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

Abstract

The invention provides a computer room energy consumption and computing power balance optimization method and system based on swarm intelligence, and relates to the technical field of data center management, and the method comprises the steps: collecting server node real-time operation data, constructing a multi-objective optimization function, predicting an energy consumption and computing power change curve through a recurrent neural network, and generating an initial task distribution scheme; and a global optimal allocation scheme is searched by adopting a parallel ant colony algorithm, the running state of the server is monitored, and task migration and energy consumption adjustment are executed, so that collaborative optimization of energy consumption reduction of the machine room and balanced allocation of computing power is realized, and the service quality and the resource utilization efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data center management, and in particular, to a method and system for optimizing the balance between computer room energy consumption and computing power based on swarm intelligence. Background Art

[0002] With the rapid development of information technology, the scale of data center computer rooms has been continuously expanding, the number of servers has increased sharply, and the problems of computer room energy consumption and computing power resource allocation have become 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 the computer room will generate different energy consumption and computing power loads when processing different tasks. How to minimize the computer room energy consumption and evenly distribute the computing power resources on the premise of ensuring service quality has become a hot issue in current research.

[0004] Traditional methods for optimizing the balance between computer room energy consumption and computing power mainly rely on static configuration and manual intervention, and task allocation and energy consumption control are performed on server nodes through preset rules. However, the existing technologies still have problems in optimizing the balance between computer room energy consumption and computing power, such as it is difficult to ensure the balanced distribution of computing power resources while reducing the total energy consumption, unable to accurately capture the dynamic change characteristics of server node energy consumption and computing power load, limited prediction accuracy, low search efficiency, and easy to fall into local optimum, etc.;

[0005] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

[0006] Embodiments of the present invention provide a method and system for optimizing the balance between computer room energy consumption and computing power based on swarm intelligence, which can at least solve some of the problems existing in the prior art.

[0007] In the first aspect of the embodiments of the present invention, a method for optimizing the balance between computer room energy consumption and computing power based on swarm intelligence is provided, including:

[0008] Collect the real-time operation data of server nodes in the computer room, perform standardization processing on the real-time operation data, and construct a total computer room energy consumption function, a computing power allocation balance function, and a service quality guarantee function, and combine them to obtain a multi-objective optimization function;

[0009] Process the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predict the energy consumption change curve and the computing power 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 plan according to the multi-objective optimization function and the task allocation weight;

[0010] Construct multiple parallel ant colony computing spaces based on the initial task allocation scheme, divide the parallel ant colony computing spaces 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 screen the global optimal task allocation scheme;

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

[0012] In an alternative embodiment,

[0013] Collect the real-time operation data of the server nodes in the computer room, perform standardization processing on the real-time operation data and construct the total energy consumption function of the computer room, the computing power allocation balance function and the service quality guarantee function, and combine them to obtain a multi-objective optimization function including:

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

[0015] Set the energy consumption index weight, the computing power index weight and the service quality index weight according to the historical operation status of each server node in the computer room, calculate the total energy consumption of the server nodes based on the energy consumption index weight to obtain the total energy consumption function of the computer room, calculate the computing power allocation deviation of the tasks among the server nodes based on the computing power index weight to obtain the computing power allocation balance function, and calculate the task processing delay degree of each server node based on the service quality index weight to obtain the service quality guarantee function;

[0016] Perform normalization processing on the total energy consumption function of the computer room, the computing power allocation balance function and the service quality guarantee function, dynamically adjust the combination coefficients of the functions, and combine the three functions into a multi-objective optimization function based on the combination coefficients.

[0017] In an alternative embodiment,

[0018] Process the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predict the energy consumption change curve and the computing power load change curve of the server nodes, 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 including:

[0019] Divide the standardized real-time operation data into time windows, organize the data in each time window into a feature matrix, the feature matrix includes the energy consumption data, the computing power load data, the temperature data and the power consumption data of the server nodes, and calculate the feature time series correlation coefficient matrix based on the feature matrix;

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

[0021] Calculate the energy consumption margin ratio sequence based on the energy consumption change curve and the maximum energy consumption threshold, calculate the computing power load margin ratio sequence based on the computing power load change curve and the maximum computing power load threshold. According to the average values of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combine the first weight coefficient and the second weight coefficient to calculate the dynamic task - bearing capacity value of the server node;

[0022] Calculate the task allocation weight based on the dynamic task - bearing capacity value and temperature parameters, and multiply the task allocation weight by the multi - objective optimization function and then find the minimum value to obtain the initial task allocation scheme.

[0023] In an alternative implementation,

[0024] Processing the operation data of each server node through the prediction model to obtain the energy consumption change curve and computing power load change curve of the server node in the next time period includes:

[0025] Add the operation data corresponding to each server node to the prediction model, perform principal component analysis to obtain a low - dimensional feature representation, perform a non - linear transformation on the low - dimensional feature representation through an activation function to obtain an initial hidden state, substitute the initial hidden state into the neural ordinary differential equation, and solve the neural ordinary differential equation to obtain a continuous - time hidden state function;

[0026] Calculate the mean and variance of the server node load data to generate a probability distribution function of the server load, sample a random perturbation sequence from the probability distribution function, and superimpose the random perturbation sequence on 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 numerical integration according to the prediction error, perform encrypted sampling on the noisy state sequence at the time points where the prediction error exceeds a preset threshold, and perform numerical integration operations on the sampling points using the fourth - order Runge - Kutta method to obtain a state evolution sequence;

[0028] Calculate the L2 norm value of the state evolution sequence and the energy loss value of the server node operation data respectively. Add the L2 norm value and the energy loss value to obtain the total energy value. Input the total energy value into the second-layer time series prediction unit, calculate the gradient of the total energy value with respect to the network parameters, update the network parameters using the backpropagation algorithm, calculate the causal relationship coefficient between the energy consumption change curve and the computing power load change curve of the server node, and generate a causal dependence graph between nodes.

[0029] Perform an intervention operation on the nodes in the causal dependence graph, calculate the energy change value caused by each intervention operation, compare the energy change value with the law of conservation of thermodynamic energy, correct the prediction data that violates the law of conservation of energy, and output the energy consumption change curve and the computing power load change curve of the server node in the next time period.

[0030] In an alternative embodiment,

[0031] Dividing the parallel ant colony computing space into multiple sub-regions and setting a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, and controlling the convergence direction of the ant colony search path and marking the optimal path includes:

[0032] Divide each parallel ant colony computing space into multiple sub-regions, calculate the temperature decay coefficient based on the server nodes within the sub-regions, dynamically adjust the temperature decay coefficient according to time changes, convert the temperature decay coefficient into a pheromone evaporation rate, perform pheromone update in each parallel ant colony computing space in combination with the task assignment paths within each sub-region, calculate the pheromone retention amount according to the pheromone evaporation rate, accumulate the newly added pheromone increments of all ant individuals on the search path, and add the pheromone retention amount and the pheromone increments to obtain the updated pheromone concentration.

[0033] Calculate the transition probabilities of ant individuals choosing different task assignment schemes in each parallel ant colony computing space respectively, control the convergence direction of the ant colony search path based on the transition probabilities, calculate the fitness values of the task assignment schemes in each parallel ant colony computing space respectively, sort the task assignment schemes in each parallel ant colony computing space in descending order of fitness value, and mark the task assignment scheme with the highest fitness value as the optimal path in each iteration cycle.

