[0091] For example, the overload ratio of node A is (1.02-0.9) / 0.9, which is approximately 0.133; the underheating ratio is (0.7-0.65) / 0.7, which is approximately 0.071.

[0092] Step c3: Based on the topology of the transmission path, backtrack and propagate the upstream influence sources and downstream influence ranges of each node along the path direction, and calculate the path influence coefficient that reflects the criticality of the node in the interactive network.

[0093] The path influence coefficient is a quantitative indicator calculated based on the transmission path topology, used to reflect the transmission link of nodes in terms of service load and heat dissipation status. The implementation of this application does not specifically limit the calculation method of the path influence coefficient, and it can be set according to the actual situation.

[0094] For example, based on the transmission path topology from A to B in the hotspot prediction results, node A is the starting point of the path and has a direct impact on the downstream node B, so the path influence coefficient is set to 0.8; node B is the end of the path and is only affected by the upstream node, so the path influence coefficient is set to 0.5.

[0095] Step c4: Weight and fuse the overload ratio, the insufficient heat dissipation ratio and the path influence coefficient to obtain the comprehensive risk assessment value of each node.

[0096] Among them, the comprehensive risk assessment value refers to the quantitative result of the combined load overload ratio, heat dissipation insufficiency ratio and path impact coefficient, which can be used as a risk weighting factor for key control of high-risk nodes in the subsequent optimization process.

[0097] In step c4, the weighted fusion formula is used as follows: Calculate the comprehensive risk assessment value, where R refers to the comprehensive risk assessment value. This refers to the weight of the overload ratio. It is the weight of the proportion of insufficient heat dissipation. These are the weights of the path influence coefficient. The sum of the three values ​​is 1, and their specific values ​​can be adjusted according to the actual optimization tendencies of the cloud lab. This refers to the percentage of load exceeding limits. C refers to the proportion of insufficient heat dissipation, and C refers to the path influence coefficient. The resulting evaluation value is the risk weight factor.

[0098] For example, setting =0.4、 =0.3、 =0.3, and then the weighted fusion formula is used to calculate the comprehensive risk assessment value of node A as 0.4×0.133+0.3×0.071+0.3×0.8, with a final result of 0.314; the comprehensive risk assessment value of node B is 0.4×0+0.3×0.114+0.3×0.5, with a final result of 0.184, and these two values ​​are used as risk weight factors for cabinet 1 and cabinet 2, respectively.

[0099] Step 1052: Generate a first penalty coefficient and a second penalty coefficient for each server rack node based on the risk weighting factor.

[0100] The penalty coefficient refers to the correction parameter generated based on the comprehensive risk assessment value. It can be divided into a first penalty coefficient corresponding to the load balancing target and a second penalty coefficient corresponding to the heat dissipation and energy efficiency target. It is used to reduce the resource allocation priority of high-risk nodes during the optimization process.

[0101] In step 1052, a first penalty coefficient and a second penalty coefficient are generated using a linear mapping method based on the risk weight factor. The larger the value of the risk weight factor, the higher the values ​​of the two penalty coefficients, which can be used to correct the subsequent optimization target.

[0102] For example, using a linear mapping method with a coefficient of 1.2 to generate penalty coefficients, the first penalty coefficient for node A is 0.314 multiplied by 1.2, resulting in approximately 0.377. Similarly, the second penalty coefficient is the same as the first penalty coefficient, also 0.377. The first penalty coefficient for node B is 0.184 multiplied by 1.2, resulting in approximately 0.221, and the second penalty coefficient is also 0.221.

[0103] Step 1053: Combining the total resource constraints of the cloud laboratory, the upper limit of the resource capacity of a single server rack, and the heat dissipation constraints of the heat dissipation system, construct a multi-objective optimization model with resource allocation weight as the decision variable. The multi-objective optimization model includes a first function that maximizes load balancing and a second function that maximizes heat dissipation efficiency.

[0104] Among them, the multi-objective optimization model refers to a mathematical model that combines cloud laboratory resource constraints, uses resource allocation weights as decision variables, and simultaneously pursues the maximization of load balancing and heat dissipation efficiency, in order to standardize the optimization direction and boundary.

[0105] In step 1053, the total resource capacity of the cloud laboratory, the resource limit of a single server rack, and the maximum heat dissipation capacity of the heat dissipation system are collected. These conditions are used as optimization constraints to construct a multi-objective optimization model with resource allocation weights as decision variables. The model contains two core objective functions: the first function maximizes load balance by minimizing the load variance of each node, and the second function maximizes heat dissipation efficiency by maximizing the average heat dissipation efficiency of each rack.

