Resource management method and system for green cloud computing and storage medium

Abstract

The invention relates to the technical field of cloud computing resources, and discloses a green cloud computing-oriented resource management method and system and a storage medium. The method comprises the following steps: performing weight optimization on CPU and memory data through a subwolf optimization algorithm to obtain a server energy consumption coefficient matrix; based on the matrix, using DDQN to decide virtual machine CPU allocation, and obtaining a resource allocation parameter table; adjusting the storage position of the data backup copy to obtain a backup node configuration table; monitoring the PUE value of each server to obtain a list of servers with power consumption exceeding the standard; and calculating a migration path of the virtual machine by using a particle swarm algorithm to obtain a load redistribution table. The problems of inaccurate energy consumption prediction, lack of intelligence of scheduling strategies, separation of backup storage and main task scheduling and low power consumption exception processing efficiency in existing cloud computing resource management are solved, and the energy utilization efficiency of a cloud computing environment and the intelligent level of resource scheduling are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing resources, and particularly relates to a resource management method and system for green cloud computing and a storage medium. BACKGROUND

[0002] The existing cloud computing resource management technology mainly adopts a static scheduling strategy based on historical load data, realizes simple allocation of virtual machines through CPU usage threshold and memory occupancy monitoring, and adopts a fixed number of data backup strategies and a time window-based load balancing algorithm for resource scheduling. These technologies can meet the basic resource allocation needs in the traditional cloud computing environment.

[0003] However, the existing technology has the following deficiencies: first, the energy consumption prediction accuracy is insufficient, the existing method only makes scheduling decisions based on a single resource index, and lacks accurate modeling of the comprehensive energy consumption characteristics of CPU and memory resources; second, there is a lack of intelligent dynamic scheduling mechanism, and the traditional rule-based scheduling algorithm cannot adapt to complex and variable cloud computing load patterns; third, the data backup strategy and the main task scheduling are independent of each other, and the selection of backup copy storage location is not optimized in coordination with the overall resource utilization efficiency.

[0004] When the power consumption of the cloud computing environment exceeds the standard, the existing technology lacks a systematic problem positioning and solving mechanism, and cannot accurately identify the specific server that causes the energy consumption anomaly, nor can it develop an optimal virtual machine migration strategy based on the global optimization goal. This limitation leads to low efficiency in handling energy consumption problems, and further affects the green energy-saving effect of the entire cloud computing system. SUMMARY

[0005] The present application provides a resource management method and system for green cloud computing and a storage medium, which are used to solve the problems of inaccurate energy consumption prediction, lack of intelligent scheduling strategy, fragmentation of backup storage and main task scheduling, and low efficiency in handling power consumption anomalies in the existing cloud computing resource management, and improve the energy utilization efficiency of the cloud computing environment and the intelligent level of resource scheduling.

[0006] In a first aspect, the application provides a resource management method for green cloud computing, which comprises: performing weight optimization processing on CPU utilization data and memory occupation data by means of a coyote optimization algorithm to obtain a server energy consumption coefficient matrix; performing DDQN decision processing on a virtual machine CPU allocation ratio according to the server energy consumption coefficient matrix to obtain a virtual machine resource allocation parameter table; performing adjustment processing on a data backup copy storage location by means of the virtual machine resource allocation parameter table to obtain a backup node distribution configuration table; performing PUE monitoring processing on real-time power consumption values of each server according to the backup node distribution configuration table to obtain a power consumption over-standard server list; and performing virtual machine migration path calculation processing on the power consumption over-standard server list by means of a particle swarm optimization algorithm to obtain a server load re-allocation table.

[0007] In a second aspect, the application provides a resource management system for green cloud computing, which comprises:

[0008] a processing module configured to perform weight optimization processing on CPU utilization data and memory occupation data by means of a coyote optimization algorithm to obtain a server energy consumption coefficient matrix;

[0009] a decision module configured to perform DDQN decision processing on a virtual machine CPU allocation ratio according to the server energy consumption coefficient matrix to obtain a virtual machine resource allocation parameter table;

[0010] an adjustment module configured to perform adjustment processing on a data backup copy storage location by means of the virtual machine resource allocation parameter table to obtain a backup node distribution configuration table;

[0011] a monitoring module configured to perform PUE monitoring processing on real-time power consumption values of each server according to the backup node distribution configuration table to obtain a power consumption over-standard server list;

[0012] a calculation module configured to perform virtual machine migration path calculation processing on the power consumption over-standard server list by means of a particle swarm optimization algorithm to obtain a server load re-allocation table.

[0013] In a third aspect, a resource management device for green cloud computing is provided, which comprises: a memory and at least one processor, the memory storing instructions; and the at least one processor invoking the instructions in the memory to cause the resource management device for green cloud computing to perform the resource management method for green cloud computing described above.

[0014] In a fourth aspect, a computer readable storage medium is provided, which stores instructions, and when the instructions are run on a computer, cause the computer to perform the resource management method for green cloud computing described above.

[0015] In the technical solution provided by this application, the technical feature of weighted optimization processing of CPU utilization data and memory occupancy data through the coyote optimization algorithm overcomes the limitations of single resource indicator evaluation in the existing technology. The algorithm simulates the social hierarchy structure of the coyote group and realizes the collaborative optimization of multi-dimensional resource data through the division of labor of Alpha, Beta, and Omega roles, thereby establishing a more accurate server energy consumption coefficient matrix. Based on the technical feature of DDQN decision processing adopted by this energy consumption coefficient matrix, the dual network architecture of the deep double Q network is used to effectively solve the over-estimation problem in traditional reinforcement learning. Through the alternating update mechanism of the target network and the evaluation network, intelligent decision-making of the virtual machine CPU allocation ratio is realized, significantly improving the accuracy and adaptability of resource allocation. The technical feature of adjusting the storage location of the data backup copy by the virtual machine resource allocation parameter table breaks the technical barrier of the traditional backup strategy and the main task scheduling being independent of each other. Through data importance rating and cross-regional distribution optimization, the deep integration of backup storage and resource scheduling is realized, which not only ensures data security but also optimizes the efficiency of storage resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A schematic diagram of an embodiment of a resource management method for green cloud computing in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of a resource management system for green cloud computing in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of a resource management device for green cloud computing in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiment of the present application provides a resource management method and system for green cloud computing and a storage medium. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0021] For ease of understanding, the specific flow of the embodiment of the present application is described below. Please refer to Figure 1 One embodiment of the resource management method for green cloud computing in the embodiment of the present application comprises the following steps.

[0022] In step S101, the CPU utilization data and the memory occupation data are subjected to weight optimization processing through the coyote optimization algorithm, and a server energy consumption coefficient matrix is obtained.

[0023] In step S102, the virtual machine CPU allocation ratio is subjected to DDQN decision processing according to the server energy consumption coefficient matrix, and a virtual machine resource allocation parameter table is obtained.

[0024] In step S103, the virtual machine resource allocation parameter table is subjected to adjustment processing on the data backup copy storage location, and a backup node distribution configuration table is obtained.

[0025] In step S104, the real-time power consumption value of each server is subjected to PUE monitoring processing according to the backup node distribution configuration table, and a power consumption exceeding standard server list is obtained.

[0026] In step S105, the virtual machine migration path of the power consumption exceeding standard server list is calculated through the particle swarm optimization algorithm, and a server load redistribution table is obtained.

