Machine room energy consumption and computing power balancing optimization method and system based on swarm intelligence

Abstract

The application provides a machine room energy consumption and computing power balancing optimization method and system based on swarm intelligence, relates to the technical field of data center management, and comprises the following steps: collecting server node real-time running data, constructing a multi-objective optimization function, predicting energy consumption and computing power change curves by using a recurrent neural network, generating an initial task allocation scheme, finding a global optimal allocation scheme by using a parallel ant colony algorithm, monitoring server running states and performing task migration and energy consumption adjustment, and realizing collaborative optimization of machine room energy consumption reduction and computing power balanced allocation, thereby improving service quality and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of data center management technology, and in particular to a method and system for optimizing data center energy consumption and computing power balance based on swarm intelligence. Background Technology

[0002] With the rapid development of information technology, the scale of data center computer rooms is constantly expanding and the number of servers is increasing dramatically. The problems of energy consumption and computing resource allocation in computer rooms are becoming increasingly prominent. As the core infrastructure for information processing and storage, the operating efficiency of modern data center computer rooms directly affects the operating costs and service quality of enterprises.

[0003] Server nodes in a data center generate different energy consumption and computing loads when processing different tasks. How to minimize data center energy consumption and balance the allocation of computing resources while ensuring service quality has become a hot research topic.

[0004] Traditional methods for optimizing data center energy consumption and computing power balance mainly rely on static configuration and manual intervention. They allocate tasks and control energy consumption of server nodes through preset rules. However, existing technologies still have problems in optimizing data center energy consumption and computing power balance, such as difficulty in ensuring balanced allocation of computing resources while reducing total energy consumption, inability to accurately capture the dynamic changes in server node energy consumption and computing load, limited prediction accuracy, low search efficiency, and easy getting trapped in local optima.

[0005] Therefore, a solution is urgently needed to address the problems existing in the current technology. Summary of the Invention

[0006] This invention provides a method and system for optimizing data center energy consumption and computing power balance based on swarm intelligence, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of this invention provides a method for optimizing data center energy consumption and computing power balance based on swarm intelligence, comprising:

[0008] Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized and the total energy consumption function, computing power allocation balancing function and service quality assurance function of the data center are constructed, and combined to obtain a multi-objective optimization function;

[0009] The time-series prediction unit of the recurrent neural network processes the standardized real-time running data to predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme based on the multi-objective optimization function and the task allocation weight.

[0010] Based on the initial task allocation scheme, multiple parallel ant colony computing spaces are constructed. The parallel ant colony computing spaces are divided into multiple sub-regions and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path, mark the optimal path, and filter the globally optimal task allocation scheme.

[0011] Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result.

[0012] In one alternative implementation,

[0013] Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized, and a total energy consumption function, a computing power allocation balancing function, and a service quality assurance function are constructed. These are combined to obtain a multi-objective optimization function, including:

[0014] Collect real-time operating data of server nodes in the computer room and perform standardized processing;

[0015] Based on the historical operating status of each server node in the data center, set the weights of energy consumption index, computing power index, and service quality index. Based on the weights of the energy consumption index, calculate the total energy consumption of the server nodes to obtain the total energy consumption function of the data center. Based on the weights of the computing power index, calculate the computing power distribution deviation of the task among the server nodes to obtain the computing power distribution balance function. Based on the weights of the service quality index, calculate the task processing latency of each server node to obtain the service quality assurance function.

[0016] The total energy consumption function of the computer room, the computing power allocation balancing function, and the service quality assurance function are normalized, and the combination coefficients of each function are dynamically adjusted. Based on the combination coefficients, the three functions are combined into a multi-objective optimization function.

[0017] In one alternative implementation,

[0018] The time-series prediction unit of the recurrent neural network processes standardized real-time operational data to predict the energy consumption and computing load change curves of server nodes, calculates dynamic task carrying capacity values, determines task allocation weights, and generates an initial task allocation scheme based on the multi-objective optimization function and the task allocation weights, including:

[0019] Standardized real-time operating data is divided into time windows, and the data within each time window is organized into a feature matrix. The feature matrix includes server node energy consumption data, computing load data, temperature data, and power consumption data. Based on the feature matrix, a feature time-series correlation coefficient matrix is ​​calculated.

[0020] A two-layer prediction model is constructed based on the temporal prediction unit in the recurrent neural network. The prediction model includes a first-layer temporal prediction unit and a second-layer temporal prediction unit. The first-layer temporal prediction unit processes the feature temporal correlation coefficient matrix to obtain the hidden state. The second-layer temporal prediction unit updates the network parameters based on the hidden state. The prediction model processes the running data of each server node to obtain the energy consumption change curve and computing load change curve of the server node in the next time period.

[0021] Based on the energy consumption change curve and the maximum energy consumption threshold, an energy consumption margin ratio sequence is calculated. Based on the computing power load change curve and the maximum computing power load threshold, a computing power load margin ratio sequence is calculated. Based on the average value of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combined with the first weighting coefficient and the second weighting coefficient, the dynamic task carrying capacity value of the server node is calculated.

[0022] The task allocation weights are calculated based on the dynamic task carrying capacity value and temperature parameters. The initial task allocation scheme is obtained by multiplying the task allocation weights by the multi-objective optimization function and finding the minimum value.

[0023] In one alternative implementation,

[0024] The predictive model is used to process the operational data of each server node to obtain the energy consumption change curve and computing load change curve of the server node for the next time period, including:

[0025] The running data corresponding to each server node is added to the prediction model, principal component analysis is performed to obtain a low-dimensional feature representation, and the low-dimensional feature representation is nonlinearly transformed by an activation function to obtain an initial hidden state. The initial hidden state is substituted into the neural ordinary differential equation, and the neural ordinary differential equation is solved to obtain a continuous-time hidden state function.

[0026] The mean and variance of the server node load data are calculated to generate a probability distribution function of the server load. A random perturbation sequence is sampled from the probability distribution function. The random perturbation sequence is superimposed with the continuous-time hidden state function to obtain a noisy state sequence.

[0027] Calculate the prediction error of the noisy state sequence, determine the step size of the numerical integration based on the prediction error, perform encrypted sampling of the noisy state sequence at the time point when the prediction error exceeds a preset threshold, and perform numerical integration on the sampling points using the fourth-order Runge-Kutta method to obtain the state evolution sequence.

[0028] The L2 norm value of the state evolution sequence and the energy loss value of the server node running data are calculated respectively. The L2 norm value and the energy loss value are added to obtain the total energy value. The total energy value is input into the second-layer time series prediction unit to calculate the gradient of the total energy value with respect to the network parameters. The network parameters are updated using the backpropagation algorithm. The causal relationship coefficient between the energy consumption change curve of the server node and the computing load change curve is calculated to generate a causal dependency graph between nodes.

[0029] Intervention operations are performed on the nodes in the causal dependency graph. The energy change value caused by each intervention operation is calculated. The energy change value is compared with the thermodynamic law of conservation of energy. The predicted data that violates the law of conservation of energy is corrected. The energy consumption change curve and computing load change curve of the server node in the next time period are output.

[0030] In one alternative implementation,

[0031] The parallel ant colony computation space is divided into multiple sub-regions, and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate. The convergence direction of the ant colony search path is controlled, and the optimal path is marked, including:

[0032] Each parallel ant colony computing space is divided into multiple sub-regions. A temperature decay coefficient is calculated based on the server nodes within each sub-region. The temperature decay coefficient is dynamically adjusted according to time changes and converted into a pheromone evaporation rate. Pheromones are updated in each parallel ant colony computing space in conjunction with the task allocation path within each sub-region. The pheromone retention amount is calculated based on the pheromone evaporation rate. The pheromone increments added by all individual ants on the search path are accumulated. The updated pheromone concentration is obtained by adding the pheromone retention amount to the pheromone increments.

[0033] In each parallel ant colony computing space, the transition probability of an individual ant choosing different task allocation schemes is calculated. Based on the transition probability, the convergence direction of the ant colony search path is controlled. The fitness value of the task allocation scheme in each parallel ant colony computing space is calculated. The task allocation schemes in each parallel ant colony computing space are sorted in descending order of fitness value. In each iteration cycle, the task allocation scheme with the highest fitness value is marked as the optimal path.

[0034] In one alternative implementation,

[0035] Converting the temperature decay coefficient into the pheromone evaporation rate, and performing pheromone updates in each parallel ant colony computation space in conjunction with the task allocation path within each sub-region, includes:

[0036] Collect the temperature decay coefficient of the server node in each sub-region, calculate the task processing capability difference coefficient of the server node in the sub-region based on the temperature decay coefficient, and set an independent pheromone evaporation rate for each sub-region according to the task processing capability difference coefficient.