[0034] In an alternative embodiment,

[0035] Converting the temperature decay coefficient into a pheromone evaporation rate and performing pheromone update in each parallel ant colony computing space in combination with the task assignment paths within each sub-region includes:

[0036] Collect the temperature decay coefficients corresponding to the server nodes in each sub-region, calculate the task processing capacity difference coefficients of the server nodes in the sub-region based on the temperature decay coefficients, and set independent pheromone evaporation rates for each sub-region according to the task processing capacity difference coefficients;

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

[0038] Detect the pheromone intensity of each task allocation path. If there is a task allocation path whose pheromone intensity exceeds the pre-set path pheromone constraint threshold, obtain the adjacent path set of the task allocation path;

[0039] Calculate the comprehensive path scores of each path in the adjacent path set, calculate the pheromone reallocation weights according to the comprehensive path scores, and allocate the pheromone amount exceeding the path pheromone constraint threshold to the paths in the adjacent path set according to the pheromone reallocation weights to obtain the increased pheromone increment of the ant individual on the search path.

[0040] In an alternative embodiment,

[0041] Calculate the target energy consumption threshold and the target computing power load threshold of each server node under the global optimal task allocation scheme, monitor the operating status of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment between the server nodes based on the judgment result. The optimal matching result includes:

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

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

[0044] Determine the source server node and the target server node for task migration based on the judgment result, perform the task migration operation, and adjust the operating frequencies and power supply voltages of each server node until the real-time energy consumption values and real-time computing power load values of the source server node and the target server node converge within a preset range to obtain an optimal matching result.

[0045] In a second aspect of the embodiments of the present invention, there is provided an optimization system for balancing energy consumption and computing power in a computer room based on swarm intelligence, including:

[0046] A first unit for collecting real-time operation data of server nodes in the computer room, performing standardized processing on the real-time operation data, and constructing a total energy consumption function of the computer room, a computing power distribution balance function, and a service quality guarantee function, and combining them to obtain a multi-objective optimization function;

[0047] A second unit for processing the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predicting the energy consumption change curve and the computing power load change curve of the server node, calculating the dynamic task carrying capacity value and determining the task allocation weight, and generating an initial task allocation plan according to the multi-objective optimization function and the task allocation weight;

[0048] A third unit for constructing a plurality of parallel ant colony calculation spaces based on the initial task allocation plan, dividing the parallel ant colony calculation space into a plurality of sub-regions, setting a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, controlling the convergence direction of the ant colony search path, marking the optimal path, and screening the global optimal task allocation plan;

[0049] A fourth unit for calculating the target energy consumption threshold and the target computing power load threshold of each server node under the global optimal task allocation plan, monitoring the operation status of the server node and generating a judgment result, and performing task migration and energy consumption level adjustment between server nodes based on the judgment result to obtain an optimal matching result.

[0050] In a third aspect of the embodiments of the present invention,

[0051] There is provided an electronic device, including:

[0052] A processor;

[0053] A memory for storing instructions executable by the processor;

[0054] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0055] In a fourth aspect of the embodiments of the present invention,

[0056] Provided is a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0057] In the present invention, by collecting real-time operation data of server nodes and constructing a multi-objective optimization function, combined with the time series prediction ability of a recurrent neural network, the changing trends of server energy consumption and computing power load can be accurately predicted, realizing forward-looking adjustment of task allocation, effectively reducing the overall energy consumption of the computer room, improving the energy utilization efficiency, adopting an innovative mechanism of combining a parallel ant colony algorithm with a temperature attenuation coefficient to dynamically adjust the pheromone evaporation rate, significantly enhancing the search efficiency and optimization quality of the task allocation scheme, avoiding the algorithm from falling into a local optimal solution, ensuring that a globally optimal task allocation scheme can still be found in a complex and changeable server load environment, by setting target energy consumption thresholds and computing power load thresholds, monitoring the running state of the server in real time and performing task migration and energy consumption adjustment, realizing dynamic balance of computer room resources, not only guaranteeing service quality, but also improving the balanced utilization rate of computing power resources, extending the service life of server devices, and reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flowchart of the method for optimizing the balance of energy consumption and computing power in a computer room based on swarm intelligence according to an embodiment of the present invention;

[0059] Figure 2 is a comparison chart of energy consumption prediction errors of the method for optimizing the balance of energy consumption and computing power in a computer room based on swarm intelligence according to an embodiment of the present invention;

[0060] Figure 3 is a comparison chart of pheromone distribution and path convergence of the method for optimizing the balance of energy consumption and computing power in a computer room based on swarm intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only some of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

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

[0063] Figure 1The figure is a schematic flowchart of the method for optimizing the energy consumption and computing power balance of a computer room based on swarm intelligence according to an embodiment of the present invention. As Figure 1 shown, the method includes:

[0064] Collect the real-time operation data of server nodes in the computer room, perform standardization processing on the real-time operation data, and construct a total computer room energy consumption function, a computing power distribution balance function, and a service quality guarantee function, and combine them to obtain a multi-objective optimization function;

[0065] Process the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predict the energy consumption change curve and the computing power load change curve of the server nodes, calculate the dynamic task carrying capacity value and determine the task allocation weight, and generate an initial task allocation plan according to the multi-objective optimization function and the task allocation weight;

[0066] Construct multiple parallel ant colony calculation spaces based on the initial task allocation plan, divide the parallel ant colony calculation spaces into multiple sub-regions, set a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path, mark the optimal path, and screen the global optimal task allocation plan;

[0067] Calculate the target energy consumption threshold and the target computing power load threshold of each server node under the global optimal task allocation plan, monitor the operation status of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment between the server nodes based on the judgment result to obtain the optimal matching result.

[0068] In an alternative embodiment,

[0069] Collect the real-time operation data of server nodes in the computer room, perform standardization processing on the real-time operation data, and construct a total computer room energy consumption function, a computing power distribution balance function, and a service quality guarantee function, and combining them to obtain a multi-objective optimization function includes:

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

[0071] Set the energy consumption index weight, the computing power index weight, and the service quality index weight according to the historical operation status of each server node in the computer room. Based on the energy consumption index weight, calculate the total energy consumption of the server nodes to obtain the total computer room energy consumption function. Based on the computing power index weight, calculate the computing power distribution deviation of tasks among the server nodes to obtain the computing power distribution balance function. Based on the service quality index weight, calculate the task processing delay degree of each server node to obtain the service quality guarantee function;

[0072] Normalize the total energy consumption function of the computer room, the computing power allocation balance function, and the service quality guarantee function, dynamically adjust the combination coefficients of each function, and combine the three functions into a multi-objective optimization function based on the combination coefficients.

[0073] Collect the real-time operation data of server nodes through the data collection module deployed in the computer room. The collected data includes but is not limited to parameters such as CPU utilization rate, memory usage rate, network throughput, storage I / O, power consumption, and temperature. For example, for a computer room with 100 servers, each server generates about 20 monitoring metrics per minute, and the system needs to process about 2000 data points per minute.

[0074] The original collected data has the problem of inconsistent dimensions and needs to be standardized. The maximum-minimum normalization method is used for standardization to convert each index into the range of [0, 1]. Taking the CPU utilization rate as an example, if the original data range is 0% - 100%, the standardized value is the original value divided by 100. For power consumption data, if the power consumption range of a certain server is 100W - 500W, the standardization process maps 100W to 0, 500W to 1, and other values are mapped proportionally.

[0075] The system sets three types of weight indicators according to the historical operation status of each server node. For the energy consumption index weight, the system analyzes the energy efficiency performance of each server in the past 30 days. Servers with a higher energy efficiency ratio are given a lower energy consumption weight, and servers with a lower energy efficiency ratio are given a higher energy consumption weight. For example, server A with an energy efficiency ratio of 2.5 is given a weight of 0.4, while server B with an energy efficiency ratio of 1.8 is given a weight of 0.7.

[0076] The computing power index weight is determined based on the computing power and stability of the server. Servers with strong computing power and stability obtain a higher computing power weight, and vice versa. For example, a server configured with 32 cores, 128GB of memory, and a stable operation time exceeding 99.9% may obtain a computing power weight of 0.9, while a server with 16 cores, 64GB of memory, and a stable operation time of 98% may obtain a computing power weight of 0.6.