[0027] It can be understood that the execution subject of the present application can be a resource management system for green cloud computing, and can also be a terminal or a server, which is not limited here. The embodiment of the present application takes the server as the execution subject for example.

[0028] Specifically, the wolf optimization algorithm collects CPU utilization raw data from each server node through the SNMP protocol or system monitoring agent, which is usually expressed in percentage. The maximum and minimum value scaling process adopts the formula to standardize the raw data, where the original value is divided by the difference between the maximum and minimum values after subtracting the minimum value, resulting in a standardized CPU usage rate vector between 0 and 1. Memory usage data is also obtained through system calls, including used memory, available memory, cache memory, and other indicators. These memory data and the standardized CPU usage rate vector are spliced by column to form a multi-dimensional state matrix containing CPU and memory dual resource indicators. The wolf optimization algorithm simulates the social hierarchy of wolf populations in nature, with Alpha wolves as group leaders responsible for exploring the global optimal solution, and the fitness function value is used to evaluate the quality of each solution. Beta wolves serve as deputy leaders to assist Alpha in local search and fine-tuning optimization based on the search direction determined by Alpha wolves. Omega wolves are ordinary members responsible for random search to prevent the algorithm from falling into local optimal solutions. The fitness function aims to minimize energy consumption, considering CPU power consumption, memory power consumption, and heat dissipation power consumption. The algorithm initializes the population generation process for the multi-dimensional state matrix, randomly generates several weight solutions as the initial population, and each weight solution represents the energy consumption weight coefficient of CPU and memory resources. Through the iterative optimization process, Alpha, Beta, and Omega wolves search collaboratively, constantly updating individual solutions in the population until the algorithm converges to the optimal weight solution. Finally, the optimal weight solution is linearly combined with the CPU base power and memory power coefficients in the server hardware specification parameters to generate a server energy consumption coefficient matrix reflecting the energy consumption characteristics of each server. The rows of the matrix represent different servers, and the columns represent the energy consumption coefficients of different resource types.

[0029] DDQN algorithm is the abbreviation of deep double Q-network, which effectively solves the overestimation bias problem in traditional Q-learning algorithm through double network architecture. The algorithm first transforms the server energy consumption coefficient matrix in dimension, and flattens the two-dimensional matrix into a one-dimensional state space vector, which is processed into a multi-dimensional state input format acceptable by the neural network through linear mapping or nonlinear mapping. The double network architecture includes two independent deep neural networks, target network and evaluation network. The target network is responsible for calculating the target Q value as the learning target, and the evaluation network is responsible for selecting the optimal action according to the current state. The two networks have the same structure but the parameters are updated independently. The evaluation network parameters are updated in real time, and the target network parameters are copied from the evaluation network periodically, which reduces the instability in the learning process. The ε-greedy strategy balances between exploring unknown states and using known optimal strategies. When the generated random number is less than the ε value, a random action is selected to explore the environment, and when the random number is greater than or equal to the ε value, the action with the maximum Q value is selected for use. The experience replay mechanism establishes a fixed capacity buffer to store state transition samples, each sample contains four elements: current state, executed action, obtained reward and next state. Through the random sampling mechanism, samples are extracted from the buffer for network training, breaking the time correlation between samples. The time difference learning calculates the time difference error between the current Q value and the target Q value, and updates the network weight parameters through the gradient descent algorithm, gradually optimizing the decision-making ability of the network. The algorithm finally outputs the CPU allocation ratio and memory allocation amount of virtual machines on each server, forming a detailed virtual machine resource allocation parameter table.

[0030] The data importance rating process classifies and evaluates data according to multiple dimensions, including business criticality, data access frequency, user sensitivity, legal compliance requirements, and other factors. Critical-level data covers core business data, user account information, transaction records, and other high-value data that requires extremely high availability and security, so five backup copies are allocated to ensure redundancy protection. General-level data includes system log files, temporary calculation results, cache data, and other relatively low-value data, and three backup copies are allocated to balance reliability and storage costs. The load balancing calculation process monitors the storage capacity usage, network bandwidth occupancy, and disk I / O load of each data center in real time, calculates the comprehensive load score of each data center, and selects the data center nodes with the lowest load score as the storage node candidate list. The cross-regional distribution optimization process calculates the geographical distance between any two data centers based on their geographical coordinates using the spherical distance calculation formula, and constructs a geographical distance matrix. The minimum spanning tree algorithm analyzes the connectivity of the distance matrix and generates a minimum connection topology graph between data centers to ensure connectivity at the minimum connection cost. The greedy algorithm traverses the storage node candidate list and selects the node combination with the minimum total transmission distance as the optimal node combination solution under the premise of meeting the cross-regional distribution requirements. Finally, the nodes are sorted according to the transmission priority to generate a complete backup storage path solution that includes the primary backup path and the secondary backup path.

[0031] The PUE value is a core indicator of data center energy use efficiency, and its calculation method is the total power consumption of the data center divided by the IT equipment power consumption. An ideal PUE value close to 1.0 indicates that almost all electrical energy is used for IT equipment operation. The power consumption sensor collects real-time three-phase current and voltage data of each server through current transformers and voltage sensors, and the sampling frequency is usually set to multiple times per second to obtain accurate power consumption measurement values. The server power consumption raw data is calculated by multiplying the voltage and current and then multiplying by the power factor, taking into account the phase difference factor in an AC circuit. The total power consumption of the data center includes multiple components, in addition to the power consumption of IT equipment such as servers, including air conditioning and refrigeration system power consumption, lighting system power consumption, UPS uninterruptible power supply loss, power distribution system loss, network equipment power consumption, and other auxiliary facility power consumption. The real-time PUE monitoring value is obtained by dividing the total power consumption of the data center by the IT equipment power consumption. When this value exceeds the pre-set PUE threshold, the anomaly detection algorithm automatically marks the corresponding server as a power consumption exceeding standard. The power consumption deviation degree is calculated by dividing the difference between the current power consumption value and the standard power consumption value by the standard power consumption value to obtain the deviation percentage. The system is arranged in descending order according to the deviation percentage, generating a power consumption exceeding standard server list sorted by power consumption exceeding standard severity, and the server with the largest deviation degree is processed first.

[0032] The particle swarm optimization algorithm abstracts the virtual machine migration problem as a particle search process in a multidimensional optimization space, where each particle represents a complete virtual machine migration solution. The particle position vector uses integer encoding, with each element corresponding to a virtual machine and the value representing the target server number for that virtual machine. This encoding establishes a mapping between source and target servers. When initializing the particle swarm, the algorithm assigns each particle an initial position and velocity using a random number generator, ensuring a diverse initial distribution of particles in the solution space. The multi-objective fitness function comprehensively considers multiple optimization objectives, including the degree of total energy consumption reduction after migration, the network transmission cost of virtual machine migration, the business impact of service interruption time, and the resource reconfiguration overhead of the target server. These objectives are then combined into a single fitness value through a weighted summation. The particle velocity update mechanism combines information about individual and global historical optimal positions and controls the particle's search behavior using three parameters: inertia weight, individual learning factor, and global learning factor. The inertia weight controls the particle's tendency to maintain its current direction of motion, the individual learning factor controls the particle's tendency to move toward its individual optimal position, and the global learning factor controls the particle's tendency to move toward the global optimal position. Particle positions are updated by adding the velocity increment to the current position to obtain the new position at the next moment. Boundary constraints are checked after the position update to ensure that the particle position is within the valid solution space. During the iterative optimization process, the algorithm continuously evaluates the fitness value of each particle, updating both the individual optimal position and the global optimal position. After a preset number of iterations, the algorithm converges and outputs the global optimal particle position as the optimal migration strategy. The resulting server load redistribution table contains detailed information for each virtual machine, including the source and target server IDs, migration start time, estimated migration completion time, and post-migration resource allocation.