[0037] Record the task allocation path of an individual ant in each sub-region, calculate the load balancing index and energy efficiency index of the task allocation path, weight the load balancing index and the energy efficiency index to obtain the path comprehensive score, map the path comprehensive score to a pheromone intensity adjustment coefficient according to a pre-set mapping relationship table, and calculate the path pheromone constraint threshold based on the pheromone intensity adjustment coefficient, wherein the path pheromone constraint threshold is dynamically adjusted with the path comprehensive score.

[0038] Detect the pheromone intensity of each task allocation path. If the pheromone intensity of a task allocation path exceeds the preset path pheromone constraint threshold, obtain the set of adjacent paths of that task allocation path.

[0039] Calculate the comprehensive path score for each path in the adjacent path set, calculate the pheromone redistribution weight based on the comprehensive path score, and distribute the pheromone amount exceeding the path pheromone constraint threshold to the paths in the adjacent path set according to the pheromone redistribution weight, thereby obtaining the pheromone increment added by the ant individual on the search path.

[0040] In one alternative implementation,

[0041] Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the operating status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching results, including:

[0042] Obtain historical task processing records of server nodes in the globally optimal task allocation scheme, extract energy consumption data and computing load data from the historical task processing records, and calculate the target energy consumption threshold and target computing load threshold for each server node.

[0043] Monitor the real-time energy consumption and real-time computing load of server nodes, calculate the difference between the real-time energy consumption and the target energy consumption threshold, and the difference between the real-time computing load and the target computing load threshold, and generate a judgment result based on the difference;

[0044] Based on the judgment result, the source server node and target server node for task migration are determined, the task migration operation is executed, and the operating frequency and power supply voltage of each server node are adjusted until the real-time energy consumption value and real-time computing load value of the source server node and the target server node converge to a preset range, thereby obtaining the optimal matching result.

[0045] A second aspect of this invention provides a data center energy consumption and computing power balancing optimization system based on swarm intelligence, comprising:

[0046] The first unit is used to collect real-time operating data of server nodes in the computer room, standardize the real-time operating data, and construct the total energy consumption function, computing power allocation balancing function, and service quality assurance function of the computer room, and combine them to obtain a multi-objective optimization function.

[0047] The second unit is used to process standardized real-time running data according to the time-series prediction unit of the recurrent neural network, predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme according to the multi-objective optimization function and the task allocation weight.

[0048] The third unit is used to construct multiple parallel ant colony computing spaces based on the initial task allocation scheme, divide the parallel ant colony computing space into multiple sub-regions and set a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path and mark the optimal path, and filter the globally optimal task allocation scheme.

[0049] The fourth unit is used to calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result.

[0050] A third aspect of the present invention,

[0051] An electronic device is provided, comprising:

[0052] processor;

[0053] Memory used to store processor-executable instructions;

[0054] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] Fourth aspect of the embodiments of the present invention,

[0056] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0057] This invention collects real-time operational data from server nodes and constructs a multi-objective optimization function. Combined with the time-series prediction capabilities of recurrent neural networks, it accurately predicts trends in server energy consumption and computing load, enabling proactive adjustments to task allocation. This effectively reduces overall data center energy consumption and improves energy efficiency. The innovative mechanism of using a parallel ant colony algorithm combined with a temperature decay coefficient to dynamically adjust pheromone evaporation rates significantly improves the search efficiency and optimization quality of task allocation schemes, preventing the algorithm from getting trapped in local optima. This ensures that a globally optimal task allocation scheme can still be found even under complex and variable server load environments. By setting target energy consumption and computing load thresholds, the invention monitors server operating status in real time and performs task migration and energy consumption adjustment, achieving dynamic balance of data center resources. This not only guarantees service quality but also improves the balanced utilization of computing resources, extends the lifespan of server equipment, and reduces maintenance costs. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the data center energy consumption and computing power balancing optimization method based on swarm intelligence, as described in an embodiment of the present invention.

[0059] Figure 2 This is a comparison chart of energy consumption prediction errors for the data center energy consumption and computing power balancing optimization method based on swarm intelligence in an embodiment of the present invention.

[0060] Figure 3 This is a comparison diagram of pheromone distribution and path convergence of the data center energy consumption and computing power balancing optimization method based on swarm intelligence in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1This is a flowchart illustrating the data center energy consumption and computing power balancing optimization method based on swarm intelligence, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized and the total energy consumption function, computing power allocation balancing function and service quality assurance function of the data center are constructed, and combined to obtain a multi-objective optimization function;

[0065] The time-series prediction unit of the recurrent neural network processes the standardized real-time running data to predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme based on the multi-objective optimization function and the task allocation weight.

[0066] Based on the initial task allocation scheme, multiple parallel ant colony computing spaces are constructed. The parallel ant colony computing spaces are divided into multiple sub-regions and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path, mark the optimal path, and filter the globally optimal task allocation scheme.

[0067] Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result.

[0068] In one alternative implementation,

[0069] Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized, and a total energy consumption function, a computing power allocation balancing function, and a service quality assurance function are constructed. These are combined to obtain a multi-objective optimization function, including:

[0070] Collect real-time operating data of server nodes in the computer room and perform standardized processing;

[0071] Based on the historical operating status of each server node in the data center, set the weights of energy consumption index, computing power index, and service quality index. Based on the weights of the energy consumption index, calculate the total energy consumption of the server nodes to obtain the total energy consumption function of the data center. Based on the weights of the computing power index, calculate the computing power distribution deviation of the task among the server nodes to obtain the computing power distribution balance function. Based on the weights of the service quality index, calculate the task processing latency of each server node to obtain the service quality assurance function.

[0072] The total energy consumption function of the computer room, the computing power allocation balancing function, and the service quality assurance function are normalized, and the combination coefficients of each function are dynamically adjusted. Based on the combination coefficients, the three functions are combined into a multi-objective optimization function.

[0073] Data acquisition modules deployed in the data center collect real-time operational data from server nodes. The collected data includes, but is not limited to, parameters such as CPU utilization, memory usage, network throughput, storage I / O, power consumption, and temperature. For example, in a data center with 100 servers, each server generates approximately 20 monitoring metrics per minute, requiring the system to process approximately 2000 data points per minute.

[0074] The collected raw data has inconsistent dimensions and needs to be standardized. Standardization uses a maximum-minimum normalization method to convert each indicator to the range [0, 1]. Taking CPU utilization as an example, if the raw data range is 0%-100%, the standardized value is the original value divided by 100. For power consumption data, if a server's power consumption ranges from 100W to 500W, standardization maps 100W to 0, 500W to 1, and other values ​​are mapped proportionally.

[0075] The system assigns three weighted metrics based on the historical operating status of each server node. For energy consumption metrics, the system analyzes the energy efficiency performance of each server over the past 30 days, assigning lower weights to servers with higher energy efficiency ratios and higher weights to servers with lower energy efficiency ratios. For example, server A with an energy efficiency ratio of 2.5 is assigned a weight of 0.4, while server B with an energy efficiency ratio of 1.8 is assigned a weight of 0.7.

[0076] The weighting of computing power metrics is determined based on the server's computing capabilities and stability. Servers with strong and stable computing capabilities receive higher computing power weights, and vice versa. For example, a server configured with 32 cores, 128GB of memory, and a stable uptime of over 99.9% may receive a computing power weight of 0.9, while a server with 16 cores, 64GB of memory, and a stable uptime of 98% may receive a computing power weight of 0.6.

[0077] Service Quality of Service (SQI) weights are determined based on the server's historical latency performance in processing tasks. Servers with shorter response times and faster processing speeds receive higher SQI weights, and vice versa. For example, a server with an average response time of 10ms might receive a SQI weight of 0.85, while a server with an average response time of 50ms might receive a SQI weight of 0.5.

[0078] Based on the aforementioned weights, three key functions are constructed. The first key function is the total energy consumption function of the data center, which is obtained by calculating the weighted sum of the energy consumption of each server node. Assuming the data center has n servers, the standardized energy consumption of the i-th server is Ei, and the energy consumption index weight is Wi, then the total energy consumption function of the data center is the sum of the products of the energy consumption of each server and its weight. For a small data center with 10 servers, if the standardized energy consumption of each server is [0.7, 0.8, 0.5, 0.6, 0.9, 0.4, 0.7, 0.8, 0.6, 0.5], and the corresponding energy consumption index weights are [0.8, 0.7, 0.6, 0.7, 0.9, 0.5, 0.7, 0.8, 0.6, 0.5], then the calculated total energy consumption function value is 4.31.

[0079] The computing power allocation balancing function is obtained by calculating the computing power allocation deviation of tasks among server nodes. The system calculates the ideal computing power ratio that each server should be allocated, calculates the deviation between the actual allocation and the ideal allocation, multiplies the deviation by the computing power index weights, and sums them. For example, for the above 10 servers, if the ideal computing power allocation ratio is [0.12, 0.11, 0.09, 0.10, 0.13, 0.08, 0.10, 0.12, 0.09, 0.06], while the actual allocation ratio is [0.15, 0.09, 0.08, 0.12, 0.14, 0.07, 0.11, 0.10, 0.08, 0.06], and the computing power index weights are [0.9, 0.8, 0.7, 0.8, 0.9, 0.6, 0.8, 0.9, 0.7, 0.6], then the calculated computing power allocation balancing function value is 0.0294.