[0106] For example, the total resource constraint of the cloud laboratory is set to 100 units of computing resources, the resource capacity limit of rack 1 is 40 units, the resource capacity limit of rack 2 is 35 units, and the heat dissipation constraint is that the heat dissipation efficiency of rack 1 is not less than 0.6 and the heat dissipation efficiency of rack 2 is not less than 0.6. Then, a multi-objective optimization model is constructed based on these conditions. The first function of the model is to maximize the difference between 1 and the load variance of each node, and the second function is to maximize the average heat dissipation efficiency of rack 1 and rack 2.

[0107] Step 1054: Using a Pareto sort-based multi-objective optimization method, perform Pareto front search within the feasible region defined by the multi-objective optimization model. During the search process, correct the function values ​​of the first function and the second function according to the first penalty coefficient and the second penalty coefficient, respectively, to obtain the target value.

[0108] In step 1054, a multi-objective optimization algorithm based on Pareto sorting, namely the NSGA-III algorithm, is adopted to initialize a batch of resource allocation weight schemes within the feasible region defined by the optimization model. The original values ​​of the first function and the second function are calculated for each scheme, and the original values ​​are corrected with the corresponding penalty coefficient. A higher penalty coefficient will reduce the objective value of the corresponding function. Then, a better weight scheme is selected through operations such as non-dominated sorting and crowding calculation, and the next round of iteration is entered.

[0109] For example, the NSGA-III algorithm is used to initialize 20 resource allocation weight schemes. The weight w1 of rack 1 and the weight w2 of rack 2 are used as decision variables, and the sum of w1 and w2 is not greater than 1. Taking one scheme w1=0.4 and w2=0.3 as an example, the original value of the first function is calculated to be 0.85 and the original value of the second function is 0.72. Then, the target value is corrected with a penalty coefficient. The target value of the first function is 0.85 minus the result of 0.4 multiplied by 0.377, and then minus the result of 0.3 multiplied by 0.221, and the final result is 0.633. The target value of the second function is 0.72 minus the result of 0.4 multiplied by 0.377, and then minus the result of 0.3 multiplied by 0.221, and the final result is 0.503. After multiple rounds of iteration, a better scheme is selected.

[0110] Step 1055: When the preset convergence condition is met, take all weight schemes corresponding to the Pareto optimal solution set of the last iteration as the weight allocation set.

[0111] Among them, the preset convergence condition refers to the preset criterion for stopping the optimization. It is usually based on the fact that the change of the optimal solution set during continuous iteration is less than the set value, so as to ensure the optimization efficiency and the stability of the results.

[0112] For example, the preset convergence condition is that the mean change of the objective function of the Pareto optimal solution set is less than 0.01 in three consecutive iterations. It is observed that when the iteration reaches the 8th round, the mean change in the three consecutive rounds is 0.008, and the convergence condition is met, so the iteration stops. The 12 Pareto optimal weight schemes obtained in this round are used as the weight allocation set, which includes several schemes with different optimization tendencies, such as w1=0.38, w2=0.32 and w1=0.42, w2=0.28.

[0113] This application incorporates the risk status of hotspot areas into the resource allocation weight optimization process, and obtains a multi-weight scheme that balances load balancing and heat dissipation efficiency through bi-objective optimization and Pareto search. This avoids overloading of high-risk nodes and provides comprehensive decision support for subsequent weight selection.

[0114] S106. According to the preset resource allocation strategy, select the final weight value from the weight allocation set to achieve dynamic load balancing of resources in the cloud laboratory.

[0115] Among them, the preset resource allocation strategy refers to the weight selection rules that the cloud laboratory has formulated in advance according to different operational needs. The preset resource allocation strategies include performance priority strategy, stability priority strategy, and balance priority strategy. The performance priority strategy is used to ensure service response efficiency when experimental tasks are urgent, the stability priority strategy is used to avoid heat dissipation failure when the equipment is under high load, and the balance priority strategy is used to balance performance and stability in cloud laboratory scenarios.

[0116] In this application example, using the weight allocation set as the selection range, the current operational needs of the cloud laboratory are first clarified, such as determining whether there are urgent experimental tasks or whether the server racks are under high load. Then, the corresponding preset resource allocation strategy is matched. If it is an urgent experimental task, a performance priority strategy is activated, and a weight scheme with better load balancing performance is selected from the weight allocation set. Alternatively, if the server is close to the heat dissipation threshold, a stability priority strategy is activated, and a weight scheme with more outstanding heat dissipation efficiency is selected. In a normal cloud laboratory scenario, a balance priority strategy is used to select a scheme with balanced performance for both objectives. Finally, after determining the weight scheme that meets the requirements, the final weight value is extracted and applied to the cloud laboratory's resource scheduling system to complete the dynamic update of resource allocation weights.

[0117] This application selects targeted weight values ​​to match actual needs, enabling precise resource allocation to fit the cloud laboratory scenario. This ensures efficient progress of experimental tasks while avoiding equipment operation risks, ultimately achieving dynamic load balancing of resources.