[0033] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0034] Collect the raw CPU utilization data of each server node and obtain the standardized CPU utilization vector after maximum and minimum value scaling.

[0035] The memory usage data and the normalized CPU usage vector are concatenated column by column to obtain a multi-dimensional state matrix containing dual resource indicators;

[0036] A fitness function with the goal of minimizing energy consumption is constructed, and the coyote algorithm is used to initialize the population generation process on the multi-dimensional state matrix to obtain the initial weight solution set;

[0037] Simulating the social hierarchy of a coyote group, the team used the Alpha, Beta, and Omega roles to perform collaborative search and optimization on the initial weight solution set, ultimately obtaining the optimal weight solution after convergence.

[0038] Based on the server CPU basic power consumption and memory power consumption coefficient, the optimal weight solution is linearly combined to obtain the server energy consumption coefficient matrix reflecting the energy consumption characteristics of each server.

[0039] Specifically, each server node collects CPU utilization raw data through the proc file system or performance counter interface provided by the operating system kernel. These data represent the current CPU workload state in percentage form. The maximum and minimum value scaling process converts the raw percentage data into standardized values using a linear transformation method. The specific calculation process is to subtract the minimum CPU utilization value in all servers from each raw CPU utilization value, and then divide by the difference between the maximum CPU utilization value and the minimum CPU utilization value. Through this normalization process, a standardized CPU usage vector with a value range of 0 to 1 is obtained. This standardization process eliminates the influence of different server hardware performance differences on data analysis, ensuring that the data processed by subsequent algorithms have a uniform numerical range and comparability. Each element of the standardized CPU usage vector corresponds to the standardized CPU utilization value of a specific server, and the vector length is equal to the total number of servers in the cloud computing environment. Memory occupation data collection includes used memory amount, total memory capacity, memory usage rate and other indicators of each server. These data are obtained through system calls or memory management units. The memory occupation data and the standardized CPU usage vector are spliced by column to form a multi-dimensional state matrix containing double resource indicators. The specific process is to take the standardized CPU usage vector as the first column, the memory usage rate data as the second column, the used memory amount data as the third column, and the total memory capacity data as the fourth column. Through column vector splicing, a multi-dimensional state matrix containing double resource indicators is formed. The number of rows of this multi-dimensional state matrix is equal to the number of servers, and the number of columns is equal to the number of monitored indicators. Each element in the matrix represents the state value of a specific server in a specific resource dimension.

[0040] The multi-dimensional state matrix establishes the correlation between CPU resources and memory resources, and the fitness function of the energy minimization objective comprehensively considers multiple energy components such as CPU energy consumption, memory energy consumption, and heat dissipation energy consumption. The fitness function value is calculated by merging various energy types through weighted summation. The wolf algorithm initializes the population generation process by assigning a corresponding weight coefficient to each resource dimension in the multi-dimensional state matrix. Each weight coefficient represents the contribution of that resource dimension to the overall energy consumption. The initial weight solution set is generated by using a random number generator to randomly assign initial weight values to each resource dimension within a predetermined weight range. Each weight solution contains two main components: CPU weight coefficient and memory weight coefficient. The population size determines the number of weight solutions that evolve simultaneously. A larger population size increases the diversity of solutions but increases the computational complexity, while a smaller population size reduces the computational overhead but reduces the search ability. Each individual solution in the initial weight solution set represents a resource weight distribution scheme, and the energy optimization effect of each scheme is evaluated by the fitness function. Alpha wolf, as the leader of the group, is responsible for global search and identifies the current optimal solution by evaluating the fitness values of all individuals in the initial weight solution set. The search behavior of Alpha wolf simulates the decision-making process of leaders in nature and searches for the weight combination with the minimum energy consumption in the solution space.

[0041] Beta wolf, as a secondary leader, assists Alpha wolf in local search and performs fine optimization based on the search direction determined by Alpha wolf. The search range of Beta wolf is relatively small but has high search accuracy. Omega wolf, as a general group member, is responsible for random search and explores other potential optimization directions by random walking in the solution space, preventing the algorithm from falling into a local optimal solution. The cooperative search optimization process of the three types of wolves continuously improves the quality of weight solutions through iterative updating mechanism. In each iteration process, Alpha, Beta, and Omega wolves update the weight solutions according to their respective search strategies, then re-evaluate the fitness values and update the social hierarchy structure. The convergence judgment conditions include reaching the preset number of iterations or the improvement amplitude of fitness value being less than the preset threshold. When the convergence conditions are met, the algorithm stops iteration and outputs the optimal weight solution. The server CPU basic power consumption data is derived from the hardware specification or actual power consumption test, reflecting the power consumption characteristics of the server under different CPU load levels. The memory power consumption coefficient is determined according to the memory type, capacity, and working frequency, and different types of memory modules have different power consumption characteristics. The linear combination operation processes the CPU weight coefficient in the optimal weight solution by multiplying it with the server CPU basic power consumption, and the memory weight coefficient by multiplying it with the memory power consumption coefficient. Then, the two product results are added to obtain the comprehensive energy consumption coefficient of each server.

[0042] The rows of the server energy consumption coefficient matrix represent different server nodes, and the columns represent different energy consumption types. Each element in the matrix represents the energy consumption coefficient value of a specific server under a specific energy consumption type. This energy consumption coefficient matrix establishes a quantitative relationship between server hardware characteristics and energy consumption performance. Taking a private cloud environment of an enterprise as an example, the environment includes 500 servers with different configurations, of which 200 are high-performance computing servers equipped with 32-core CPUs and 64GB of memory, 200 are storage servers equipped with 16-core CPUs and 32GB of memory, and 100 are general-purpose servers equipped with 8-core CPUs and 16GB of memory. During the CPU utilization collection process, it was found that the CPU utilization of high-performance computing servers was generally high and fluctuated greatly, the CPU utilization of storage servers was relatively stable but the memory usage was high, and the resource utilization of general-purpose servers was relatively balanced. The maximum and minimum value scaling process standardizes the CPU utilization of high-performance computing servers from the original data, and the CPU utilization of storage servers and general-purpose servers is also standardized. Through standardization, the performance differences between different server types are eliminated.

[0043] During the memory occupation data splicing process, the standardized CPU utilization, memory utilization, used memory absolute value, and total memory capacity of each server are combined by column to form a 500-row and 4-column multi-dimensional state matrix. This matrix completely describes the resource usage state of the entire cloud environment. The wolf algorithm initializes 50 weight solutions as the initial population, each weight solution contains two parameters of CPU weight and memory weight. The Alpha wolf identifies the weight combination with the lowest overall energy consumption through fitness evaluation. The Beta wolf performs local optimization on the basis of this to fine-tune the weight parameters. The Omega wolf explores other weight combinations through random search. After multiple iterations, the algorithm converges. The optimal weight solution is multiplied by the hardware power consumption parameters of each server through linear combination operation. The CPU basic power consumption of high-performance computing servers is multiplied by the CPU weight coefficient to obtain the CPU energy consumption contribution value. The memory power consumption is multiplied by the memory weight coefficient to obtain the memory energy consumption contribution value. The sum of the two values is the comprehensive energy consumption coefficient of the server. The energy consumption coefficients of storage servers and general-purpose servers are calculated in the same way. Finally, a complete energy consumption coefficient matrix reflecting the energy consumption characteristics of all servers is formed.