[0080] The Quality of Service (QoS) assurance function is obtained by calculating the task processing latency of each server node. The system determines the expected response time of each server, calculates the deviation between the actual response time and the expected response time, multiplies the deviation by the QoS indicator weights, and sums them. For example, if the standardized latency deviation of 10 servers is [0.05, 0.08, 0.12, 0.07, 0.04, 0.15, 0.09, 0.06, 0.11, 0.14], and the QoS indicator weights are [0.9, 0.85, 0.75, 0.8, 0.9, 0.7, 0.8, 0.85, 0.75, 0.7], then the calculated QoS assurance function value is 0.0783.

[0081] To construct a multi-objective optimization function, the three functions mentioned above are normalized. The normalization process uses the maximum and minimum values ​​of historical data as a benchmark, mapping the outputs of the three functions to the same numerical range. For example, if the historical range of the total energy consumption function for the computer room is [3.5, 5.2], then the current value of 4.31 is normalized to 0.4765.

[0082] The system dynamically adjusts the combination coefficients of each function based on the current operating status of the data center. During periods of high load, more emphasis may be placed on service quality, increasing the weight of the service quality assurance function; during periods of low load, more emphasis may be placed on energy efficiency, increasing the weight of the total data center energy consumption function. For example, during peak daytime hours on weekdays, the combination coefficients might be set to energy consumption: computing power: service quality = 0.3:0.3:0.4; while during off-peak hours at night, the combination coefficients might be adjusted to 0.5:0.3:0.2.

[0083] The system combines three normalized functions into a multi-objective optimization function based on combination coefficients. Assuming the normalized total energy consumption function of the data center is 0.4765, the computing power distribution balancing function is 0.3520, the service quality assurance function is 0.2890, and the combination coefficients are [0.4, 0.3, 0.3], then the multi-objective optimization function value is 0.3759.

[0084] In this embodiment, the data center can reduce energy consumption and balance the allocation of computing resources while ensuring service quality, thereby improving overall operating efficiency. By minimizing the value of the multi-objective optimization function, the data center achieves comprehensive optimization of energy consumption, computing power allocation, and service quality, providing a solid theoretical foundation for subsequent optimization.

[0085] In one alternative implementation,

[0086] The time-series prediction unit of the recurrent neural network processes standardized real-time operational data to predict the energy consumption and computing load change curves of server nodes, calculates dynamic task carrying capacity values, determines task allocation weights, and generates an initial task allocation scheme based on the multi-objective optimization function and the task allocation weights, including:

[0087] Standardized real-time operating data is divided into time windows, and the data within each time window is organized into a feature matrix. The feature matrix includes server node energy consumption data, computing load data, temperature data, and power consumption data. Based on the feature matrix, a feature time-series correlation coefficient matrix is ​​calculated.

[0088] A two-layer prediction model is constructed based on the temporal prediction unit in the recurrent neural network. The prediction model includes a first-layer temporal prediction unit and a second-layer temporal prediction unit. The first-layer temporal prediction unit processes the feature temporal correlation coefficient matrix to obtain the hidden state. The second-layer temporal prediction unit updates the network parameters based on the hidden state. The prediction model processes the running data of each server node to obtain the energy consumption change curve and computing load change curve of the server node in the next time period.

[0089] Based on the energy consumption change curve and the maximum energy consumption threshold, an energy consumption margin ratio sequence is calculated. Based on the computing power load change curve and the maximum computing power load threshold, a computing power load margin ratio sequence is calculated. Based on the average value of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combined with the first weighting coefficient and the second weighting coefficient, the dynamic task carrying capacity value of the server node is calculated.

[0090] The task allocation weights are calculated based on the dynamic task carrying capacity value and temperature parameters. The initial task allocation scheme is obtained by multiplying the task allocation weights by the multi-objective optimization function and finding the minimum value.

[0091] The standardized real-time operational data is divided into time windows, with the length of each window determined based on the server task processing cycle. Within each time window, the system continuously collects operational status data from server nodes, including real-time energy consumption, CPU utilization reflecting computing load, temperature data collected by temperature sensors, and power consumption data calculated by power consumption calculations. This data is arranged in sampling time sequence to construct an m×n feature matrix, where m represents the number of sampling points and n represents the feature dimension. Pearson correlation coefficients are calculated for each indicator in the feature matrix to obtain a feature time-series correlation coefficient matrix reflecting the correlation strength between the indicators.

[0092] The prediction model employs a two-layer recurrent neural network structure. The first-layer temporal prediction unit receives a feature temporal correlation coefficient matrix, controls the proportion of data input at the current time step through an input gate, controls the retention of historical information through a forget gate, and controls the output proportion of the current state. These gating units work together to convert the input data into a hidden state vector containing temporal features. The second-layer temporal prediction unit receives the hidden states from the first layer and dynamically adjusts the network parameters, including the weight matrix and bias vector, based on the state information. Parameter updates utilize a backpropagation algorithm to minimize prediction error. After model training, inputting data from the current time window allows prediction of the energy consumption and computing load trends of server nodes in the next cycle.

[0093] Dynamic task carrying capacity calculation involves obtaining the predicted energy consumption change curve. The predicted value at each time point on the curve is subtracted from the preset maximum energy consumption threshold, and then divided by the maximum threshold to obtain a ratio sequence reflecting the remaining energy capacity. Similarly, the predicted value of the computing power load change curve is subtracted from the maximum load threshold and normalized to obtain a ratio sequence of remaining computing power capacity. The arithmetic mean of the two sequences over the prediction time period is calculated, representing the average energy consumption margin ratio and the average computing power margin ratio, respectively. Energy consumption weighting coefficients and computing power weighting coefficients are set, and a weighted sum is used to obtain a comprehensive dynamic task carrying capacity value. A larger value indicates more abundant remaining resources on the server nodes.

[0094] The task allocation scheme is generated based on the calculated dynamic task carrying capacity value, combined with an exponential decay function reflecting the impact of server temperature (the higher the temperature, the faster the decay), to calculate the task allocation weight for each node. The task allocation weights are then multiplied by a multi-objective optimization function that considers minimizing energy consumption and load balancing. By solving the minimum value problem, an initial task allocation ratio scheme that satisfies resource constraints is obtained.

[0095] For example, assume a data center has three server nodes, and a 5-minute time window is selected. The sampled data for server 1 within a certain time window are: energy consumption [80, 85, 82, 88, 86] W, CPU utilization [60%, 65%, 63%, 68%, 66%], temperature [45, 46, 45, 47, 46] °C, and power consumption [90, 95, 92, 98, 96] W. The temporal correlation coefficients between these features are calculated to form a 9×9 matrix (3 nodes × 3 feature indicators). This correlation coefficient matrix is ​​input into a two-layer prediction model. The first layer outputs a hidden state vector [0.75, 0.82, 0.79] through a gating mechanism. The second layer updates the network parameters based on this state and predicts the energy consumption curve [87, 89, 86, 90, 88] W and the load curve [67%, 69%, 66%, 70%, 68%] for server 1 in the next period.

[0096] With an energy consumption threshold of 100W and a load threshold of 80%, the calculated energy consumption margin ratio sequence [(100-87) / 100, ..., (100-88) / 100] = [0.13, 0.11, 0.14, 0.10, 0.12], with an average value of 0.12; and the load margin ratio sequence [(80-67) / 80, ..., (80-68) / 80] = [0.16, 0.14, 0.18, 0.13, 0.15], with an average value of 0.15. The weighting coefficients are set to 0.6 and 0.4 respectively, and the dynamic task carrying capacity value of server 1 is calculated as 0.6 × 0.12 + 0.4 × 0.15 = 0.132. Considering the temperature effect of 46℃, the compensation coefficient is set to 0.95, and the final task allocation weights for the three nodes are [0.125, 0.142, 0.138]. The task allocation scheme [35%, 33%, 32%] is obtained by solving the multi-objective optimization function.

[0097] In this embodiment, the dynamic changes in the server node's operating status are effectively captured through time window segmentation and feature extraction. By constructing a feature time-series correlation coefficient matrix, the correlation between multi-dimensional indicators such as energy consumption, computing load, temperature, and power consumption is comprehensively reflected, providing a reliable data foundation for subsequent predictions. The hierarchical structure not only improves the model's ability to perceive changes in server operating status but also enhances the reliability of prediction results. The introduction of a temperature compensation mechanism can effectively avoid performance loss and stability risks caused by excessively high server node temperatures. It considers both the real-time operating status of server nodes and the changing trends over a future period, effectively improving the resource utilization efficiency of the data center, reducing overall energy consumption, and ensuring the stable operation of the server cluster.