[0077] The service quality index weight is determined based on the historical delay performance of the server in processing tasks. Servers with short response times and fast processing speeds obtain a higher service quality weight, and vice versa. For example, a server with an average response time of 10ms may obtain a service quality weight of 0.85, while a server with an average response time of 50ms may obtain a service quality weight of 0.5.

[0078] Based on the above weights, three key functions are constructed. The first key function is the total energy consumption function of the computer room, which is obtained by calculating the weighted sum of the energy consumption of each server node. Suppose there are n servers in the computer room, the normalized energy consumption of the i-th server is Ei, and the weight of the energy consumption index is Wi. Then the total energy consumption function of the computer room is the sum of the products of the energy consumption of each server and its weight. For a small computer room with 10 servers, if the normalized 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 weights of the energy consumption index are [0.8, 0.7, 0.6, 0.7, 0.9, 0.5, 0.7, 0.8, 0.6, 0.5], then the calculated value of the total energy consumption function of the computer room is 4.31.

[0079] The computing power allocation balance function is obtained by calculating the deviation of the computing power allocation among each server node. The system calculates the proportion of computing power that should be allocated to each server under ideal conditions, calculates the deviation between the actual allocation and the ideal allocation, multiplies the deviation by the weight of the computing power index, and sums them up. 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], and 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 weight of the computing power index is [0.9, 0.8, 0.7, 0.8, 0.9, 0.6, 0.8, 0.9, 0.7, 0.6], then the calculated value of the computing power allocation balance function is 0.0294.

[0080] The service quality guarantee function is obtained by calculating the task processing delay 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 value, multiplies the deviation by the weight of the service quality index, and sums them up. For example, if the normalized delay 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 weight of the service quality index is [0.9, 0.85, 0.75, 0.8, 0.9, 0.7, 0.8, 0.85, 0.75, 0.7], then the calculated value of the service quality guarantee function is 0.0783.

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

[0082] The system dynamically adjusts the combination coefficients of each function according to the current operating status of the computer room. During high load periods, more attention may be paid to service quality, increasing the weight of the service quality guarantee function; during low load periods, more attention may be paid to energy efficiency, increasing the weight of the total computer room energy consumption function. For example, during the peak period during the day on weekdays, the combination coefficient may be set as energy consumption: computing power: service quality = 0.3:0.3:0.4; while during the low period at night, the combination coefficient may be adjusted to 0.5:0.3:0.2.

[0083] The system combines the three normalized functions into a multi-objective optimization function based on the combination coefficient. Assuming that the normalized value of the total computer room energy consumption function is 0.4765, the computing power allocation balance function value is 0.3520, the service quality guarantee function value is 0.2890, and the combination coefficient is [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 evenly allocate computing power resources while ensuring service quality, improve the overall operation efficiency, and realize the comprehensive optimization of computer room energy consumption, computing power allocation, and service quality by minimizing the multi-objective optimization function value, providing a solid theoretical basis for subsequent optimization.

[0085] In an alternative embodiment,

[0086] Processing the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predicting the energy consumption change curve and computing power load change curve of the server node, calculating the dynamic task carrying capacity value and determining the task allocation weight, and generating an initial task allocation plan according to the multi-objective optimization function and the task allocation weight includes:

[0087] Dividing the standardized real-time operation data into time windows, organizing the data within each time window into a feature matrix, the feature matrix including the energy consumption data, computing power load data, temperature data, and power consumption data of the server node, and calculating the feature time series correlation coefficient matrix based on the feature matrix;

[0088] Constructing a two-layer prediction model according to the time series prediction unit in the recurrent neural network, the prediction model including a first-layer time series prediction unit and a second-layer time series prediction unit, the first-layer time series prediction unit processing the feature time series correlation coefficient matrix to obtain a hidden state, the second-layer time series prediction unit updating network parameters based on the hidden state, and processing the operation data of each server node through the prediction model to obtain the energy consumption change curve and computing power 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. According to the average values of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combined with the first weight coefficient and the second weight coefficient, the dynamic task carrying capacity value of the server node is calculated;

[0090] Based on the dynamic task carrying capacity value and the temperature parameter, a task allocation weight is calculated. After multiplying the task allocation weight by the multi-objective optimization function, the minimum value is obtained to get the initial task allocation scheme.

[0091] The standardized real-time operation data is divided into time windows, and the length of each time window is determined according to the server task processing cycle. In each time window, the system continuously collects the operation status data of the server node, including the real-time energy consumption value, the computing power load value reflected by the CPU usage rate, the temperature data collected by the temperature sensor, and the power consumption data calculated by the power. These data are arranged according to the sampling time sequence to construct an m×n feature matrix, where m represents the number of sampling points and n represents the feature dimension. The Pearson correlation coefficient is calculated for each index data in the feature matrix to obtain a feature time series correlation coefficient matrix reflecting the correlation intensity between each index.

[0092] The prediction model is constructed using a double-layer recurrent neural network structure. The first-layer time series prediction unit receives the feature time series correlation coefficient matrix, controls the input proportion of the current moment data through the input gate, controls the retention degree of historical information through the forget gate, and controls the output proportion of the current state through the output gate. The gated units work together to convert the input data into a hidden state vector containing time series features. The second-layer time series prediction unit receives the hidden state of the first layer and dynamically adjusts the network parameters according to the state information, including the weight matrix and the bias vector. The parameter update uses the backpropagation algorithm to minimize the prediction error. After the model training is completed, inputting the data of the current time window can predict the energy consumption change trend and the computing power load change trend of the server node in the next period.

[0093] Calculation of dynamic task carrying capacity: Obtain the predicted energy consumption change curve, subtract the predicted value of each time point on the curve from the preset maximum energy consumption threshold and then divide by the maximum threshold to obtain a ratio sequence reflecting the remaining energy consumption capacity. Normalize the difference between the predicted value of the computing power load change curve and the load maximum threshold to obtain a ratio sequence of the remaining computing power capacity. Calculate the arithmetic mean of the two sequences in the prediction time period, which respectively represent the average energy consumption margin ratio and the average computing power margin ratio. Set the energy consumption weight coefficient and the computing power weight coefficient, and obtain the comprehensive dynamic task carrying capacity value through weighted summation. The larger this value is, the more sufficient the remaining resources of the server node are.

[0094] The generation of the task allocation scheme is based on the calculated dynamic task carrying capacity value. Combining with the exponential decay function reflecting the influence of server temperature (the higher the temperature, the faster the decay), the task allocation weights of each node are calculated. Multiply the task allocation weights by the multi-objective optimization function considering objectives such as minimizing energy consumption and load balancing. By solving the minimum value problem, the initial task allocation ratio scheme that meets the resource constraints is obtained.

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

[0096] Set the energy consumption threshold to 100 W and the load threshold to 80%. Calculate the 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; 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 weight coefficients are taken as 0.6 and 0.4 respectively. Calculate the dynamic task carrying capacity value of Server 1 = 0.6×0.12 + 0.4×0.15 = 0.132. Considering the temperature influence of 46 °C, the compensation coefficient is taken as 0.95. Finally, the task allocation weights of the three nodes are [0.125, 0.142, 0.138]. Combining with the multi-objective optimization function to solve, the task allocation scheme [35%, 33%, 32%] is obtained.

[0097] In this embodiment, through the methods of time window division and feature extraction, the dynamic change features of the running state of server nodes are effectively captured. By constructing a feature time series correlation coefficient matrix, the correlation relationships among multi-dimensional indicators such as energy consumption, computing power load, temperature, and power consumption are comprehensively reflected, providing a reliable data basis for subsequent prediction. The hierarchical structure not only improves the model's perception ability of changes in the server running state but also enhances the reliability of the prediction results. Introducing a temperature compensation mechanism can effectively avoid performance losses and stability risks caused by overheating of server nodes. Considering both the real-time running state of server nodes and the change trends in a future period of time can effectively improve the resource utilization efficiency of the data center, reduce the overall energy consumption level, and ensure the stable operation of the server cluster.