[0044] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0045] Convert the server energy consumption coefficient matrix to a state space vector, and obtain a multi-dimensional state input after dimension mapping processing;

[0046] Build a double-network architecture including a target network and an evaluation network, perform forward propagation processing on the multi-dimensional state input, and obtain a Q-value action matrix;

[0047] The CPU allocation action is selected from the Q-value action matrix based on an epsilon-greedy strategy, sample storage processing is performed through an experience replay mechanism, and a training buffer is obtained;

[0048] The training buffer is subjected to parameter update processing through temporal difference learning, and a DDQN decision model is obtained;

[0049] The CPU allocation proportion and the memory allocation amount output by the DDQN decision model are subjected to parameter combination processing according to the virtual machine load demand, and a virtual machine resource allocation parameter table is obtained.

[0050] Specifically, the server energy consumption coefficient matrix is converted into a state space vector through a matrix flattening operation, which rearranges the two-dimensional energy consumption coefficient matrix into a one-dimensional vector form in row or column order, and each energy consumption coefficient value in the matrix occupies a fixed position index in the state space vector. Dimension mapping processing converts the state space vector into a format acceptable to the neural network through linear transformation or nonlinear transformation methods. Linear transformation adjusts the vector dimension through weight matrix multiplication, and nonlinear transformation introduces nonlinear characteristics through an activation function to enhance the expression ability of the network. The multi-dimensional state input contains the energy consumption coefficient information of each server and the current virtual machine distribution state, forming a comprehensive input vector that describes the resource state of the entire cloud environment. The double-network architecture of the DDQN algorithm includes two deep neural networks with the same structure but independent parameters, namely the target network and the evaluation network. The target network is responsible for calculating the target Q-value as the reference standard for learning, and the evaluation network is responsible for selecting the optimal action according to the current state. During the forward propagation process, the multi-dimensional state input first enters the network through the input layer of the evaluation network, and then passes through multiple hidden layer neuron calculations and activation function processing, finally producing Q-values corresponding to different CPU allocation actions in the output layer. The rows of the Q-value action matrix represent different virtual machines, and the columns represent different CPU allocation proportion options. Each element in the matrix represents the expected reward value of executing a specific CPU allocation action under the current state. The parameter update strategies of the target network and the evaluation network are different. The parameters of the evaluation network are updated in real time to adapt to environmental changes, and the parameters of the target network are periodically copied from the evaluation network to maintain the stability of the learning process.

[0051] The e-greedy strategy balances exploration of unknown actions and exploitation of known optimal actions. When the generated random number is less than the preset e value, the algorithm selects a random CPU allocation action for environment exploration. When the random number is greater than or equal to the e value, the algorithm selects the CPU allocation action with the maximum Q value in the Q value action matrix for exploitation. The CPU allocation action includes allocating a virtual machine to different servers, adjusting the CPU usage ratio of the virtual machine, modifying the priority of the virtual machine, and the like. The experience replay mechanism establishes a fixed-capacity circular buffer to store state transition experiences. Each experience includes four key elements: the current state, the executed action, the obtained reward, and the next state. When the buffer capacity reaches the upper limit, the newest experience will overwrite the oldest experience. The sample storage process writes the current state transition experience into the training buffer in a fixed format, while maintaining the timestamp and priority information of the experience to support subsequent sampling operations.

[0052] The temporal difference learning calculates the temporal difference error between the current Q value and the target Q value using the Bellman equation. The target Q value is calculated by the target network, and the current Q value is calculated by the evaluation network. The parameter update process uses the gradient descent algorithm to minimize the square of the temporal difference error. The gradient of the network parameters is calculated by the backpropagation algorithm, and then the optimizer is used to update the weight and bias parameters of the evaluation network. The experience samples in the training buffer are selected for network training by the random sampling mechanism. The batch training method processes multiple experience samples simultaneously to improve training efficiency and stability. The DDQN decision model gradually learns the optimal CPU allocation strategy during the training process. After the model parameters converge, the model can output the optimal virtual machine resource allocation scheme based on the current server energy consumption state.

[0053] The virtual machine load demand includes CPU computing demand, memory usage demand, network bandwidth demand, storage I / O demand, and other resource requirements in multiple dimensions. These demand information is derived from the historical running data and current task load of the virtual machine. The parameter combination process matches the CPU allocation ratio output by the DDQN decision model with the CPU demand of the virtual machine, ensuring that the allocated CPU resources meet the computing requirements of the virtual machine. At the same time, the memory allocation is compared with the memory demand of the virtual machine to adjust the memory allocation to meet the memory usage requirements of the virtual machine. Each row of the virtual machine resource allocation parameter table corresponds to a virtual machine, and each column corresponds to a resource type. The values in the table represent the specific resource quantities allocated to a specific virtual machine, including CPU core number, CPU usage ratio, memory capacity, storage space, and other detailed parameters.

[0054] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0055] The virtual machine identifier and CPU allocation ratio are extracted from the virtual machine resource allocation parameter table, and a data priority classification matrix is obtained through data importance rating processing;

[0056] The importance level of each data block in the data priority classification matrix is determined, 5 backup copies are allocated when the importance level is critical, and 3 backup copies are allocated when the importance level is general, to obtain a copy number allocation table;

[0057] Based on the copy number allocation table, load balancing calculation and processing are performed on the storage capacity of each data center to obtain a storage node candidate list;

[0058] According to the data center geographical location distance matrix, cross-regional distribution optimization processing is performed on the storage node candidate list to obtain a backup storage path scheme;

[0059] The backup storage path scheme is associated and mapped with the virtual machine resource allocation parameter to obtain a backup node distribution configuration table containing specific storage node addresses.

[0060] Specifically, the virtual machine identifier in the virtual machine resource allocation parameter table identifies each virtual machine using a unique coding method. The identifier contains key information such as virtual machine type, creation time, and business affiliation. The CPU allocation ratio extraction process obtains the CPU resource percentage occupied by each virtual machine by analyzing the CPU-related fields in the parameter table. These values reflect the computational intensity and business importance of the virtual machine. The data importance rating processing classifies the data carried by the virtual machine based on multiple evaluation dimensions, including business criticality, data access frequency, user size, legal compliance requirements, and data recovery time targets. The rating algorithm assigns appropriate weight coefficients to each evaluation dimension, calculates the comprehensive importance score of each virtual machine data through weighted summation, and sets the score range to a numerical interval of 0 to 100. The rows of the data priority classification matrix correspond to different virtual machines, and the columns correspond to different evaluation dimensions. The matrix element value represents the score of a specific virtual machine in a specific evaluation dimension. Through matrix operations, the final importance level is obtained.