[0098] In one alternative implementation,

[0099] The predictive model is used to process the operational data of each server node to obtain the energy consumption change curve and computing load change curve of the server node for the next time period, including:

[0100] The running data corresponding to each server node is added to the prediction model, principal component analysis is performed to obtain a low-dimensional feature representation, and the low-dimensional feature representation is nonlinearly transformed by an activation function to obtain an initial hidden state. The initial hidden state is substituted into the neural ordinary differential equation, and the neural ordinary differential equation is solved to obtain a continuous-time hidden state function.

[0101] The mean and variance of the server node load data are calculated to generate a probability distribution function of the server load. A random perturbation sequence is sampled from the probability distribution function. The random perturbation sequence is superimposed with the continuous-time hidden state function to obtain a noisy state sequence.

[0102] Calculate the prediction error of the noisy state sequence, determine the step size of the numerical integration based on the prediction error, perform encrypted sampling of the noisy state sequence at the time point when the prediction error exceeds a preset threshold, and perform numerical integration on the sampling points using the fourth-order Runge-Kutta method to obtain the state evolution sequence.

[0103] The L2 norm value of the state evolution sequence and the energy loss value of the server node running data are calculated respectively. The L2 norm value and the energy loss value are added to obtain the total energy value. The total energy value is input into the second-layer time series prediction unit to calculate the gradient of the total energy value with respect to the network parameters. The network parameters are updated using the backpropagation algorithm. The causal relationship coefficient between the energy consumption change curve of the server node and the computing load change curve is calculated to generate a causal dependency graph between nodes.

[0104] Intervention operations are performed on the nodes in the causal dependency graph. The energy change value caused by each intervention operation is calculated. The energy change value is compared with the thermodynamic law of conservation of energy. The predicted data that violates the law of conservation of energy is corrected. The energy consumption change curve and computing load change curve of the server node in the next time period are output.

[0105] The operational data for each server node is added to the predictive model. This operational data includes CPU utilization, memory usage, network traffic, disk read / write speed, and power consumption. For example, taking a group of servers in a data center, operational data from 10 servers is collected every 5 minutes over the past 24 hours to form a multidimensional time series dataset.

[0106] Principal component analysis is performed on the collected operational data to reduce the dimensionality of the original high-dimensional data (such as 20 monitoring indicators of a server) to an 8-dimensional feature representation. For example, for the original data of server A, the low-dimensional feature representation obtained after principal component analysis is [0.56, -0.23, 0.78, 0.12, -0.45, 0.33, -0.67, 0.21].

[0107] The initial hidden state is obtained by performing a nonlinear transformation on the low-dimensional feature representation using the ReLU activation function. For example, after applying the ReLU activation function to the above low-dimensional features, the initial hidden state is [0.56, 0, 0.78, 0.12, 0, 0.33, 0, 0.21].

[0108] The initial hidden state is substituted into the neural constant differential equation, which is composed of a multi-layer fully connected network. The input is the hidden state at the current time point, and the output is the rate of change of the hidden state. By solving this equation, the continuous-time hidden state function is obtained. Specifically, the time interval is set to [0, 1], and it is evenly divided into 100 time points. The value of the hidden state is calculated at each time point to form the continuous-time hidden state function.

[0109] Calculate the mean and variance of server node load data to generate a probability distribution function for server load. For example, for the CPU load data of server A, the calculated mean is 65% and the variance is 8%, based on which a normal distribution function is constructed. A random perturbation sequence is sampled from this probability distribution function, such as [1.2%, -0.8%, 2.3%, -1.5%, 0.7%]. This random perturbation sequence is superimposed on a continuous-time hidden state function to obtain a noisy state sequence.

[0110] Calculate the prediction error of the noisy state sequence by averaging the sum of squares of the differences between the predicted and observed values. Determine the step size for numerical integration based on the prediction error; the larger the prediction error, the smaller the step size. For example, when the prediction error is 0.05, set the step size to 0.01; when the prediction error is 0.02, set the step size to 0.02.

[0111] At time points where the prediction error exceeds a preset threshold (e.g., 0.08), the noisy state sequence is encrypted and sampled. A fourth-order Runge-Kutta method is used to perform numerical integration on the sampled points to obtain the state evolution sequence. The fourth-order Runge-Kutta method improves the accuracy of numerical integration by estimating the state value at the next time point through a weighted average of four intermediate points.

[0112] Calculate the L2 norm of the state evolution sequence and the energy loss of the server node's running data, respectively. The L2 norm is calculated by taking the square root of the sum of the squares of each element in the state evolution sequence. For example, for the state sequence [0.5, 0.3, -0.2, 0.1, -0.4], its L2 norm is 0.726. The energy loss is obtained by calculating the mean square error between the predicted and actual values, such as 0.023.

[0113] The L2 norm value is added to the energy loss value to obtain the total energy value, such as 0.749. This total energy value is then input into the second-layer temporal prediction unit, which is composed of a Long Short-Term Memory (LSTM) network containing 64 hidden units. The gradient of the total energy value with respect to the network parameters is calculated, and the network parameters are updated using the backpropagation algorithm. The learning rate is set to 0.001, with 32 samples processed per batch and 100 training epochs.

[0114] The causal relationship coefficient between the energy consumption change curve and the computing load change curve of the server nodes is calculated. The Granger causality test is used to calculate the F-statistic and p-value at different time lags. For example, for server A, the causal relationship coefficient between energy consumption and CPU load is 0.78, and the p-value is 0.002, indicating that CPU load has a significant causal impact on energy consumption. Based on the causal relationship coefficients between nodes, a causal dependency graph between nodes is generated.

[0115] Perform intervention operations on nodes in the causal dependency graph, such as increasing the CPU load of server B by 10%, and observe the impact on its own and other servers' energy consumption. Calculate the energy change value caused by each intervention operation, such as an increase of 5.2% in server B's energy consumption and 1.3% in server C's energy consumption. Compare the energy change values ​​with the thermodynamic law of conservation of energy and correct any prediction data that violates the law. For example, if the prediction results show that the total energy increase exceeds the input energy, scale the energy change of each node proportionally to ensure that the total energy is conserved.

[0116] The output shows the energy consumption and computing load change curves for the server nodes in the next time period. For example, it is predicted that in the next 4 hours, the energy consumption of server A will first increase and then decrease, with the peak occurring in the 2nd hour, with an increase of 8%; the CPU load will fluctuate steadily within the range of 75% ± 3%. These predictions can be used for energy management and load balancing in data centers to improve energy efficiency.

[0117] In this embodiment, principal component analysis is used to reduce the dimensionality of high-dimensional operational data, effectively extracting key feature information of server nodes and reducing the complexity of subsequent calculations. The nonlinear transformation of the ReLU activation function enhances the model's ability to express nonlinear features and improves the effectiveness of feature extraction. A neural ordinary differential equation is introduced to model the server state, which, compared to traditional discrete time series models, can more accurately describe the continuous change process of the server state. Adding random perturbations to the state sequence improves the model's adaptability to unknown disturbances. An adaptive step-size numerical integration strategy optimizes computational efficiency while ensuring computational accuracy.

[0118] In existing technologies, the prediction of server node operating status usually adopts a single prediction model, which directly processes the raw high-dimensional data. This not only has high computational complexity, but is also easily affected by data noise, resulting in unstable prediction accuracy. At the same time, traditional methods often ignore the mutual influence between server nodes, making it difficult to accurately depict the dynamic changes of the data center as a whole.

[0119] This embodiment achieves accurate prediction of server node energy consumption and computing load through the synergistic effect of two-layer time-series prediction units. It improves the efficiency of feature extraction through dimensionality reduction and nonlinear transformation, enhances the reliability of prediction results by introducing physical model constraints, and strengthens the adaptability of the model by adopting a two-layer prediction structure. This enables more accurate prediction of the operating status of server nodes and provides a more reliable basis for decision-making in data center energy management and load balancing.

[0120] Figure 2 This chart compares the energy consumption prediction errors of the data center energy consumption and computing power balancing optimization method based on swarm intelligence, as described in this invention. It shows the energy consumption prediction errors of different prediction models across various prediction time ranges. A visualization combining bar charts and lines clearly illustrates the performance differences of each method at different prediction durations. This technical solution employs a method combining neural ordinary differential equations and causal inference, demonstrating significant advantages across all prediction time ranges.