[0118] Figure 3 This application provides a schematic diagram of a specific implementation of a cloud laboratory resource allocation weight optimization system based on dynamic load balancing, as illustrated in the embodiments of this application. Figure 3 The system may include: The acquisition module 31 is used to collect response time and throughput data of the call links between microservices running in the cloud laboratory, and to capture power consumption and temperature data of each server rack.

[0119] The calculation module 32 is used to calculate the heat dissipation efficiency data of the corresponding server rack based on the power consumption and temperature data.

[0120] The construction module 33 is used to construct a service dependency topology based on the response time and throughput data, and to perform spatiotemporal fusion of the service dependency topology and the heat dissipation efficiency data to construct a coupled graph model.

[0121] Analysis module 34 is used to perform evolutionary analysis on the coupled graph model using an attention mechanism to obtain hotspot prediction results.

[0122] Search module 35 is used to perform Pareto optimal search on resource allocation weights based on the hotspot prediction results, with load balancing maximization and heat dissipation efficiency maximization as optimization objectives, and adopt a multi-objective optimization method to obtain a weight allocation set.

[0123] Selection module 36 is used to select the final weight value from the weight allocation set according to the preset resource allocation strategy, so as to realize dynamic load balancing of resources in the cloud laboratory.

[0124] The cloud laboratory resource allocation weight optimization system based on dynamic load balancing in this application embodiment is used to implement the aforementioned cloud laboratory resource allocation weight optimization method based on dynamic load balancing. Therefore, the specific implementation of the cloud laboratory resource allocation weight optimization system based on dynamic load balancing can be found in the embodiment section of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing mentioned above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0125] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described methods for optimizing the resource allocation weight of a cloud laboratory based on dynamic load balancing.

[0126] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for optimizing the resource allocation weights of a cloud laboratory based on dynamic load balancing.

[0127] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0128] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing.

[0129] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0130] The above provides a detailed description of a cloud laboratory resource allocation weight optimization method and system based on dynamic load balancing, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for optimizing resource allocation weights in a cloud laboratory based on dynamic load balancing, characterized in that, include: Collect response time and throughput data of inter-service call links running in the cloud laboratory, and capture power consumption and temperature data of each server rack; Based on the power consumption and temperature data, calculate the heat dissipation efficiency data of the corresponding server rack; Based on the response time and throughput data, a service dependency topology is constructed, and the service dependency topology is spatiotemporally fused with the heat dissipation efficiency data to construct a coupled graph model. An attention mechanism is used to perform evolutionary analysis on the coupled graph model to obtain hotspot prediction results; Based on the hotspot prediction results, with load balancing maximization and heat dissipation efficiency maximization as optimization objectives, a multi-objective optimization method is adopted to perform Pareto optimal search on the resource allocation weights to obtain the weight allocation set. According to the preset resource allocation strategy, the final weight value is selected from the weight allocation set to achieve dynamic load balancing of resources in the cloud laboratory.

2. The method according to claim 1, characterized in that, The step of using an attention mechanism to perform evolutionary analysis on the coupled graph model to obtain hotspot prediction results includes: The load feature vectors and heat dissipation features of each node in the coupled graph model are concatenated to form a fused feature vector; Based on the graph attention network, the attention score between each pair of adjacent nodes in the coupled graph model is calculated, and the weight distribution between nodes is generated according to the attention scores of all pairs of adjacent nodes. Based on the weight distribution, the fused feature vectors of all neighboring nodes of each node are weighted and aggregated to update the state representation of the node. The propagation process of the node state in the coupled graph model is simulated through multi-round graph convolution operations to obtain the evolution law of each node state. Based on the aforementioned evolution pattern, a graph search algorithm is used to identify the transmission path between changes in service load and fluctuations in heat dissipation status, and an interaction network is constructed based on the transmission path. Using the evolutionary pattern as input and the interaction network as constraints, the state of each node in the coupled graph model within a preset time period is deduced through a pre-trained time-series prediction model to obtain hotspot prediction results.

3. The method according to claim 2, characterized in that, The process involves using the evolutionary pattern as input and the interaction network as constraints, then using a pre-trained time-series prediction model to deduce the state of each node in the coupled graph model within a preset future time period, to obtain hotspot prediction results, including: The evolution pattern is represented as the state time series of each node, and input together with the adjacency matrix of the interaction network into the pre-trained time series prediction model. Based on the node state and interaction network of the previous moment, the time-series prediction model iteratively calculates the node state of the next moment to deduce the continuous change trajectory of the load level and heat dissipation efficiency of each node within a future preset time period. The continuously changing trajectory is scanned according to the preset load threshold and heat dissipation efficiency threshold. During the scanning process, nodes with load greater than the load threshold and heat dissipation efficiency less than the heat dissipation efficiency threshold are marked. Cluster all the nodes in the labeled nodes to form hotspot regions, and output the prediction results containing the hotspot regions and the corresponding transmission paths.