[0061] The importance level judgment of each data block in the data priority classification matrix is realized by threshold comparison method, and the algorithm sets multiple importance level thresholds to divide the data into different categories. The key level data usually refers to the virtual machine data with a comprehensive importance score of more than 80 points, including core business systems, user account information, transaction records, compliance audit data and other high-value information. The loss or damage of this kind of data will have a serious impact on the business. The general level data refers to the virtual machine data with a comprehensive importance score between 40 and 80 points, including business logs, cache data, report files, temporary calculation results and other medium-value information. The loss of this kind of data will affect the business continuity but will not cause fatal loss. The low level data refers to the virtual machine data with a comprehensive importance score of less than 40 points, including test data, expired files, temporary storage and other low-value information. The replica number allocation rule determines the backup strategy according to the importance level of the data. The key level data is allocated 5 backup replicas to ensure high data security, the general level data is allocated 3 backup replicas to balance security and storage cost, and the low level data is allocated 1 backup replica or not backed up. The replica number allocation table records the replica number requirements corresponding to each virtual machine, including virtual machine identifier, importance level, replica number, storage capacity requirement and other detailed information.

[0062] The load balancing calculation process optimizes the allocation based on the storage requirements in the replica number allocation table and the current storage status of each data center. The storage capacity monitoring includes used storage space, available storage space, disk I / O load, network bandwidth occupation and other indicators, which are obtained in real time through the monitoring interface of the storage management system. The load balancing algorithm calculates the comprehensive load score of each data center, and the score formula considers factors such as storage capacity utilization, I / O load level, network transmission delay, etc. In the process of generating the storage node candidate list, the algorithm preferentially selects data centers with lower load scores as backup storage targets, and also considers whether the storage capacity is sufficient to accommodate the required backup data. Each storage node in the candidate list contains data center identifier, available storage capacity, current load level, geographic location information and other attributes.

[0063] The data center geographical position distance matrix records the geographical distance information between any two data centers, and the distance calculation adopts a spherical distance formula to determine the actual geographical distance based on latitude and longitude coordinates. The cross-regional distribution optimization processing ensures that backup copies are distributed in different geographical regions to reduce the impact of regional disasters on data security. The optimization algorithm sets a minimum geographical distance threshold to require the distance between backup copies to exceed a preset value. During the backup storage path scheme generation process, the algorithm selects a node combination that meets the geographical distribution requirements from the storage node candidate list. The optimization objectives include minimizing the total transmission distance, balancing the storage load of each data center, meeting data sovereignty and compliance requirements, and multiple constraint conditions. The path optimization algorithm uses a greedy strategy or dynamic programming method to find the optimal storage node combination under the premise of meeting the constraint conditions. The generated backup storage path scheme includes complete planning of the primary backup path and the secondary backup path.

[0064] The association mapping processing establishes a corresponding relationship between the abstract storage nodes in the backup storage path scheme and the specific virtual machines in the virtual machine resource allocation parameter table. The mapping process determines the most suitable storage node based on factors such as the data importance level of the virtual machine, storage capacity requirements, geographical location preferences, and the current running physical server location of the virtual machine to optimize data transmission efficiency. The generation of the backup node distribution configuration table combines virtual machine resource allocation information and backup storage path schemes to specify specific backup storage node addresses for each virtual machine, including primary data center addresses, backup data center addresses, storage device identifiers, access paths, and other detailed information. The configuration table also contains backup scheduling strategies, data synchronization frequencies, fault switching schemes, and other operation and maintenance management information to ensure the efficient operation of the backup system.

[0065] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0066] Perform spherical distance calculation processing on the latitude and longitude coordinates of each data center to obtain distance measurement values between data centers;

[0067] Construct the distance measurement values into a symmetric matrix form and perform connectivity analysis processing through a minimum spanning tree algorithm to obtain a regional connection topology graph;

[0068] Set a preset distance threshold based on the regional connection topology graph, and mark two data centers as a cross-regional node pair when the distance between them exceeds the threshold to obtain a cross-regional node grouping;

[0069] Perform greedy algorithm traversal processing on the storage node candidate list based on the cross-regional node grouping to select a node combination that minimizes the total transmission distance to obtain an optimal node combination scheme;

[0070] The nodes in the optimal node combination scheme are sorted according to the transmission priority, and a backup storage path scheme containing a primary backup path and a secondary backup path is generated.

[0071] Specifically, the latitude and longitude coordinates of each data center are obtained through a GPS positioning system or a geographic information database, and the coordinate information includes two key parameters, i.e., a latitude value and a longitude value, which represent the geographic position in the form of a decimal degree. The spherical distance calculation process adopts the great circle distance algorithm to consider the influence of the curvature of the earth. The calculation process first converts the latitude and longitude coordinates from degrees to radians, and then calculates the shortest distance between any two points by applying the principles of spherical trigonometry. The distance calculation formula considers parameters such as the radius of the earth, the latitude difference between two points, the longitude difference, and the latitude cosine value, and obtains the accurate geographic distance through the inverse cosine function and square root operation. The distance measurement value between data centers is expressed in kilometers, and the calculation result reflects the actual geographic distance between data centers, providing a quantitative basis for subsequent network transmission delay estimation and regional division.

[0072] The symmetric matrix constructed by the distance measurement value adopts a square matrix form, and the number of rows and columns of the matrix is equal to the total number of data centers. The element in the ith row and jth column of the matrix represents the distance from the ith data center to the jth data center. The symmetric matrix has the property of symmetry, i.e., the distance from A data center to B data center is equal to the distance from B data center to A data center, and the diagonal elements of the matrix are all zero, indicating that the distance from a data center to itself is zero. The minimum spanning tree algorithm processes the distance matrix using Kruskal's algorithm or Prim's algorithm. The algorithm aims to find a tree-like connection structure that connects all data centers with the minimum total distance. The connectivity analysis process ensures that all data centers maintain network connectivity at the minimum connection cost through the minimum spanning tree algorithm. The generated regional connection topology graph describes the optimal connection relationship between data centers in the form of graph theory, where the nodes represent data centers, the edges represent connection relationships, and the weights of the edges represent connection distances.

[0073] The preset distance threshold in the regional connection topology graph is determined according to business requirements and geographical distribution characteristics. The threshold setting considers multiple factors such as data transmission delay, network bandwidth cost, disaster recovery requirements, etc. The cross-regional node pair marking process traverses all edges in the topology graph. When the weight of an edge, i.e., the distance between two data centers, exceeds the preset threshold, the algorithm marks the pair of data centers as a cross-regional node pair, indicating that they are located in different geographical regions. The cross-regional node grouping uses the connected component algorithm in graph theory to group data centers with distances less than the threshold into the same regional group and data centers with distances greater than the threshold into different regional groups. The grouping result generates multiple data center sets, each representing a geographical region. Data centers within a set can be connected through high-speed networks, and data centers between sets need to be connected through long-distance networks.

[0074] The greedy algorithm iteratively selects one or more storage nodes from each regional group as backup targets based on cross-regional node grouping and storage node candidate list. The selection criteria include sufficient storage capacity, low current load, good network connection quality, etc. The greedy strategy makes a local optimal decision at each selection, prioritizing the storage node with the shortest transmission distance that meets the constraints, and gradually builds a complete node combination through the iterative selection process. The total transmission distance calculation includes the sum of distances from the main data center to each backup data center and the transmission distance for data synchronization between backup data centers. The algorithm finds the node combination that minimizes the sum of these distances. The optimal node combination scheme includes the complete configuration of the main storage node and multiple backup storage nodes, with each node having a clear role definition and storage responsibility division.