[0121] Specifically, in short-term (1-hour) predictions, the prediction error of this technique is only 2.1%, while the traditional LSTM method is 4.2% and the GRU method is 5.2%. As the prediction time increases, the prediction errors of all methods show an upward trend, but the error growth of this technique is the slowest. In medium-term predictions (6 hours), the prediction error of this technique is 6.0%, far lower than LSTM's 12.3% and GRU's 13.7%. In long-term predictions (24 hours), the prediction error of this technique is 10.8%, while LSTM reaches 21.9% and GRU reaches 24.1%, indicating that this technique has significant resistance to error accumulation in long-term predictions. This is attributed to the use of the neural ordinary differential equation, which accurately captures the dynamic characteristics of the system, and the energy conservation constraint correction mechanism, which effectively prevents prediction deviation. The height of the bar chart intuitively shows the error difference between different methods, while the lines clearly show the error growth trend of each method as the prediction time increases. The line for this technique is significantly lower than that of other methods, proving its superior prediction stability.

[0122] In one alternative implementation,

[0123] The parallel ant colony computation space is divided into multiple sub-regions, and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate. The convergence direction of the ant colony search path is controlled, and the optimal path is marked, including:

[0124] Each parallel ant colony computing space is divided into multiple sub-regions. A temperature decay coefficient is calculated based on the server nodes within each sub-region. The temperature decay coefficient is dynamically adjusted according to time changes and converted into a pheromone evaporation rate. Pheromones are updated in each parallel ant colony computing space in conjunction with the task allocation path within each sub-region. The pheromone retention amount is calculated based on the pheromone evaporation rate. The pheromone increments added by all individual ants on the search path are accumulated. The updated pheromone concentration is obtained by adding the pheromone retention amount to the pheromone increments.

[0125] In each parallel ant colony computing space, the transition probability of an individual ant choosing different task allocation schemes is calculated. Based on the transition probability, the convergence direction of the ant colony search path is controlled. The fitness value of the task allocation scheme in each parallel ant colony computing space is calculated. The task allocation schemes in each parallel ant colony computing space are sorted in descending order of fitness value. In each iteration cycle, the task allocation scheme with the highest fitness value is marked as the optimal path.

[0126] The computational space of each parallel ant colony is divided into multiple sub-regions. Assuming there are N server nodes, they can be divided into M sub-regions based on their physical location or logical relationships. For example, in a data center environment containing 100 server nodes, it can be divided into 10 sub-regions, each containing 10 server nodes. The partitioning can be achieved using the K-means clustering algorithm, grouping nodes based on network latency or geographical location.

[0127] The temperature decay coefficient is calculated based on the server nodes within a sub-region. The temperature decay coefficient is related to the computing power, load status, and energy consumption of the server nodes within the sub-region. For sub-region i, its temperature decay coefficient Ti can be calculated by comprehensively considering the average CPU utilization, memory utilization, and energy consumption level of all server nodes within the sub-region. For example, when the average CPU utilization in the sub-region is 60%, the average memory utilization is 50%, and the average energy consumption is 300W, the initial temperature decay coefficient can be set to 0.85.

[0128] The temperature decay coefficient is dynamically adjusted based on time changes. As the number of iterations increases, the temperature decay coefficient needs to be gradually decreased to accelerate the convergence speed of the algorithm. Assuming the initial temperature decay coefficient is T0, the current iteration number is iter, and the maximum number of iterations is MAX_ITER, the temperature decay coefficient can be updated in a linear decreasing manner. For example, in the 50th iteration, if the initial temperature decay coefficient is 0.85 and the maximum number of iterations is 200, then the current temperature decay coefficient can be updated to 0.7375.

[0129] The temperature decay coefficient is converted into the pheromone evaporation rate. The pheromone evaporation rate ρ is positively correlated with the temperature decay coefficient T. When the temperature decay coefficient is high, the pheromone evaporation rate is low, which is beneficial for global search; when the temperature decay coefficient is low, the pheromone evaporation rate is high, which is beneficial for local search. For example, when the temperature decay coefficient is 0.7375, the pheromone evaporation rate can be set to 0.2625.

[0130] Pheromones are updated in each parallel ant colony computation space based on the task allocation paths within each sub-region. Assume there are K tasks to be assigned to N server nodes, and each ant constructs a complete task allocation path from the starting point to the ending point. The pheromone concentration on the path where the k-th task is assigned to the n-th server node needs to be updated.

[0131] The pheromone retention amount is calculated based on the pheromone evaporation rate. For a pheromone concentration τ(k,n) on the path (k,n), the retention amount is (1-ρ)×τ(k,n). For example, when the original pheromone concentration on the path is 2.5 and the pheromone evaporation rate is 0.2625, the pheromone retention amount is (1-0.2625)×2.5=1.84375.

[0132] The total pheromone increment is the sum of the increments made by all individual ants along the search path. Assuming there are ANT_NUM ants, and the pheromone increment left by the a-th ant on path (k, n) is Δτa(k, n), then the total increment is the sum of the increments of all ants. The pheromone increment is proportional to the quality of the solution constructed by the ants. For example, if the increment of the first ant on the path is 0.3, and the increment of the second ant is 0.25, then the total increment is 0.55.

[0133] The updated pheromone concentration is obtained by adding the pheromone retention amount to the pheromone increment. The updated pheromone concentration τnew(k, n) is equal to the pheromone retention amount plus the pheromone increment. Continuing the previous example, the updated pheromone concentration is 1.84375 + 0.55 = 2.39375.

[0134] In each parallel ant colony's computational space, the transition probability of an individual ant choosing different task assignment schemes is calculated. The transition probability is related to the pheromone concentration along the path and heuristic information. Heuristic information can be the matching degree between the server node's processing power and the task requirements. For example, for CPU-intensive tasks, server nodes with stronger CPU performance can be preferentially selected. Assuming the pheromone concentration of path (k, n1) is 2.39375 and the heuristic value is 0.8, and the pheromone concentration of path (k, n2) is 1.5 and the heuristic value is 0.6, then the probability of choosing n1 is approximately 0.68, and the probability of choosing n2 is approximately 0.32.

[0135] The convergence direction of the ant colony's search path is controlled based on transition probabilities. During the search process, ants choose their next movement direction according to the transition probabilities. By adjusting the pheromone evaporation rate, the convergence speed and direction of the algorithm can be controlled. A lower evaporation rate helps maintain diversity, while a higher evaporation rate helps accelerate convergence.

[0136] Calculate the fitness value of each task allocation scheme in the computational space of the parallel ant colony. The fitness value can be comprehensively evaluated based on indicators such as task completion time, system load balancing, and energy consumption. For example, if the average task completion time of a certain task allocation scheme is 150ms, the load imbalance is 0.15, and the total energy consumption is 2500W, then its fitness value can be calculated as a weighted sum of these indicators.

[0137] The task allocation schemes in each parallel ant colony computing space are sorted in descending order of fitness value. Assuming there are 5 parallel ant colony computing spaces, each with 20 task allocation schemes, then the 20 schemes in each space need to be sorted separately.

[0138] In each iteration, the task allocation scheme with the highest fitness value is marked as the optimal path. For example, in a certain iteration, the 7th task allocation scheme in the 3rd parallel computing space has the highest fitness value of 0.92, so it is marked as the optimal path for the current iteration. This optimal path will serve as a reference for the next iteration and may become the global optimal solution.

[0139] In this embodiment, by dividing the parallel ant colony computing space into multiple sub-regions, the search space dimension of the task allocation problem is effectively reduced, and the computational efficiency of the algorithm is improved. The sub-region division strategy fully considers the physical location and logical relationship of the server nodes, making the task allocation more in line with the characteristics of the actual network environment. The temperature decay coefficient calculated based on the sub-region can accurately reflect the operating status of the local server cluster. By dynamically adjusting the temperature decay coefficient, the adaptive adjustment of the algorithm's search intensity is achieved. By performing ant colony search simultaneously in multiple parallel computing spaces, the search efficiency of the algorithm is significantly improved.

[0140] In one alternative implementation,

[0141] Converting the temperature decay coefficient into the pheromone evaporation rate, and performing pheromone updates in each parallel ant colony computation space in conjunction with the task allocation path within each sub-region, includes:

[0142] Collect the temperature decay coefficient of the server node in each sub-region, calculate the task processing capability difference coefficient of the server node in the sub-region based on the temperature decay coefficient, and set an independent pheromone evaporation rate for each sub-region according to the task processing capability difference coefficient.

[0143] Record the task allocation path of an individual ant in each sub-region, calculate the load balancing index and energy efficiency index of the task allocation path, weight the load balancing index and the energy efficiency index to obtain the path comprehensive score, map the path comprehensive score to a pheromone intensity adjustment coefficient according to a pre-set mapping relationship table, and calculate the path pheromone constraint threshold based on the pheromone intensity adjustment coefficient, wherein the path pheromone constraint threshold is dynamically adjusted with the path comprehensive score.

[0144] Detect the pheromone intensity of each task allocation path. If the pheromone intensity of a task allocation path exceeds the preset path pheromone constraint threshold, obtain the set of adjacent paths of that task allocation path.