[0098] In an alternative embodiment,

[0099] Processing the running data of each server node through the prediction model to obtain the energy consumption change curve and computing power load change curve of the server node in the next time period includes:

[0100] Adding the running data corresponding to each server node to the prediction model, performing principal component analysis to obtain a low-dimensional feature representation, performing a non-linear transformation on the low-dimensional feature representation through an activation function to obtain an initial hidden state, substituting the initial hidden state into a neural ordinary differential equation, and solving the neural ordinary differential equation to obtain a continuous-time hidden state function;

[0101] Calculating the mean and variance of the server node load data to generate a probability distribution function of the server load, sampling a random perturbation sequence from the probability distribution function, and superimposing the random perturbation sequence on the continuous-time hidden state function to obtain a noisy state sequence;

[0102] Calculating the prediction error of the noisy state sequence, determining the step size of numerical integration according to the prediction error, performing encrypted sampling on the noisy state sequence at the time points where the prediction error exceeds a preset threshold, and performing a numerical integration operation on the sampling points using the fourth-order Runge-Kutta method to obtain a state evolution sequence;

[0103] Calculating the L2 norm value of the state evolution sequence and the energy loss value of the server node running data respectively, adding the L2 norm value and the energy loss value to obtain a total energy value, inputting the total energy value into the second-layer time series prediction unit, calculating the gradient of the total energy value with respect to the network parameters, updating the network parameters using the backpropagation algorithm, calculating the causal relationship coefficient between the energy consumption change curve and the computing power load change curve of the server node, and generating a causal dependence relationship graph between nodes;

[0104] Perform intervention operations on the nodes in the causal dependency graph, calculate the energy change values caused by each intervention operation, compare the energy change values with the law of conservation of thermodynamic energy, correct the predicted data that violates the law of conservation of energy, and output the energy consumption change curve and computing power load change curve of the server node in the next time period.

[0105] Add the operation data corresponding to each server node to the prediction model. The operation data includes CPU usage rate, memory occupancy rate, network traffic, disk read and write speed, power consumption data, etc. Taking a group of servers in a certain data center as an example, collect the operation data of 10 servers every 5 minutes in the past 24 hours to form a multi-dimensional time series data set.

[0106] Perform principal component analysis on the collected operation data to reduce 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 through principal component analysis is [0.56, -0.23, 0.78, 0.12, -0.45, 0.33, -0.67, 0.21].

[0107] Perform a non-linear transformation on the low-dimensional feature representation through the ReLU activation function to obtain the initial hidden state. For example, after applying the ReLU activation function to the above low-dimensional features, the initial hidden state obtained is [0.56, 0, 0.78, 0.12, 0, 0.33, 0, 0.21].

[0108] Substitute the initial hidden state into the neural ordinary differential equation. The neural ordinary differential equation consists of a multi-layer fully connected network, with the input being the hidden state at the current time point and the output being the change rate of the hidden state. By solving this equation, a continuous-time hidden state function is obtained. Specifically, set the time interval as [0, 1], divide it evenly into 100 time points, calculate the values of the hidden state at each time point, and form a continuous-time hidden state function.

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

[0110] Calculate the prediction error of the noisy state sequence, and take the average of the sum of the squares of the differences between the predicted values and the actual observed values. Determine the step size of numerical integration according to 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 the time points when the prediction error exceeds a preset threshold (such as 0.08), perform encrypted sampling on the noisy state sequence. Use the fourth-order Runge-Kutta method to perform numerical integration operations on the sampling points to obtain the state evolution sequence. The fourth-order Runge-Kutta method estimates the state value at the next time point through the weighted average of four intermediate points, improving the accuracy of numerical integration.

[0112] Calculate the L2 norm value of the state evolution sequence and the energy loss value of the server node operation data respectively. The calculation method of the L2 norm value is to take 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 value is 0.726. The energy loss value is obtained by calculating the mean square error between the predicted value and the actual value, such as 0.023.

[0113] Add the L2 norm value and the energy loss value to obtain the total energy value, such as 0.749. Input the total energy value into the second-layer time series prediction unit, which is composed of a long short-term memory network (LSTM) and contains 64 hidden units. Calculate the gradient of the total energy value with respect to the network parameters, and use the backpropagation algorithm to update the network parameters. The learning rate is set to 0.001, 32 samples are processed per batch, and 100 epochs are trained.

[0114] Calculate the causal relationship coefficient between the energy consumption change curve and the computing power load change curve of the server node. Use the Granger causality test method 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 the CPU load has a significant causal impact on energy consumption. Generate a causal dependency graph between nodes according to the causal relationship coefficients between nodes.

[0115] Perform intervention operations on the nodes in the causal dependency graph, such as increasing the CPU load of server B by 10%, and observe its impact on its own and other servers' energy consumption. Calculate the energy change value caused by each intervention operation. For example, the energy consumption of server B increases by 5.2%, and the energy consumption of server C increases by 1.3%. Compare the energy change value with the law of conservation of thermodynamic energy, and correct the prediction data that violates the law of conservation of energy. For example, if the prediction result shows that the total energy increases more than the input energy, scale the energy changes of each node proportionally to make the total energy conserved.

[0116] Output the energy consumption change curve and computing power load change curve of the server node in the next time period. For example, it is predicted that within the next 4 hours, the energy consumption of Server A will first increase and then decrease, with the peak occurring at the 2nd hour and the increase rate being 8%; the CPU load will fluctuate stably within the range of 75% ± 3%. These prediction results can be used for energy management and load balancing in the data center to improve energy utilization efficiency.

[0117] In this embodiment, dimensionality reduction processing of high-dimensional operation data is achieved through principal component analysis, effectively extracting the key feature information of the server node and reducing the complexity of subsequent calculations. Through the non-linear transformation of the ReLU activation function, the expression ability of the model for non-linear features is enhanced, and the effectiveness of feature extraction is improved. A neural ordinary differential equation is introduced to model the server state. Compared with the traditional discrete time series model, it can more accurately describe the continuous change process of the server state. By adding random perturbations to the state sequence, the adaptability of the model to unknown disturbances is improved, and a numerical integration strategy with an adaptive step size is adopted to optimize the calculation efficiency while ensuring the calculation accuracy.

[0118] In the prior art, the prediction of the operating state of server nodes usually uses a single prediction model to directly process the original high-dimensional data, which not only has a high calculation 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 relationship between server nodes and are difficult to accurately describe the overall dynamic change characteristics of the data center.

[0119] In this embodiment, through the collaborative action of the double-layer time series prediction unit, accurate prediction of the energy consumption and computing power load of the server node is achieved. The efficiency of feature extraction is improved through dimensionality reduction and non-linear transformation. The reliability of the prediction results is enhanced by introducing physical model constraints. The double-layer prediction structure is adopted to enhance the adaptability of the model, and it can more accurately predict the operating state of the server node, providing a more reliable decision-making basis for energy management and load balancing in the data center.

[0120] Figure 2 This is the comparison chart of energy consumption prediction errors for the energy consumption and computing power balance optimization method of the computer room based on swarm intelligence in the embodiment of the present invention, showing the comparison of energy consumption prediction errors of different prediction models in each prediction time range. Using a visualization method combining a bar chart and a line, the performance differences of each method under different prediction durations are clearly shown. The technical solution of this application uses a method combining neural ordinary differential equations and causal inference, showing obvious advantages in all prediction time ranges.