[0075] The transmission priority ranking process evaluates the priority of each node in the optimal node combination scheme based on multiple evaluation indicators, including geographical distance, network bandwidth, storage performance, historical reliability, maintenance cost, etc. The main backup path points to the backup storage node with the highest transmission priority, which usually has the best network connection quality and the highest reliability guarantee, serving as the main data backup target. The secondary backup path points to multiple backup storage nodes with the next highest transmission priority, which serve as backup data backup targets and provide redundancy protection when the main backup path fails. The backup storage path scheme organizes each backup path in a hierarchical structure, including path priority, transmission protocol, bandwidth allocation, fault switching strategy, etc., ensuring efficient execution of data backup operations and fast recovery in fault conditions.

[0076] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0077] The server node identifier and the storage load distribution are extracted from the backup node distribution configuration table, real-time current and voltage data of each server are collected through a power consumption sensor, and a server power consumption original data set is obtained;

[0078] The server power consumption original data set, the cooling system power consumption, and the lighting system power consumption are accumulated and calculated to obtain a total power consumption value of the data center;

[0079] The total power consumption value of the data center and the IT equipment power consumption are subjected to ratio operation processing to obtain a real-time PUE monitoring value;

[0080] It is judged whether the real-time PUE monitoring value exceeds a preset PUE threshold value, and when the threshold value is exceeded, the corresponding server is marked as an over-standard state, and an abnormal power consumption server set is obtained through an abnormality detection algorithm for screening processing;

[0081] The servers in the abnormal power consumption server set are arranged in descending order according to the power consumption deviation degree to obtain a power consumption over-standard server list sorted according to the severity of power consumption over-standard.

[0082] Specifically, the server node identifier in the backup node distribution configuration table adopts a hierarchical coding structure to contain data center number, rack number, server position number and other hierarchical information, and each identifier uniquely corresponds to one physical server. The storage load distribution information records the current backup data amount, storage device utilization rate, disk I / O load and other storage-related indicators of each server, and these data are updated regularly through the monitoring interface of the storage management system. The power consumption sensor includes current transformers and voltage sensors. The current transformer measures the alternating current flowing through the server power line through electromagnetic induction principle, and the voltage sensor measures the voltage amplitude of the server power supply line through a voltage dividing circuit. During the real-time current and voltage data collection process, the sensor samples at a high frequency, usually multiple times per second to capture the instantaneous changes in power consumption. The analog signals collected are converted into digital signals by an analog-to-digital converter for subsequent processing. The server power consumption original data set converts the current and voltage data into power consumption values through an electric power calculation formula, and the calculation process considers the power factor correction in the alternating current circuit to obtain the instantaneous power consumption and average power consumption data of each server.

[0083] The cooling system power consumption includes the total power consumption of air conditioning equipment, fans, cooling towers, water pumps and other refrigeration equipment. The power consumption of these devices is collected in real time by dedicated energy consumption monitoring devices. The cooling system power consumption is closely related to the thermal load of the data center. An increase in server power consumption leads to an increase in cooling demand, and the cooling system power consumption increases accordingly to maintain the appropriate operating temperature. The lighting system power consumption includes the power consumption of all lighting devices in the machine room, such as LED lamps, emergency lighting, safety lighting, etc. The lighting power consumption is relatively stable but adjusted at different times. The cumulative calculation process sums the power consumption values of all servers in the server power consumption raw data set to obtain the total IT equipment power consumption. Then, the total IT equipment power consumption is added to the cooling system power consumption, lighting system power consumption, UPS loss, power distribution loss and other auxiliary facility power consumption to obtain the total power consumption value reflecting the energy consumption status of the entire data center. The total power consumption value of the data center is expressed in kilowatts. This value fluctuates dynamically with changes in business load and is an important indicator for evaluating the energy efficiency of the data center.

[0084] The PUE value is the abbreviation of Power Usage Effectiveness, which represents the energy usage efficiency of the data center. An ideal PUE value close to 1.0 indicates that almost all electrical energy is used for IT equipment operation. The ratio operation process divides the total power consumption value of the data center by the IT equipment power consumption to obtain the real-time PUE monitoring value. The calculation formula is PUE equals the total power consumption of the data center divided by the power consumption of the IT equipment. The real-time PUE monitoring value usually varies between 1.2 and 2.5. The closer the value is to 1.0, the higher the energy utilization efficiency. A larger value indicates that auxiliary facilities consume too much electrical energy. Real-time calculation of the PUE monitoring value provides immediate feedback on energy efficiency status for data center operation and maintenance personnel, helping to identify energy consumption abnormalities and optimization opportunities.

[0085] The preset PUE threshold is determined according to the design standards and operation targets of the data center. It is usually set to a value between 1.5 and 2.0. If the threshold is exceeded, it indicates that the energy efficiency is low and optimization measures need to be taken. The judgment process checks whether the real-time PUE monitoring value exceeds the preset threshold through a comparison operation. When the monitoring value is greater than the threshold, the algorithm triggers an abnormal alarm process. The server over-standard state marking process identifies the specific servers that cause the PUE value to exceed the standard. The responsibility for exceeding the standard is determined by analyzing the power consumption contribution of each server. The abnormality detection algorithm uses statistical methods to identify servers with abnormal power consumption, including the 3-sigma rule based on mean standard deviation, time series analysis based on historical data, and abnormal pattern recognition based on machine learning. The screening process filters out devices with abnormal power consumption from all servers. The set of servers with abnormal power consumption includes all server nodes with power consumption significantly deviating from the normal range, providing target objects for subsequent load adjustment.

[0086] The power consumption deviation degree calculation obtains a deviation value by comparing the current power consumption of each server with its historical average power consumption or theoretical standard power consumption, and the deviation degree is expressed in percentage form as the degree of power consumption exceeding the normal range. The descending order arrangement process sorts all servers in the abnormal power consumption server set in descending order of power consumption deviation degree, with the server having the largest deviation degree ranked first in the list and the server having a smaller deviation degree ranked at the end of the list. The power consumption exceeding standard server list provides the priority processing order for the operation and maintenance personnel, and the priority processing of the server with the largest deviation degree can most effectively reduce the overall PUE value. The list contains detailed information such as server identification, current power consumption value, standard power consumption value, deviation degree, and over-standard duration.

[0087] In a specific embodiment, the process of performing step S106 can specifically include the following steps:

[0088] Extract the virtual machine distribution information of the over-standard server from the power consumption over-standard server list, and quantitatively process the resource occupation of each virtual machine by the load calculation algorithm to obtain a virtual machine migration candidate set;

[0089] Initialize the particle swarm population, encode each migration scheme in the virtual machine migration candidate set as a particle position vector, and obtain the initial particle swarm population through random generation processing;

[0090] Construct a multi-objective fitness function targeting energy consumption reduction and migration cost, and perform fitness evaluation processing on each particle in the initial particle swarm population to obtain a particle fitness score matrix;

[0091] Based on the particle velocity update and position update mechanism, perform iterative optimization processing on the particle fitness score matrix, select the global optimal particle position as the optimal migration strategy, and obtain a virtual machine migration path configuration;

[0092] Map and associate the source server and the target server in the virtual machine migration path configuration, and generate a server load reallocation table containing a migration time sequence and resource reallocation.