[0145] Calculate the comprehensive path score for each path in the adjacent path set, calculate the pheromone redistribution weight based on the comprehensive path score, and distribute the pheromone amount exceeding the path pheromone constraint threshold to the paths in the adjacent path set according to the pheromone redistribution weight, thereby obtaining the pheromone increment added by the ant individual on the search path.

[0146] The system collects the temperature decay coefficient corresponding to the server nodes in each sub-region. The temperature decay coefficient reflects the thermal characteristics of the server nodes and can be obtained in real time through temperature sensors. For example, for sub-region A containing 5 server nodes, its temperature decay coefficients are [0.92, 0.88, 0.95, 0.90, 0.87], representing the differences in heat dissipation capabilities of different nodes.

[0147] The task processing capability difference coefficient of server nodes within a sub-region is calculated based on the temperature decay coefficient. In practice, the temperature decay coefficient can be compared with a preset benchmark value (such as 0.90) to calculate the relative deviation value. For example, for the server nodes in sub-region A above, the task processing capability difference coefficient can be calculated as [0.022, -0.022, 0.056, 0.000, -0.033].

[0148] An independent pheromone evaporation rate is set for each sub-region based on the task processing capacity difference coefficient. The pheromone evaporation rate is positively correlated with the task processing capacity difference coefficient; the greater the difference, the higher the evaporation rate is set, thus promoting dynamic adjustment of task allocation. For example, the pheromone evaporation rate for sub-region A can be set to 0.15, while that for sub-region B, where the temperature decay coefficient is more uniform, it can be set to 0.10.

[0149] Record the task allocation path of an individual ant in each sub-region. The task allocation path is represented as a sequence of accesses to server nodes, such as path P1=[S1→S3→S5→S2], which means that the task is assigned to server nodes S1, S3, S5 and S2 in sequence.

[0150] The system calculates the load balancing index and energy efficiency index for task allocation paths. The load balancing index reflects the uniformity of load distribution among nodes on the path, while the energy efficiency index reflects the energy consumption efficiency of task execution on the path. For example, the load balancing index for path P1 is 0.78, and the energy efficiency index is 0.65.

[0151] The overall score of the path is obtained by weighting the load balancing index and the energy efficiency index. The weighting coefficients can be dynamically adjusted according to the system optimization objectives. For example, when the load balancing weight is 0.6 and the energy efficiency weight is 0.4, the overall score of path P1 is 0.78×0.6+0.65×0.4=0.728.

[0152] The path's overall score is mapped to a pheromone intensity adjustment coefficient based on a pre-defined mapping table. This mapping table can be designed as a piecewise linear mapping; for example, a path with an overall score in the range [0.7, 0.8] has a pheromone intensity adjustment coefficient of 1.2, while a path with an overall score in the range [0.8, 0.9] has a pheromone intensity adjustment coefficient of 1.5. For path P1, its overall score of 0.728 corresponds to a pheromone intensity adjustment coefficient of 1.2.

[0153] The path pheromone constraint threshold is calculated based on the pheromone intensity adjustment coefficient. The path pheromone constraint threshold is dynamically adjusted according to the overall path score; the higher the overall score, the higher the constraint threshold. For example, a baseline threshold of 10 can be set, and the pheromone constraint threshold for path P1 can be 10 × 1.2 = 12.

[0154] The system detects the pheromone intensity of each task allocation path. If the pheromone intensity of a task allocation path exceeds a pre-set path pheromone constraint threshold, the system obtains the set of adjacent paths for that task allocation path. Adjacent paths are defined as paths that differ from the current path by no more than a certain degree, such as paths that differ by only one node. For example, the set of adjacent paths for path P1=[S1→S3→S5→S2] may include P2=[S1→S3→S5→S4], P3=[S1→S3→S4→S2], etc.

[0155] Assuming the pheromone intensity of path P1 is 15, exceeding its constraint threshold of 12, the system will calculate the comprehensive score of each path in its neighboring path set. For example, the comprehensive score of P2 is 0.685, and the comprehensive score of P3 is 0.710.

[0156] The pheromone redistribution weights are calculated based on the overall path score. The higher the overall score, the greater the redistribution weight. For example, the redistribution weights for P2 and P3 are 0.49 and 0.51, respectively.

[0157] The amount of pheromone exceeding the path pheromone constraint threshold is redistributed to the paths in the adjacent path set according to the pheromone redistribution weight. For path P1, the amount of pheromone exceeding the threshold is 15-12=3, which is redistributed to P2 and P3 according to the redistribution weight, that is, P2 increases by 3×0.49=1.47 pheromone, and P3 increases by 3×0.51=1.53 pheromone.

[0158] Through a pheromone redistribution mechanism, the system obtains the incremental pheromone content of individual ants along their search paths. This mechanism ensures the rationality of pheromone distribution, prevents excessive concentration of pheromones on specific paths, and improves the ant colony algorithm's exploration capability and convergence efficiency in task allocation.

[0159] In this embodiment, a task processing capability evaluation mechanism based on temperature decay coefficient is introduced. By monitoring the temperature characteristics of server nodes in real time, the task processing capability difference coefficient between nodes is calculated, providing a reliable basis for subsequent pheromone adjustment. The evaluation method based on physical characteristics makes task allocation more in line with the actual operating environment, and a differentiated pheromone evaporation rate adjustment mechanism is designed. According to the difference in task processing capability of server nodes in a sub-region, the pheromone evaporation rate is dynamically adjusted, realizing adaptive optimization of task allocation.

[0160] In existing technologies, server task allocation typically employs a fixed pheromone update strategy, which fails to fully consider the physical characteristics and dynamic changes of server nodes. This can easily lead to uneven pheromone distribution, affecting the convergence performance of the algorithm. Furthermore, traditional solutions often treat load balancing and energy efficiency as independent optimization objectives, making it difficult to achieve a dynamic balance between multiple objectives.

[0161] This embodiment improves the targeting of task allocation through temperature characteristic analysis, enhances the algorithm's adaptability through differentiated pheromone adjustment strategies, improves the overall effect of task allocation through multi-objective joint optimization and pheromone redistribution mechanisms, and significantly improves the accuracy and efficiency of task allocation through multi-level optimization mechanisms, providing strong support for the intelligent operation and maintenance of data centers.

[0162] Figure 3 The diagram shows the pheromone distribution and path convergence of the data center energy consumption and computing power balancing optimization method based on swarm intelligence in this embodiment of the invention, illustrating the pheromone distribution and path convergence of four algorithms in server task allocation.

[0163] The diagram shows the temperature decay coefficients of the 11 server nodes (S1-S11), ranging from 0.85 to 0.95, indicating significant differences in temperature characteristics. The path chosen by this technical solution (thick solid line) is S1→S3→S8→S9→S11, with pheromone intensities of 8.5, 9.2, 7.8, and 8.1 along the way, all below the set constraint threshold of 10.0. This demonstrates the effective control of pheromone intensity by this solution. The standard ACO algorithm selects the path S1→S2→S7→S11 (medium dashed line) with pheromone intensities as high as 12.6, 11.8, and 14.2, far exceeding the reasonable threshold, leading to premature convergence to a local optimum.

[0164] Although the Max-Min ant colony algorithm improves this problem by setting upper and lower limit constraints, with pheromone intensities of 10.2, 9.8, 9.5 and 9.0 on the selected path S1→S6→S7→S8→S11 (thin dashed line), it does not take into account the differences in server temperature characteristics.

[0165] The temperature-sensing priority algorithm considers temperature factors, and the pheromone intensities on the path S1→S3→S4→S9→S11 (dotted line) are 9.7, 8.9, 8.3, and 8.5, respectively. However, it lacks the dynamic pheromone redistribution mechanism of this technical solution. As can be clearly seen from the figure, this technical solution, by integrating the temperature attenuation coefficient and the dynamic pheromone adjustment mechanism, ensures a moderate pheromone distribution while selecting a more reasonable server access path, avoiding server nodes with low temperature attenuation coefficients (poor heat dissipation capabilities) (such as S10 and S5), thus achieving dual optimization of load balancing and energy efficiency.

[0166] Table 1 compares the multi-dimensional performance indicators of the data center energy consumption and computing power balancing optimization methods based on swarm intelligence, showing the comparison results of five algorithms on six key performance indicators. This technical solution performs best across all indicators, achieving a comprehensive score of 9.2, significantly higher than the standard ACO algorithm's 7.1.