4. The method according to claim 1, characterized in that, Based on the hotspot prediction results, and with the optimization objectives of maximizing load balancing and maximizing heat dissipation efficiency, a multi-objective optimization method is employed to perform a Pareto optimal search on the resource allocation weights, resulting in a weight allocation set, including: Based on the hotspot prediction results, the overload ratio and underheating ratio of each node in the hotspot area are extracted, and combined with the topology of the conduction path, the comprehensive risk assessment value of each node is calculated as a risk weighting factor. Based on the aforementioned risk weighting factor, a first penalty coefficient and a second penalty coefficient are generated for each server rack node; Combining the total resource constraints of the cloud laboratory, the upper limit of the resource capacity of a single server rack, and the heat dissipation constraints of the heat dissipation system, a multi-objective optimization model with resource allocation weight as the decision variable is constructed. The multi-objective optimization model includes a first function that maximizes load balancing and a second function that maximizes heat dissipation efficiency. A Pareto sort-based multi-objective optimization method is adopted to perform Pareto front search within the feasible region defined by the multi-objective optimization model. During the search process, the function values ​​of the first function and the second function are corrected according to the first penalty coefficient and the second penalty coefficient, respectively, to obtain the target value. When the preset convergence condition is met, all weight schemes corresponding to the Pareto optimal solution set of the last iteration are used as the weight allocation set.

5. The method according to claim 4, characterized in that, Based on the hotspot prediction results, the overload ratio and underheating ratio of each node within the hotspot area are extracted, and combined with the topology of the conduction path, a comprehensive risk assessment value for each node is calculated, including: Based on the hotspot prediction results, the load value and heat dissipation efficiency value of each node in the hotspot area are extracted; The load value and heat dissipation efficiency value are compared with the preset load safety threshold and heat dissipation efficiency safety threshold, respectively, to calculate the overload ratio and underheating ratio of each node. Based on the topology of the transmission path, the upstream influence source and downstream influence range of each node are traced and propagated along the path direction to calculate the path influence coefficient, which reflects the criticality of the node in the interactive network. The overload ratio, insufficient heat dissipation ratio, and path influence coefficient are weighted and fused to obtain the comprehensive risk assessment value for each node.

6. The method according to claim 1, characterized in that, The process of constructing a service dependency topology based on the response time and throughput data, and then spatiotemporally fusing the service dependency topology with the heat dissipation efficiency data to construct a coupled graph model includes: Based on the response time and throughput data, the connection tightness between microservice nodes is determined to obtain the node connection strength; A node relationship network is constructed based on the node connection strength, and a service dependency topology is constructed based on the node relationship network; By combining the deployment records of the cloud laboratory microservices, the service dependency topology and the heat dissipation efficiency data are spatiotemporally fused to obtain a coupled graph model.

7. The method according to claim 6, characterized in that, The deployment records of the cloud lab microservices are combined, and the service dependency topology and the heat dissipation efficiency data are spatiotemporally fused to obtain a coupled graph model, including: By combining the deployment records of microservices in the cloud laboratory, each microservice in the service dependency topology is matched and associated with its corresponding server rack to obtain a mapping relationship; Based on the heat dissipation efficiency data, heat dissipation characteristics are added to each server rack node. Based on the mapping relationship, all the heat dissipation characteristics are integrated to construct an attribute graph that includes service call relationships and physical heat dissipation status. A correlation analysis between service load and heat dissipation status is performed on the attribute graph to obtain a coupled graph model.

8. A cloud laboratory resource allocation weight optimization system based on dynamic load balancing, characterized in that, include: The data acquisition module is used to collect response time and throughput data of the call links between microservices running in the cloud laboratory, and to capture power consumption and temperature data of each server rack. The calculation module is used to calculate the heat dissipation efficiency data of the corresponding server rack based on the power consumption and temperature data. The construction module is used to construct a service dependency topology based on the response time and throughput data, and to perform spatiotemporal fusion of the service dependency topology and the heat dissipation efficiency data to construct a coupled graph model. The analysis module is used to perform evolutionary analysis on the coupled graph model using an attention mechanism to obtain hotspot prediction results; The search module is used to perform Pareto optimal search on the resource allocation weights based on the hotspot prediction results, with the optimization objectives of maximizing load balancing and maximizing heat dissipation efficiency, and to obtain the weight allocation set by adopting a multi-objective optimization method. The selection module is used to select the final weight value from the weight allocation set according to the preset resource allocation strategy, so as to realize dynamic load balancing of resources in the cloud laboratory.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the cloud laboratory resource allocation weight optimization method based on dynamic load balancing as described in any one of claims 1 to 7.

Images