[0121] Specifically, in the short-term 1-hour prediction, the prediction error of this technical solution is only 2.1%, while that of the traditional LSTM method is 4.2% and that of the GRU method is 5.2%. As the prediction time extends, the prediction errors of all methods show an upward trend, but the error growth of this technical solution is the slowest. In the medium-term prediction (6 hours), the prediction error of this technical solution is 6.0%, far lower than 12.3% of LSTM and 13.7% of GRU. In the long-term prediction (24 hours), the prediction error of this technical solution is 10.8%, while that of LSTM reaches 21.9% and that of GRU reaches 24.1%. This indicates that this technical solution has a significant anti-error accumulation ability in long-term prediction, thanks to the fact that the neural ordinary differential equation adopted can accurately capture the dynamic characteristics of the system, and the energy conservation constraint correction mechanism effectively prevents prediction deviation. The height of the bar chart intuitively shows the error gap between different methods, while the connecting lines clearly indicate the error growth trend of each method as the prediction time extends. The connecting line of this technical solution is significantly lower than that of other methods, proving its excellent prediction stability.

[0122] In an alternative embodiment,

[0123] Dividing the parallel ant colony calculation space into multiple sub-regions and setting a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, and controlling the convergence direction of the ant colony search path and marking the optimal path includes:

[0124] Dividing each parallel ant colony calculation space into multiple sub-regions, calculating the temperature decay coefficient based on the server nodes within the sub-regions, dynamically adjusting the temperature decay coefficient according to the time change, converting the temperature decay coefficient into the pheromone evaporation rate, performing pheromone update in each parallel ant colony calculation space in combination with the task assignment paths within each sub-region, calculating the pheromone retention amount according to the pheromone evaporation rate, accumulating the pheromone increment newly added by all ant individuals on the search path, and adding the pheromone retention amount and the pheromone increment to obtain the updated pheromone concentration;

[0125] Calculating the transition probabilities of ant individuals selecting different task assignment schemes in each parallel ant colony calculation space respectively, controlling the convergence direction of the ant colony search path based on the transition probabilities, calculating the fitness values of the task assignment schemes in each parallel ant colony calculation space respectively, sorting the task assignment schemes in each parallel ant colony calculation space in descending order of the fitness values, and marking the task assignment scheme with the highest fitness value as the optimal path in each iteration cycle.

[0126] Divide each parallel ant colony computing space into multiple sub-regions. Suppose there are N server nodes, which can be divided into M sub-regions according to the physical location or logical relationship of the nodes. For example, in a data center environment with 100 server nodes, it can be divided into 10 sub-regions, with each sub-region containing 10 server nodes. The division method can adopt the K-means clustering algorithm and group based on the network latency or geographical location between nodes.

[0127] Calculate the temperature decay coefficient based on the server nodes within the 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 comprehensively calculated through the average CPU utilization rate, memory usage rate, and energy consumption level of all server nodes within the sub-region. For example, when the average CPU utilization rate within the sub-region is 60%, the average memory usage rate is 50%, and the average energy consumption is 300W, the initial temperature decay coefficient can be set to 0.85.

[0128] Dynamically adjust the temperature decay coefficient according to the change of time. As the number of iterations increases, the temperature decay coefficient needs to gradually decrease to accelerate the convergence speed of the algorithm. Suppose the initial temperature decay coefficient is T0, the current number of iterations 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, at the 50th iteration, if the initial temperature decay coefficient is 0.85 and the maximum number of iterations is 200, the current temperature decay coefficient can be updated to 0.7375.

[0129] Convert the temperature decay coefficient into the pheromone evaporation rate. The pheromone evaporation rate ρ has a positive correlation 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] Combine the task assignment paths within each sub-region to perform pheromone update in each parallel ant colony computing space. Suppose there are K tasks to be assigned to N server nodes, and each ant constructs a complete task assignment path from the starting point to the ending point. The pheromone concentration on the path where the kth task is assigned to the nth server node needs to be updated.

[0131] Calculate the pheromone retention amount according to the pheromone evaporation rate. For the pheromone concentration τ(k, n) on the path (k, n), its 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] Accumulate the pheromone increment newly added by all ant individuals on the search path. Suppose there are ANT_NUM ants, and the pheromone increment left by the a-th ant on the path (k, n) is Δτa(k, n), then the total increment is the accumulation of the increments of all ants. The pheromone increment is proportional to the quality of the solution constructed by the ant. For example, if the increment of the first ant on this path is 0.3 and the increment of the second ant is 0.25, then the total increment is 0.55.

[0133] Add the pheromone retention amount and the pheromone increment to obtain the updated pheromone concentration. The updated pheromone concentration τnew(k, n) is equal to the pheromone retention amount plus the pheromone increment. Continuing with the above example, the updated pheromone concentration is 1.84375 + 0.55 = 2.39375.

[0134] Calculate the transition probabilities for ant individuals to select different task allocation schemes in each parallel ant colony computing space respectively. The transition probability is related to the pheromone concentration and heuristic information on the path. The heuristic information can be the matching degree between the processing capacity of the server node and the task requirements. For example, for CPU-intensive tasks, servers with stronger CPU performance can be preferentially selected. Suppose the pheromone concentration of the path (k, n1) is 2.39375 and the heuristic value is 0.8, and the pheromone concentration of the path (k, n2) is 1.5 and the heuristic value is 0.6, then the probability of selecting n1 is approximately 0.68 and the probability of selecting n2 is approximately 0.32.

[0135] Control the convergence direction of the ant colony search path based on the transition probability. During the search process, the ant selects the next moving direction according to the transition probability. By adjusting the pheromone evaporation rate, the convergence speed and direction of the algorithm can be controlled. A lower evaporation rate is beneficial to maintaining diversity, and a higher evaporation rate is beneficial to accelerating convergence.

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

[0137] Sort the task allocation schemes in each parallel ant colony computing space in descending order of fitness value. Suppose there are 5 parallel ant colony computing spaces, and there are 20 task allocation schemes in each space. Then, it is necessary to sort the 20 schemes in each space separately.

[0138] In each iteration cycle, mark the task allocation scheme with the highest fitness value 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, then mark it as the optimal path of the current iteration. This optimal path will be used 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 computing efficiency of the algorithm is improved. The sub-region division strategy fully considers the physical location and logical relationship of 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 state of the local server cluster. By dynamically adjusting the temperature decay coefficient, the adaptive adjustment of the algorithm search intensity is realized. By simultaneously performing ant colony search in multiple parallel computing spaces, the search efficiency of the algorithm is significantly improved.

[0140] In an alternative embodiment,

[0141] Converting the temperature decay coefficient into a pheromone evaporation rate and performing pheromone update in each parallel ant colony computing space in combination with the task allocation paths in each sub-region includes:

[0142] Collect the temperature decay coefficients corresponding to the server nodes in each sub-region, calculate the task processing capacity difference coefficient of the server nodes 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 capacity difference coefficient;

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

[0144] Detect the pheromone intensity of each task allocation path. If there is a task allocation path whose pheromone intensity exceeds the pre-set path pheromone constraint threshold, then obtain the adjacent path set of the task allocation path;

[0145] Calculate the comprehensive path scores of each path in the set of adjacent paths, calculate the pheromone reallocation weights based on the comprehensive path scores, and allocate the pheromone amount exceeding the pheromone constraint threshold to the paths in the set of adjacent paths according to the pheromone reallocation weights, so as to obtain the newly added pheromone increment of the ant individual on the search path.

[0146] The system collects the temperature decay coefficients 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 through real-time monitoring by 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], indicating the differences in the heat dissipation capabilities of different nodes.