[0167] Table 1 is a comparison of multi-dimensional performance indicators of data center energy consumption and computing power balancing optimization methods based on swarm intelligence:

[0168]

[0169] Specifically, this technical solution achieves a load balancing score of 0.935, which is 0.09 percentage points higher than the standard ACO algorithm; an energy efficiency of 0.882, which is 0.087 higher than the standard ACO algorithm; a convergence speed of only 42 iterations, almost half that of the standard ACO algorithm (78 iterations); a high-temperature adaptability score of 0.875, which is 0.16 higher than the standard ACO algorithm; a task throughput of 3450 tasks / minute, which is 21% higher than the standard ACO algorithm; and a scheduling fairness score of 0.915, which is 0.13 higher than the standard ACO algorithm. Compared with the Max-Min ant colony algorithm, this technical solution is 0.05 higher in load balancing, 0.057 higher in energy efficiency, and 35% faster in convergence speed. Compared with the temperature sensing priority algorithm, although the gap is smaller, this technical solution still has significant advantages in load balancing, energy efficiency, and scheduling fairness, exceeding them by 0.033, 0.037, and 0.05 respectively. Compared with the dynamic temperature sensing algorithm, this technical solution maintains a comprehensive but small leading advantage.

[0170] The above data fully demonstrates the effectiveness of the temperature attenuation coefficient calculation and pheromone redistribution mechanism in this technical solution, enabling it to maintain stable high performance under various operating conditions. Particularly in terms of high-temperature adaptability, the gap between this technical solution and the temperature-sensing priority algorithm and the dynamic temperature-sensing algorithm is relatively small (0.045 and 0.030 respectively), indicating that all three algorithms effectively consider temperature factors. However, this technical solution, through a superior pheromone adjustment strategy, still maintains a significant advantage in overall performance.

[0171] In one alternative implementation,

[0172] Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the operating status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching results, including:

[0173] Obtain historical task processing records of server nodes in the globally optimal task allocation scheme, extract energy consumption data and computing load data from the historical task processing records, and calculate the target energy consumption threshold and target computing load threshold for each server node.

[0174] Monitor the real-time energy consumption and real-time computing load of server nodes, calculate the difference between the real-time energy consumption and the target energy consumption threshold, and the difference between the real-time computing load and the target computing load threshold, and generate a judgment result based on the difference;

[0175] Based on the judgment result, the source server node and target server node for task migration are determined, the task migration operation is executed, and the operating frequency and power supply voltage of each server node are adjusted until the real-time energy consumption value and real-time computing load value of the source server node and the target server node converge to a preset range, thereby obtaining the optimal matching result.

[0176] The system retrieves historical task processing records for server nodes within the globally optimal task allocation scheme. It extracts operational data from the database for each server node under optimal task allocation conditions over the past 30 days, including hourly energy consumption and computing load data. For example, in a cluster with 10 server nodes, each node would have 720 historical records (30 days x 24 hours).

[0177] Extract energy consumption and computing load data from historical task processing records. The system organizes the energy consumption data (in kilowatt-hours) and computing load data (in percentages) of each server node into time series. For example, server node 1 has an average energy consumption of 0.85 kWh / hour and an average computing load of 65% over the past 30 days.

[0178] The system calculates the target energy consumption threshold and target computing load threshold for each server node. It statistically analyzes the historical energy consumption data of each server node, taking the average energy consumption under optimal operating conditions as the target energy consumption threshold, while considering energy consumption fluctuations and setting an upper and lower fluctuation range of ±5%. Similarly, it analyzes historical computing load data, taking the average load under optimal operating conditions as the target computing load threshold, setting an upper and lower fluctuation range of ±10%. For example, the target energy consumption threshold for server node 1 is 0.85 kWh / hour (fluctuation range 0.81-0.89 kWh / hour), and the target computing load threshold is 65% (fluctuation range 58.5%-71.5%).

[0179] During the real-time monitoring phase, the system collects the real-time energy consumption and real-time computing load values ​​of the server nodes every minute. Data collection methods include obtaining power data, CPU utilization, memory usage, network throughput, and other metrics through the server management interface. For example, at a certain moment, server node 1's real-time energy consumption is 0.95 kWh / hour, and its real-time computing load is 75%.

[0180] The system calculates the difference between the real-time energy consumption value and the target energy consumption threshold, as well as the difference between the real-time computing load value and the target computing load threshold. For server node 1, the energy consumption difference is 0.1 kWh / hour (0.95-0.85), and the computing load difference is 10% (75%-65%). The system compares these differences with preset fluctuation ranges to generate a judgment result.

[0181] The judgment results include: Normal state: The real-time value is within the fluctuation range of the target threshold and no adjustment is required; High energy consumption state: The real-time energy consumption value exceeds the upper limit of the target energy consumption threshold; Low energy consumption state: The real-time energy consumption value is lower than the lower limit of the target energy consumption threshold; High load state: The real-time computing power load value exceeds the upper limit of the target computing power load threshold; Low load state: The real-time computing power load value is lower than the lower limit of the target computing power load threshold; Composite abnormal state: Both energy consumption and load are abnormal at the same time.

[0182] Based on the judgment results, the system determines the source server node and target server node for task migration. For example, when server node 1 is in a state of "high energy consumption and high load" and server node 2 is in a state of "low energy consumption and low load", the system sets node 1 as the source node and node 2 as the target node, preparing to perform task migration.

[0183] During task migration, the system selects suitable tasks from the source node's task queue. Selection criteria include task priority, resource requirements, and migration cost. For example, for 10 tasks running on server node 1, the system selects two tasks with higher resource requirements but lower priority for migration.

[0184] Assess the resource availability of the target node to ensure it has sufficient resources to receive the migration tasks. For example, if server node 2 currently has a CPU utilization of 40% and a memory utilization of 35%, it is sufficient to receive the two tasks from node 1 (which are expected to increase CPU utilization by 15% and memory utilization by 10%).

[0185] Performing task migration transfers selected tasks from the source node to the target node. The migration process includes steps such as saving task state, network transmission, and task recovery on the target node. For example, the system uses real-time migration technology to complete the migration of two tasks within 100 milliseconds, during which task execution is uninterrupted.

[0186] The operating frequency and power supply voltage of each server node are adjusted to optimize energy consumption. For the source node (server node 1) after the load is reduced, the system reduces the CPU frequency from 3.2GHz to 2.8GHz and the power supply voltage from 1.2V to 1.1V; for the target node (server node 2) after the load is increased, the system increases the CPU frequency from 2.4GHz to 2.7GHz and the power supply voltage from 1.0V to 1.05V.

[0187] Continuously monitor the real-time energy consumption and computing load of the source and target nodes until they converge to a preset range. For example, after migration and adjustment, the energy consumption of server node 1 drops to 0.87 kWh / hour (within the range of 0.81-0.89), and the computing load drops to 68% (within the range of 58.5%-71.5%); the energy consumption of server node 2 rises to 0.78 kWh / hour (close to its target threshold of 0.8), and the computing load rises to 62% (close to its target threshold of 60%).

[0188] The optimal matching results are recorded, including the task allocation, energy consumption level, and computing load level of each server node. This data serves as new historical data for future threshold calculations and optimizations. For example, after this optimization, the average energy efficiency of the entire cluster improved by 8%, and the task processing capacity improved by 5%, achieving the expected optimization goals.

[0189] In this embodiment, through real-time monitoring and dynamic judgment mechanisms, abnormal operating conditions of server nodes can be accurately identified, and the source and target nodes that need to be migrated can be determined in a timely manner, ensuring the relevance and necessity of task migration. By adopting multi-dimensional task evaluation criteria and comprehensively considering factors such as task priority, resource requirements, and migration costs, the most suitable tasks for migration are selected from the source nodes, effectively reducing the impact of migration on the overall system performance. Through accurate evaluation of the resource availability of the target node, the system can accurately predict the resource usage after migration before migration, avoiding resource overload problems that may be caused by blind migration, and significantly improving the success rate of task migration.

[0190] A data center energy consumption and computing power balancing optimization system based on swarm intelligence includes:

[0191] The first unit is used to collect real-time operating data of server nodes in the computer room, standardize the real-time operating data, and construct the total energy consumption function, computing power allocation balancing function, and service quality assurance function of the computer room, and combine them to obtain a multi-objective optimization function.

[0192] The second unit is used to process standardized real-time running data according to the time-series prediction unit of the recurrent neural network, predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme according to the multi-objective optimization function and the task allocation weight.

[0193] The third unit is used to construct multiple parallel ant colony computing spaces based on the initial task allocation scheme, divide the parallel ant colony computing space into multiple sub-regions and set a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path and mark the optimal path, and filter the globally optimal task allocation scheme.

[0194] The fourth unit is used to calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result.