[0093] Specifically, the virtual machine distribution information in the power consumption over-standard server list contains the number of virtual machines running on each over-standard server, virtual machine identifier, virtual machine type, creation time, and other basic information. The virtual machine distribution information extraction process queries the virtual machine list of each over-standard server through the API interface of the virtualization management platform to obtain the mapping relationship between the virtual machine and the physical server. The load calculation algorithm monitors and statistically analyzes the CPU usage, memory occupancy, disk I / O rate, and network bandwidth consumption of the virtual machine in real time. Through the calculation of average value, peak value, fluctuation amplitude, and other statistical indicators, the resource consumption characteristics of each virtual machine are quantified. The resource occupancy quantification process uses a weighted scoring method to unify different types of resource consumption into comparable numerical values. The CPU occupancy is multiplied by the CPU weight coefficient, the memory occupancy is multiplied by the memory weight coefficient, the storage and network occupancy are multiplied by the corresponding weight coefficients, and finally the comprehensive resource occupancy score of each virtual machine is obtained. The virtual machine migration candidate set contains all virtual machines running on the power consumption over-standard server and suitable for migration. The set excludes virtual machines that are executing critical tasks, have hardware dependencies, or are configured with migration restrictions to ensure the feasibility and safety of the migration operation.

[0094] The population initialization process of the particle swarm optimization algorithm sets the population size parameter to determine the number of particles evolving simultaneously, usually set to 30 to 100 particles to balance the search ability and computational complexity. The particle position vector encoding uses integer encoding to represent the virtual machine migration scheme. The vector length is equal to the number of virtual machines in the virtual machine migration candidate set, and each element in the vector represents the target server number of the corresponding virtual machine. Random generation assigns each particle an initial position and an initial velocity. The initial position randomly selects a target server for each virtual machine within the valid server number range using a random number generator, and each element of the initial velocity vector is set to a small random value to control the initial movement amplitude of the particle. The initial particle population forms a solution set containing multiple migration schemes. Each particle represents a complete virtual machine allocation strategy, and the diversity of the population ensures that the algorithm can explore different regions in the solution space to find the global optimal solution.

[0095] The multi-objective fitness function considers both energy consumption reduction and migration cost, which are two conflicting optimization objectives. The energy consumption reduction objective is quantified by calculating the difference in power consumption before and after migration. The migration cost objective includes network transmission time, service interruption time, resource reconfiguration overhead, and other cost factors. The fitness function combines the two objectives into a single evaluation index using a weighted summation method. The weight coefficients are determined based on business priority and operation strategy, with the energy consumption weight typically set to 0.6 to 0.7 and the migration cost weight set to 0.3 to 0.4. During the fitness evaluation process, the algorithm calculates the objective function value of each particle position vector corresponding to the migration scheme. The evaluation process needs to consider constraints such as the remaining capacity of the target server, network connection quality, and compatibility constraints. The particle fitness score matrix records the fitness value of each particle at the current position. The matrix rows correspond to different particles, and the columns contain detailed evaluation results such as fitness score, energy consumption reduction, and migration cost. This provides decision-making basis for subsequent particle updating.

[0096] The particle velocity update mechanism uses the standard particle swarm optimization algorithm's velocity update formula, which includes inertia term, cognitive term, and social term. The inertia term makes the particle maintain the current motion direction, the cognitive term makes the particle move towards the individual historical optimal position, and the social term makes the particle move towards the global optimal position. The position update mechanism adds the updated velocity to the current position to obtain the new position of the particle. After position update, boundary constraint checking is needed to ensure that the particle position is within the effective solution space. During the iterative optimization process, the algorithm repeatedly performs fitness evaluation, velocity update, position update, and global optimal update. After each iteration, the individual optimal position and global optimal position are updated. The iteration process continues until the convergence condition is met or the maximum iteration number is reached. The global optimal particle position is selected based on the optimal fitness value in the fitness score matrix, and the corresponding particle position vector represents the best virtual machine migration strategy found in the current population. The virtual machine migration path configuration includes detailed information such as the source server, target server, migration priority, and estimated migration time for each virtual machine. The configuration scheme ensures the orderly execution of migration operations and avoids resource conflicts.

[0097] The mapping association process establishes a detailed correspondence between the source server and the target server in the virtual machine migration path configuration. The associated information includes technical elements such as server hardware configuration, network connection topology, storage sharing relationship, and virtualization platform compatibility. The migration time series calculates the migration start time and expected completion time of each virtual machine based on factors such as the virtual machine's migration priority, data transmission volume, and network bandwidth capacity. The sequence arrangement avoids network congestion caused by migrating too many virtual machines at the same time. Resource reallocation calculates the new resource allocation status of the target server after receiving the migrated virtual machine, including the readjustment of resource configurations such as CPU allocation ratio, memory allocation, storage space allocation, and network bandwidth allocation. The server load reallocation table records the load changes of all servers involved in the migration operation in a tabular form. The table contains complete information such as server identification, load status before migration, load status after migration, load change, and new virtual machine distribution, providing detailed guidance for the execution and monitoring of the migration operation.

[0098] The resource management method for green cloud computing in the embodiment of the present application is described above. The resource management system for green cloud computing in the embodiment of the present application is described below. Figure 2 In one embodiment of the present application, a resource management system for green cloud computing includes:

[0099] A processing module is used to perform weight optimization processing on CPU utilization data and memory usage data using the Coyote optimization algorithm to obtain a server energy consumption coefficient matrix;

[0100] A decision module is used to perform DDQN decision processing on the virtual machine CPU allocation ratio according to the server energy consumption coefficient matrix to obtain a virtual machine resource allocation parameter table;

[0101] An adjustment module, configured to adjust the virtual machine resource allocation parameter table to the data backup copy storage location to obtain a backup node distribution configuration table;

[0102] A monitoring module is used to perform PUE monitoring on the real-time power consumption value of each server according to the backup node distribution configuration table to obtain a list of servers with excessive power consumption;

[0103] The calculation module is used to calculate the virtual machine migration path of the server list with excessive power consumption by using a particle swarm optimization algorithm to obtain a server load redistribution table.

[0104] above Figure 2 The resource management system for green cloud computing in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The resource management device for green cloud computing in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0105] ReferenceFigure 3 The embodiment of the present application also provides a resource management device for green cloud computing. The resource management device for green cloud computing can be a server, and its internal structure can be as shown in the figure. Figure 3 The resource management device for green cloud computing comprises a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. The processor of the computer is used to provide computing and control capabilities. The memory of the resource management device for green cloud computing comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the resource management device for green cloud computing is used to store corresponding data in the embodiment. The network interface of the resource management device for green cloud computing is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement the above method.

[0106] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the present application, and does not constitute a limitation on the resource management device for green cloud computing to which the present application is applied.

[0107] The present application also provides a computer readable storage medium. The computer readable storage medium can be a non-volatile computer readable storage medium, and can also be a volatile computer readable storage medium. The computer readable storage medium stores instructions. When the instructions are run on a computer, the computer executes the steps of the resource management method for green cloud computing.

[0108] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system, system and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0109] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a resource management device for green cloud computing (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A resource management method for green cloud computing, characterized in that, The method comprises: The CPU utilization data and the memory occupation data are processed by the coyote optimization algorithm to obtain a server energy consumption coefficient matrix; According to the server energy consumption coefficient matrix, the virtual machine CPU allocation ratio is processed by DDQN decision to obtain a virtual machine resource allocation parameter table; The virtual machine resource allocation parameter table is used to adjust the data backup copy storage location to obtain a backup node distribution configuration table; According to the backup node distribution configuration table, the real-time power consumption value of each server is monitored by PUE to obtain a power consumption exceeding standard server list; The power consumption exceeding standard server list is calculated by the particle swarm optimization algorithm to obtain a server load redistribution table.