[0147] Calculate the task processing ability difference coefficients of the server nodes in the sub-region based on the temperature decay coefficients. Specifically, in implementation, the temperature decay coefficients can be compared with a preset reference value (such as 0.90) to calculate the relative deviation values. For example, for the server nodes in the above-mentioned sub-region A, its task processing ability difference coefficients can be calculated as [0.022, -0.022, 0.056, 0.000, -0.033].

[0148] Set independent pheromone evaporation rates for each sub-region according to the task processing ability difference coefficients. The pheromone evaporation rate is positively correlated with the task processing ability difference coefficient. The greater the difference, the higher the evaporation rate is set to promote the dynamic adjustment of task allocation. For example, the pheromone evaporation rate of sub-region A can be set to 0.15, while that of sub-region B with relatively balanced temperature decay coefficients can be set to 0.10.

[0149] Record the task allocation paths of the ant individuals in each sub-region. The task allocation path is represented as a sequence of accesses to a series of server nodes. For example, the path P1 = [S1 → S3 → S5 → S2] means that the task is sequentially allocated to the server nodes S1, S3, S5, and S2.

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

[0151] Weight the load balancing index and the energy consumption efficiency index to obtain the comprehensive path score. 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 consumption efficiency weight is 0.4, the comprehensive score of path P1 is 0.78×0.6 + 0.65×0.4 = 0.728.

[0152] Map the path comprehensive score to the pheromone intensity adjustment coefficient according to a pre-set mapping relation table. The mapping relation table can be designed as a piecewise linear mapping. For example, for a path with a comprehensive score in the interval [0.7, 0.8], its pheromone intensity adjustment coefficient is 1.2; for a path with a comprehensive score in the interval [0.8, 0.9], its pheromone intensity adjustment coefficient is 1.5. For path P1, the pheromone intensity adjustment coefficient corresponding to its comprehensive score of 0.728 is 1.2.

[0153] Calculate the path pheromone constraint threshold based on the pheromone intensity adjustment coefficient. The path pheromone constraint threshold is dynamically adjusted with the path comprehensive score. The higher the comprehensive score, the higher the constraint threshold. For example, the benchmark threshold can be set to 10, and the pheromone constraint threshold of path P1 is 10×1.2 = 12.

[0154] The system detects the pheromone intensity of each task assignment path. If there is a task assignment path whose pheromone intensity exceeds the pre-set path pheromone constraint threshold, obtain the set of adjacent paths of this task assignment path. An adjacent path is defined as a path that differs from the current path by no more than a certain degree, such as a path that differs by only one node. For example, the set of adjacent paths of path P1 = [S1→S3→S5→S2] can include P2 = [S1→S3→S5→S4], P3 = [S1→S3→S4→S2], etc.

[0155] Suppose the pheromone intensity of path P1 is 15, which exceeds its constraint threshold of 12. The system will calculate the path comprehensive score of each path in its set of adjacent paths. For example, the comprehensive score of P2 is 0.685, and the comprehensive score of P3 is 0.710.

[0156] Calculate the pheromone reallocation weight according to the path comprehensive score. The higher the comprehensive score, the greater the reallocation weight. For example, the reallocation weights of P2 and P3 are 0.49 and 0.51 respectively.

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

[0158] Through the pheromone reallocation mechanism, the system obtains the new pheromone increment of the ant individuals on the search path. This mechanism ensures the rationality of the pheromone distribution, prevents the over-concentration of pheromone on specific paths, and improves the exploration ability and convergence efficiency of the ant colony algorithm in task assignment.

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

[0160] In the prior art, the server task allocation usually adopts a fixed pheromone update strategy, which fails to fully consider the physical characteristics and dynamic change characteristics of server nodes, easily leading to unbalanced pheromone distribution and affecting the convergence performance of the algorithm. At the same time, traditional solutions often take load balancing and energy consumption efficiency as independent optimization goals, making it difficult to achieve dynamic balance among multiple goals;

[0161] This embodiment improves the pertinence of task allocation through temperature characteristic analysis, enhances the adaptive ability of the algorithm through differential pheromone regulation strategies, improves the overall effect of task allocation through multi-objective joint optimization and pheromone reallocation mechanisms, and significantly improves the accuracy and efficiency of task allocation through a multi-level optimization mechanism, providing strong support for the intelligent operation and maintenance of data centers.

[0162] Figure 3 This is a comparison chart of pheromone distribution and path convergence for the method of optimizing the balance of computer room energy consumption and computing power based on swarm intelligence in the embodiments of the present invention, showing the pheromone distribution and path convergence of four algorithms in server task allocation.

[0163] The temperature attenuation coefficients of the 11 server nodes (S1 - S11) in the figure are marked, with the value range between 0.85 and 0.95, and the temperature characteristics are significantly different. The selected path (thick solid line) of this technical solution is S1→S3→S8→S9→S11, and the pheromone intensities along the way are 8.5, 9.2, 7.8, and 8.1 respectively, all lower than the set constraint threshold of 10.0, reflecting the effective control of pheromone intensity by this solution. The pheromone intensities on the path S1→S2→S7→S11 (medium dashed line) selected by the standard ACO algorithm are as high as 12.6, 11.8, and 14.2, far exceeding the reasonable threshold, resulting in premature convergence to a local optimal solution.

[0164] Although the Max - Min ant colony algorithm improves this problem by setting upper and lower bounds, the pheromone intensities on the selected path S1→S6→S7→S8→S11 (thin dashed line) are 10.2, 9.8, 9.5, and 9.0, without considering the temperature characteristic differences of the servers.

[0165] The temperature-aware priority algorithm takes into account the temperature factor. The pheromone intensities on the path S1→S3→S4→S9→S11 (dotted line) are 9.7, 8.9, 8.3, and 8.5, but it lacks the dynamic pheromone reallocation mechanism of this technical solution. It can be clearly seen from the figure that by integrating the temperature decay coefficient and the pheromone dynamic adjustment mechanism, this technical solution selects a more reasonable server access path while ensuring a moderate distribution of pheromones, avoiding server nodes with a lower temperature decay coefficient (poor heat dissipation ability) (such as S10 and S5), and achieving a dual optimization of load balancing and energy efficiency.

[0166] Table 1 is a comparison table of multi-dimensional performance indicators for the optimization method of computer room energy consumption and computing power balance based on swarm intelligence, showing the comparison results of five algorithms on six key performance indicators. This technical solution performs best in all indicators, with a comprehensive score of 9.2 points, far higher than the 7.1 points of the standard ACO algorithm.

[0167] Table 1 is a comparison table of multi-dimensional performance indicators for the optimization method of computer room energy consumption and computing power balance based on swarm intelligence:

[0168]

[0169] Specifically, this technical solution reaches 0.935 in terms of load balancing degree, 0.09 percentage points higher than the standard ACO algorithm; the energy consumption efficiency is 0.882, 0.087 higher than the standard ACO; the convergence speed only requires 42 iterations, almost half of the standard ACO algorithm (78 times); the high-temperature adaptability reaches 0.875, 0.16 higher than the standard ACO algorithm; the task throughput is 3450 tasks / minute, 21% higher than the standard ACO algorithm; the scheduling fairness reaches 0.915, 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 degree, 0.057 higher in energy consumption efficiency, and 35% faster in convergence speed; compared with the temperature-aware priority algorithm, although the gap is reduced, this technical solution still has obvious advantages in terms of load balancing degree, energy consumption efficiency, and scheduling fairness, being 0.033, 0.037, and 0.05 higher respectively; compared with the dynamic temperature-aware algorithm, this technical solution maintains an overall but small leading advantage.

[0170] The above data fully demonstrate the effectiveness of the temperature decay coefficient calculation and pheromone reallocation mechanism in this technical solution, enabling it to maintain stable high performance under various working conditions. Especially in terms of high-temperature adaptability, the gaps between this technical solution and the temperature-aware priority algorithm and the dynamic temperature-aware algorithm are relatively small (0.045 and 0.030 respectively), indicating that all three algorithms effectively consider the temperature factor, but this technical solution still has a significant advantage in overall performance through a better pheromone adjustment strategy.