[0195] A third aspect of the present invention,

[0196] An electronic device is provided, comprising:

[0197] processor;

[0198] Memory used to store processor-executable instructions;

[0199] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0200] Fourth aspect of the embodiments of the present invention,

[0201] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0202] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0203] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing data center energy consumption and computing power balance based on swarm intelligence, characterized in that, include: Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized and the total energy consumption function, computing power allocation balancing function and service quality assurance function of the data center are constructed, and combined to obtain a multi-objective optimization function; The time-series prediction unit of the recurrent neural network processes the standardized real-time running data to predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme based on the multi-objective optimization function and the task allocation weight. Based on the initial task allocation scheme, multiple parallel ant colony computing spaces are constructed. The parallel ant colony computing spaces are divided into multiple sub-regions and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path, mark the optimal path, and filter the globally optimal task allocation scheme. Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result. The parallel ant colony computation space is divided into multiple sub-regions, and a temperature decay coefficient is set to dynamically adjust the pheromone evaporation rate. The convergence direction of the ant colony search path is controlled, and the optimal path is marked, including: Each parallel ant colony computing space is divided into multiple sub-regions. A temperature decay coefficient is calculated based on the server nodes within each sub-region. The temperature decay coefficient is dynamically adjusted according to time changes and converted into a pheromone evaporation rate. Pheromones are updated in each parallel ant colony computing space in conjunction with the task allocation path within each sub-region. The pheromone retention amount is calculated based on the pheromone evaporation rate. The pheromone increments added by all individual ants on the search path are accumulated. The updated pheromone concentration is obtained by adding the pheromone retention amount to the pheromone increments. In each parallel ant colony computing space, the transition probability of an individual ant choosing different task allocation schemes is calculated. Based on the transition probability, the convergence direction of the ant colony search path is controlled. The fitness value of the task allocation scheme in each parallel ant colony computing space is calculated. The task allocation schemes in each parallel ant colony computing space are sorted in descending order of fitness value. In each iteration cycle, the task allocation scheme with the highest fitness value is marked as the optimal path.

2. The method according to claim 1, characterized in that, Real-time operating data of server nodes in the data center is collected, the real-time operating data is standardized, and a total energy consumption function, a computing power allocation balancing function, and a service quality assurance function are constructed. These are combined to obtain a multi-objective optimization function, including: Collect real-time operating data of server nodes in the computer room and perform standardized processing; Based on the historical operating status of each server node in the data center, set the weights of energy consumption index, computing power index, and service quality index. Based on the weights of the energy consumption index, calculate the total energy consumption of the server nodes to obtain the total energy consumption function of the data center. Based on the weights of the computing power index, calculate the computing power distribution deviation of the task among the server nodes to obtain the computing power distribution balance function. Based on the weights of the service quality index, calculate the task processing latency of each server node to obtain the service quality assurance function. The total energy consumption function of the computer room, the computing power allocation balancing function, and the service quality assurance function are normalized, and the combination coefficients of each function are dynamically adjusted. Based on the combination coefficients, the three functions are combined into a multi-objective optimization function.

3. The method according to claim 1, characterized in that, The time-series prediction unit of the recurrent neural network processes standardized real-time operational data to predict the energy consumption and computing load change curves of server nodes, calculates dynamic task carrying capacity values, determines task allocation weights, and generates an initial task allocation scheme based on the multi-objective optimization function and the task allocation weights, including: Standardized real-time operating data is divided into time windows, and the data within each time window is organized into a feature matrix. The feature matrix includes server node energy consumption data, computing load data, temperature data, and power consumption data. Based on the feature matrix, a feature time-series correlation coefficient matrix is ​​calculated. A two-layer prediction model is constructed based on the temporal prediction unit in the recurrent neural network. The prediction model includes a first-layer temporal prediction unit and a second-layer temporal prediction unit. The first-layer temporal prediction unit processes the feature temporal correlation coefficient matrix to obtain the hidden state. The second-layer temporal prediction unit updates the network parameters based on the hidden state. The prediction model processes the running data of each server node to obtain the energy consumption change curve and computing load change curve of the server node in the next time period. Based on the energy consumption change curve and the maximum energy consumption threshold, an energy consumption margin ratio sequence is calculated. Based on the computing power load change curve and the maximum computing power load threshold, a computing power load margin ratio sequence is calculated. Based on the average value of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combined with the first weighting coefficient and the second weighting coefficient, the dynamic task carrying capacity value of the server node is calculated. The task allocation weights are calculated based on the dynamic task carrying capacity value and temperature parameters. The initial task allocation scheme is obtained by multiplying the task allocation weights by the multi-objective optimization function and finding the minimum value.

4. The method according to claim 3, characterized in that, The predictive model is used to process the operational data of each server node to obtain the energy consumption change curve and computing load change curve of the server node for the next time period, including: The running data corresponding to each server node is added to the prediction model, principal component analysis is performed to obtain a low-dimensional feature representation, and the low-dimensional feature representation is nonlinearly transformed by an activation function to obtain an initial hidden state. The initial hidden state is substituted into the neural ordinary differential equation, and the neural ordinary differential equation is solved to obtain a continuous-time hidden state function. The mean and variance of the server node load data are calculated to generate a probability distribution function of the server load. A random perturbation sequence is sampled from the probability distribution function. The random perturbation sequence is superimposed with the continuous-time hidden state function to obtain a noisy state sequence. Calculate the prediction error of the noisy state sequence, determine the step size of the numerical integration based on the prediction error, perform encrypted sampling of the noisy state sequence at the time point when the prediction error exceeds a preset threshold, and perform numerical integration on the sampling points using the fourth-order Runge-Kutta method to obtain the state evolution sequence. The L2 norm value of the state evolution sequence and the energy loss value of the server node running data are calculated respectively. The L2 norm value and the energy loss value are added to obtain the total energy value. The total energy value is input into the second-layer time series prediction unit to calculate the gradient of the total energy value with respect to the network parameters. The network parameters are updated using the backpropagation algorithm. The causal relationship coefficient between the energy consumption change curve of the server node and the computing load change curve is calculated to generate a causal dependency graph between nodes. Intervention operations are performed on the nodes in the causal dependency graph. The energy change value caused by each intervention operation is calculated. The energy change value is compared with the thermodynamic law of conservation of energy. The predicted data that violates the law of conservation of energy is corrected. The energy consumption change curve and computing load change curve of the server node in the next time period are output.

5. The method according to claim 1, characterized in that, Converting the temperature decay coefficient into the pheromone evaporation rate, and performing pheromone updates in each parallel ant colony computation space in conjunction with the task allocation path within each sub-region, includes: Collect the temperature decay coefficient of the server node in each sub-region, calculate the task processing capability difference coefficient of the server node in the sub-region based on the temperature decay coefficient, and set an independent pheromone evaporation rate for each sub-region according to the task processing capability difference coefficient. Record the task allocation path of an individual ant in each sub-region, calculate the load balancing index and energy efficiency index of the task allocation path, weight the load balancing index and the energy efficiency index to obtain the path comprehensive score, map the path comprehensive score to a pheromone intensity adjustment coefficient according to a pre-set mapping relationship table, and calculate the path pheromone constraint threshold based on the pheromone intensity adjustment coefficient, wherein the path pheromone constraint threshold is dynamically adjusted with the path comprehensive score. Detect the pheromone intensity of each task allocation path. If the pheromone intensity of a task allocation path exceeds the preset path pheromone constraint threshold, obtain the set of adjacent paths of that task allocation path. Calculate the comprehensive path score for each path in the adjacent path set, calculate the pheromone redistribution weight based on the comprehensive path score, and distribute the pheromone amount exceeding the path pheromone constraint threshold to the paths in the adjacent path set according to the pheromone redistribution weight, thereby obtaining the pheromone increment added by the ant individual on the search path.

6. The method according to claim 1, characterized in that, Calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the operating status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching results, including: Obtain historical task processing records of server nodes in the globally optimal task allocation scheme, extract energy consumption data and computing load data from the historical task processing records, and calculate the target energy consumption threshold and target computing load threshold for each server node. Monitor the real-time energy consumption and real-time computing load of server nodes, calculate the difference between the real-time energy consumption and the target energy consumption threshold, and the difference between the real-time computing load and the target computing load threshold, and generate a judgment result based on the difference; Based on the judgment result, the source server node and target server node for task migration are determined, the task migration operation is executed, and the operating frequency and power supply voltage of each server node are adjusted until the real-time energy consumption value and real-time computing load value of the source server node and the target server node converge to a preset range, thereby obtaining the optimal matching result.

7. A data center energy consumption and computing power balancing optimization system based on swarm intelligence, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect real-time operating data of server nodes in the computer room, standardize the real-time operating data, and construct the total energy consumption function, computing power allocation balancing function, and service quality assurance function of the computer room, and combine them to obtain a multi-objective optimization function. The second unit is used to process standardized real-time running data according to the time-series prediction unit of the recurrent neural network, predict the energy consumption change curve and computing load change curve of the server node, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation scheme according to the multi-objective optimization function and the task allocation weight. The third unit is used to construct multiple parallel ant colony computing spaces based on the initial task allocation scheme, divide the parallel ant colony computing space into multiple sub-regions and set a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path and mark the optimal path, and filter the globally optimal task allocation scheme. The fourth unit is used to calculate the target energy consumption threshold and target computing load threshold for each server node under the globally optimal task allocation scheme, monitor the running status of server nodes and generate judgment results, and perform task migration and energy consumption level adjustment between server nodes based on the judgment results to obtain the optimal matching result.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images