[0171] In an alternative embodiment,

[0172] Calculate the target energy consumption threshold and the target computing power load threshold for each server node under the global optimal task allocation scheme, monitor the operating status of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment among the server nodes based on the judgment result. The optimal matching result includes:

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

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

[0175] Determine the source server node and the target server node for task migration based on the judgment result, perform the task migration operation and adjust the operating frequency and power supply voltage of each server node until the real-time energy consumption value and the real-time computing power load value of the source server node and the target server node converge within a preset interval to obtain the optimal matching result.

[0176] Obtain the historical task processing records of the server nodes in the global optimal task allocation scheme. The system extracts the operating data of each server node in the optimal task allocation state in the past 30 days from the database, including the energy consumption data and the computing power load data per hour. For example, for a cluster containing 10 server nodes, each node will have 720 historical records (30 days × 24 hours).

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

[0178] Calculate the target energy consumption threshold and the target computing power load threshold for each server node. The system statistically analyzes the historical energy consumption data of each server node and takes the average energy consumption in its optimal operating state as the target energy consumption threshold. At the same time, considering the energy consumption fluctuation range, the upper and lower floating intervals are set at ±5%. Similarly, analyze the historical computing power load data, take the average load in its optimal operating state as the target computing power load threshold, and set the upper and lower floating intervals at ±10%. For example, the target energy consumption threshold of server node 1 is 0.85 kWh / h (the floating interval is 0.81 - 0.89 kWh / h), and the target computing power load threshold is 65% (the floating interval is 58.5% - 71.5%).

[0179] In the real-time monitoring stage, the system collects the real-time energy consumption value and the real-time computing power load value of the server node every minute. The collection methods include obtaining indicators such as power data, CPU utilization rate, memory usage rate, and network throughput through the server management interface. For example, at a certain moment, the real-time energy consumption value of server node 1 is 0.95 kWh / h, and the real-time computing power load value is 75%.

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

[0181] Among them, the judgment results include: normal state: the real-time value is within the floating interval 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; compound abnormal state: both energy consumption and load anomalies occur simultaneously.

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

[0183] During the task migration operation, the system selects tasks suitable for migration from the task queue of the source node. The selection criteria include factors such as task priority, resource requirements, and migration cost. For example, for the 10 tasks running on server node 1, the system selects 2 tasks with high resource requirements but low priority for migration.

[0184] Evaluate the resource availability of the target node to ensure that the target node has sufficient resources to receive the migration task. For example, the current CPU utilization rate of server node 2 is 40% and the memory usage rate is 35%, which is sufficient to receive 2 tasks from node 1 (it is expected to increase the CPU utilization rate by 15% and the memory usage rate by 10%).

[0185] Execute task migration to transfer the selected tasks from the source node to the target node. The migration process includes steps such as task status saving, network transmission, and task recovery on the target node. For example, the system used live migration technology to complete the migration of 2 tasks within 100 milliseconds, during which the task execution was not interrupted.

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

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

[0188] Record the optimal matching results, including the task allocation, energy consumption level, and computing power load level of each server node, as new historical data for future threshold calculation and optimization. For example, after this optimization, the average energy consumption efficiency of the entire cluster increased by 8%, and the task processing capacity increased by 5%, achieving the expected optimization goal.

[0189] In this embodiment, through the real-time monitoring and dynamic judgment mechanism, it is possible to accurately identify abnormal operating states of server nodes, timely determine the source node and target node that need to perform task migration, ensuring the pertinence and necessity of task migration. By adopting a multi-dimensional task evaluation criterion, comprehensively considering factors such as task priority, resource requirements, and migration cost, the most suitable tasks for migration are screened from the source nodes, effectively reducing the impact of migration on the overall system performance. Through the precise evaluation of the resource availability of the target node, the system can accurately predict the resource usage status 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] An energy consumption and computing power balance optimization system for computer rooms based on swarm intelligence, comprising:

[0191] A first unit, configured to collect real-time operation data of server nodes in a computer room, perform standardization processing on the real-time operation data, construct a total energy consumption function, a computing power distribution balance function, and a service quality guarantee function for the computer room, and combine them to obtain a multi-objective optimization function;

[0192] A second unit, configured to process the standardized real-time operation data according to the time series prediction unit of a recurrent neural network, predict the energy consumption change curve and the computing power load change curve of server nodes, 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] A third unit, configured to construct a plurality of parallel ant colony computing spaces based on the initial task allocation scheme, divide the parallel ant colony computing spaces into a plurality of sub-regions, set a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path, mark the optimal path, and screen the global optimal task allocation scheme;

[0194] A fourth unit, configured to calculate the target energy consumption threshold and the target computing power load threshold of each server node under the global optimal task allocation scheme, monitor the operation state of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment among the server nodes based on the judgment result to obtain an optimal matching result.

[0195] In the third aspect of the embodiments of the present invention,

[0196] There is provided an electronic device, comprising:

[0197] A processor;

[0198] A memory for storing instructions executable by the processor; [[ID=URL]]

[0199] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0200] In the fourth aspect of the embodiments of the present invention,

[0201] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0202] The present invention can be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

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

Claims

1. An optimization method for balancing the energy consumption and computing power of a computer room based on swarm intelligence, characterized in that, Including: Collecting the real-time operation data of server nodes in the computer room, performing standardized processing on the real-time operation data, constructing the total energy consumption function, computing power allocation balance function, and service quality guarantee function of the computer room, and combining them to obtain a multi-objective optimization function; Processing the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predicting the energy consumption change curve and computing power load change curve of the server nodes, calculating the dynamic task carrying capacity value and determining the task allocation weight, and generating an initial task allocation plan according to the multi-objective optimization function and the task allocation weight; Constructing multiple parallel ant colony computing spaces based on the initial task allocation plan, dividing the parallel ant colony computing spaces into multiple sub-regions, setting the temperature decay coefficient to dynamically adjust the pheromone evaporation rate, controlling the convergence direction of the ant colony search path and marking the optimal path, and screening the global optimal task allocation plan; Calculating the target energy consumption threshold and target computing power load threshold of each server node under the global optimal task allocation plan, monitoring the operation status of the server nodes and generating a judgment result, and performing task migration and energy consumption level adjustment between the server nodes based on the judgment result to obtain the optimal matching result.

2. The method according to claim 1, wherein Collecting the real-time operation data of server nodes in the computer room, performing standardized processing on the real-time operation data, and constructing the total energy consumption function, computing power allocation balance function, and service quality guarantee function of the computer room, and combining them to obtain a multi-objective optimization function including: Collecting the real-time operation data of server nodes in the computer room and performing standardized processing; Setting the energy consumption index weight, computing power index weight, and service quality index weight according to the historical operation status of each server node in the computer room, calculating the total energy consumption of the server nodes based on the energy consumption index weight to obtain the total energy consumption function of the computer room, calculating the computing power allocation deviation of the tasks among the server nodes based on the computing power index weight to obtain the computing power allocation balance function, and calculating the task processing delay degree of each server node based on the service quality index weight to obtain the service quality guarantee function; Performing normalization processing on the total energy consumption function, the computing power allocation balance function, and the service quality guarantee function of the computer room, dynamically adjusting the combination coefficients of the functions, and combining the three functions into a multi-objective optimization function based on the combination coefficients.

3. The method according to claim 1, characterized in that, Processing the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predicting the energy consumption change curve and computing power load change curve of the server nodes, calculating the dynamic task carrying capacity value and determining the task allocation weight, and generating an initial task allocation plan according to the multi-objective optimization function and the task allocation weight including: Dividing the standardized real-time operation data into time windows, organizing the data in each time window into a feature matrix, where the feature matrix includes the energy consumption data, computing power load data, temperature data, and power consumption data of the server nodes, and calculating the feature time series correlation coefficient matrix based on the feature matrix; Construct a two - layer prediction model based on the time - series prediction unit in the recurrent neural network. The prediction model includes a first - layer time - series prediction unit and a second - layer time - series prediction unit. The first - layer time - series prediction unit processes the feature time - series correlation coefficient matrix to obtain a hidden state. The second - layer time - series prediction unit updates network parameters based on the hidden state. Process the operation data of each server node through the prediction model to obtain the energy consumption change curve and computing power load change curve of the server node in the next time period; Calculate the energy consumption margin ratio sequence based on the energy consumption change curve and the maximum energy consumption threshold, calculate the computing power load margin ratio sequence based on the computing power load change curve and the maximum computing power load threshold. According to the average values of the energy consumption margin ratio sequence and the computing power load margin ratio sequence in the next time period, combine the first weight coefficient and the second weight coefficient to calculate the dynamic task - carrying capacity value of the server node; Calculate the task allocation weight based on the dynamic task - carrying capacity value and temperature parameters, multiply the task allocation weight by the multi - objective optimization function and then find the minimum value to obtain the initial task allocation scheme.

4. The method according to claim 3, wherein Processing the operation data of each server node through the prediction model to obtain the energy consumption change curve and the computing power load change curve of the server node in the next time period includes: Add the operation data corresponding to each server node to the prediction model, perform principal component analysis to obtain a low - dimensional feature representation, perform a non - linear transformation on the low - dimensional feature representation through an activation function to obtain an initial hidden state, substitute the initial hidden state into the neural ordinary differential equation, and solve the neural ordinary differential equation to obtain a continuous - time hidden state function; Calculate the mean and variance of the server node load data to generate a probability distribution function of the server load, sample a random perturbation sequence from the probability distribution function, and superimpose the random perturbation sequence on 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 numerical integration according to the prediction error, perform encrypted sampling on the noisy state sequence at the time points where the prediction error exceeds the preset threshold, and perform numerical integration operations on the sampling points using the fourth - order Runge - Kutta method to obtain a state evolution sequence; Calculate the L2 - norm value of the state evolution sequence and the energy loss value of the server node operation data respectively, add the L2 - norm value and the energy loss value to obtain the total energy value, input the total energy value into the second - layer time - series prediction unit, calculate the gradient of the total energy value with respect to the network parameters, use the backpropagation algorithm to update the network parameters, calculate the causal relationship coefficient between the energy consumption change curve and the computing power load change curve of the server node, and generate a causal dependence relationship graph between nodes; Perform intervention operations on the nodes in the causal dependency graph, calculate the energy change values caused by each intervention operation, compare the energy change values with the law of conservation of thermodynamic energy, correct the prediction data that violates the law of conservation of energy, and output the energy consumption change curve and computing power load change curve of the server node in the next time period.

5. The method according to claim 1, wherein Dividing the parallel ant colony calculation space into multiple sub-regions and setting a temperature decay coefficient to dynamically adjust the pheromone evaporation rate, and controlling the convergence direction of the ant colony search path and marking the optimal path includes: Dividing each parallel ant colony calculation space into multiple sub-regions, calculating the temperature decay coefficient based on the server nodes in the sub-regions, dynamically adjusting the temperature decay coefficient according to time changes, converting the temperature decay coefficient into a pheromone evaporation rate, performing pheromone update in each parallel ant colony calculation space in combination with the task assignment paths in each sub-region, calculating the pheromone retention amount according to the pheromone evaporation rate, accumulating the newly added pheromone increments of all ant individuals on the search path, and adding the pheromone retention amount and the pheromone increments to obtain the updated pheromone concentration; Calculating the transition probabilities of ant individuals choosing different task assignment schemes in each parallel ant colony calculation space respectively, controlling the convergence direction of the ant colony search path based on the transition probabilities, calculating the fitness values of the task assignment schemes in each parallel ant colony calculation space respectively, sorting the task assignment schemes in each parallel ant colony calculation space in descending order according to the fitness values, and marking the task assignment scheme with the highest fitness value as the optimal path in each iteration cycle.

6. The method according to claim 5, characterized in that, Converting the temperature decay coefficient into a pheromone evaporation rate, and performing pheromone update in each parallel ant colony calculation space in combination with the task assignment paths in each sub-region includes: Collecting the temperature decay coefficients corresponding to the server nodes in each sub-region, calculating the task processing ability difference coefficient of the server nodes in the sub-region based on the temperature decay coefficient, and setting an independent pheromone evaporation rate for each sub-region according to the task processing ability difference coefficient; Recording the task assignment paths of ant individuals in each sub-region, calculating the load balancing index and energy consumption efficiency index of the task assignment paths, weighting the load balancing index and the energy consumption efficiency index to obtain a comprehensive path score, mapping the comprehensive path score to a pheromone intensity adjustment coefficient according to a pre-set mapping relation table, and calculating a path pheromone constraint threshold based on the pheromone intensity adjustment coefficient, where the path pheromone constraint threshold is dynamically adjusted with the comprehensive path score; Detecting the pheromone intensity of each task assignment path, and if there is a task assignment path whose pheromone intensity exceeds the pre-set path pheromone constraint threshold, obtaining the adjacent path set of the task assignment path; Calculating the comprehensive path scores of each path in the adjacent path set, calculating the pheromone reallocation weight according to the comprehensive path scores, and allocating the pheromone amount exceeding the path pheromone constraint threshold to the paths in the adjacent path set according to the pheromone reallocation weight to obtain the newly added pheromone increments of the ant individuals on the search path.

7. The method according to claim 1, wherein Calculate the target energy consumption threshold and target computing power load threshold for each server node under the global optimal task allocation scheme, monitor the operating status of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment among the server nodes based on the judgment result to obtain the optimal matching result, including: Obtain the historical task processing records of the server nodes in the global optimal task allocation scheme, extract the energy consumption data and computing power load data from the historical task processing records, and calculate the target energy consumption threshold and target computing power load threshold for each server node; Monitor the real-time energy consumption value and real-time computing power load value of the server nodes, calculate the difference between the real-time energy consumption value and the target energy consumption threshold and the difference between the real-time computing power load value and the target computing power load threshold, and generate a judgment result based on the differences; Determine the source server node and target server node for task migration based on the judgment result, perform the task migration operation and adjust the operating frequency and power supply voltage of each server node until the real-time energy consumption value and real-time computing power load value of the source server node and the target server node converge within a preset interval to obtain the optimal matching result.

8. An energy consumption and computing power balance optimization system for computer rooms based on swarm intelligence, which is used to implement the method described in any one of the foregoing claims 1-7, and is characterized in that, Including: The first unit is used to collect the real-time operation data of the server nodes in the computer room, perform standardization processing on the real-time operation data and construct the total energy consumption function, computing power allocation balance function and service quality guarantee function of the computer room, and combine them to obtain a multi-objective optimization function; The second unit is used to process the standardized real-time operation data according to the time series prediction unit of the recurrent neural network, predict the energy consumption change curve and computing power load change curve of the server nodes, 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 calculation spaces based on the initial task allocation scheme, divide the parallel ant colony calculation spaces into multiple sub-regions and set the temperature attenuation coefficient to dynamically adjust the pheromone evaporation rate, control the convergence direction of the ant colony search path and mark the optimal path, and screen the global optimal task allocation scheme; The fourth unit is used to calculate the target energy consumption threshold and target computing power load threshold for each server node under the global optimal task allocation scheme, monitor the operating status of the server nodes and generate a judgment result, and perform task migration and energy consumption level adjustment among the server nodes based on the judgment result to obtain the optimal matching result.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images