2. The resource management method for green cloud computing according to claim 1, wherein, The CPU utilization data and the memory occupation data are processed by the coyote optimization algorithm to obtain a server energy consumption coefficient matrix, comprising: The CPU utilization rate of each server node is collected, and the standardized CPU utilization rate vector is obtained after maximum and minimum value scaling processing; The memory occupation data and the standardized CPU utilization rate vector are spliced by column to obtain a multi-dimensional state matrix containing double resource indicators; The fitness function with the minimum energy consumption as the target is constructed, the multi-dimensional state matrix is initialized by the coyote algorithm to generate a population, and the initial weight solution set is obtained; The social hierarchy structure of the coyote population is simulated, the Alpha, Beta and Omega roles are divided to cooperatively search and optimize the initial weight solution set, and the optimal weight solution after convergence is obtained; Based on the server CPU basic power consumption and the memory power consumption coefficient, the optimal weight solution is linearly combined to obtain a server energy consumption coefficient matrix reflecting the energy consumption characteristics of each server. 3.The resource management method for green cloud computing according to claim 1, wherein, The CPU allocation ratio of the virtual machine is processed by DDQN decision according to the server energy consumption coefficient matrix to obtain a virtual machine resource allocation parameter table, comprising: The server energy consumption coefficient matrix is converted into a state space vector, and a multi-dimensional state input is obtained after dimension mapping processing; The double network architecture containing the target network and the evaluation network is constructed, the multi-dimensional state input is processed by forward propagation to obtain a Q value action matrix; Based on the ε-greedy strategy, the CPU allocation action is selected from the Q value action matrix, and the sample storage is processed by the experience replay mechanism to obtain a training buffer; The training buffer is updated by the time difference learning to obtain a DDQN decision model; According to the virtual machine load demand, the CPU allocation ratio and the memory allocation amount output by the DDQN decision model are combined to obtain a virtual machine resource allocation parameter table.

4. The resource management method for green cloud computing according to claim 1, wherein, The virtual machine resource allocation parameter table is used to adjust the data backup copy storage location to obtain a backup node distribution configuration table, comprising: The virtual machine identifier and the CPU allocation ratio are extracted from the virtual machine resource allocation parameter table, and the data priority classification matrix is obtained by data importance rating processing; judging importance levels of each data block in the data priority classification matrix, assigning 5 backup copies when the importance level is critical, and assigning 3 backup copies when the importance level is general, to obtain a copy number assignment table; performing load balancing calculation processing on storage capacities of each data center based on the copy number assignment table, to obtain a storage node candidate list; performing cross-region distribution optimization processing on the storage node candidate list according to a data center geographical location distance matrix, to obtain a backup storage path scheme; performing association mapping processing on the backup storage path scheme and virtual machine resource allocation parameters, to obtain a backup node distribution configuration table containing specific storage node addresses.

5. The resource management method for green cloud computing according to claim 4, wherein, The cross-region distribution optimization processing on the storage node candidate list according to the data center geographical location distance matrix to obtain the backup storage path scheme includes: performing spherical distance calculation processing on the latitude and longitude coordinates of each data center, to obtain distance measurement values between data centers; constructing the distance measurement values into a symmetric matrix form, performing connectivity analysis processing through a minimum spanning tree algorithm, to obtain a regional connection topology graph; setting a preset distance threshold based on the regional connection topology graph, marking two data centers as a cross-region node pair when the distance between them exceeds the threshold, to obtain a cross-region node grouping; performing greedy algorithm traversal processing on the storage node candidate list according to the cross-region node grouping, selecting a node combination that minimizes the total transmission distance, to obtain an optimal node combination scheme; sorting each node in the optimal node combination scheme according to transmission priority, to generate a backup storage path scheme containing a primary backup path and a secondary backup path.

6. The resource management method for green cloud computing according to claim 1, wherein, The PUE monitoring processing on real-time power consumption values of each server according to the backup node distribution configuration table to obtain a power consumption over-standard server list includes: extracting server node identifiers and storage load distribution from the backup node distribution configuration table, collecting real-time current and voltage data of each server through a power consumption sensor, to obtain a server power consumption original data set; performing cumulative calculation processing on the server power consumption original data set, cooling system power consumption, and lighting system power consumption, to obtain a data center total power consumption value; performing ratio operation processing on the data center total power consumption value and IT equipment power consumption, to obtain a real-time PUE monitoring value; judging whether the real-time PUE monitoring value exceeds a preset PUE threshold, marking the corresponding server as an over-standard state when the threshold is exceeded, performing screening processing through an anomaly detection algorithm, to obtain an abnormal power consumption server set; performing descending order arrangement processing on the power consumption deviation degrees of each server in the abnormal power consumption server set, to obtain a power consumption over-standard server list sorted by the severity of power consumption over-standard.

7. The resource management method for green cloud computing according to claim 1, wherein, The virtual machine migration path calculation processing on the power consumption over-standard server list through the particle swarm optimization algorithm to obtain a server load reallocation table includes: extracting virtual machine distribution information of over-standard servers from the power consumption over-standard server list, quantifying resource occupancy of each virtual machine through a load calculation algorithm, to obtain a virtual machine migration candidate set; Initialize a particle swarm population, encode each migration scheme in the virtual machine migration candidate set as a particle position vector, and obtain an initial particle swarm body through a random generation process; Construct a multi-objective fitness function targeting energy consumption reduction and migration cost, perform fitness evaluation processing on each particle in the initial particle swarm body, and obtain a particle fitness score matrix; Based on particle velocity updating and position updating mechanisms, perform iterative optimization processing on the particle fitness score matrix, select the global optimal particle position as the optimal migration strategy, and obtain a virtual machine migration path configuration; Map and associate the source server and target server in the virtual machine migration path configuration, and generate a server load redistribution table containing migration time series and resource reallocation.

8. A resource management system for green cloud computing, characterized by, The green cloud computing-oriented resource management system for implementing the green cloud computing-oriented resource management method according to any one of claims 1-7 comprises: A processing module configured to perform weight optimization processing on CPU utilization data and memory occupation data through a wolf optimization algorithm to obtain a server energy consumption coefficient matrix; A decision module configured to perform DDQN decision processing on a virtual machine CPU allocation ratio based on the server energy consumption coefficient matrix to obtain a virtual machine resource allocation parameter table; An adjustment module configured to perform adjustment processing on a data backup copy storage location based on the virtual machine resource allocation parameter table to obtain a backup node distribution configuration table; A monitoring module configured to perform PUE monitoring processing on real-time power consumption values of each server based on the backup node distribution configuration table to obtain a power consumption exceeding standard server list; A calculation module configured to perform virtual machine migration path calculation processing on the power consumption exceeding standard server list through a particle swarm optimization algorithm to obtain a server load redistribution table.

9. A resource management device for green cloud computing, characterized by, The computer program, when executed by the processor, causes the processor to perform the green cloud computing-oriented resource management method according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, causes the processor to perform the green cloud computing-oriented resource management method according to any one of claims 1-7.

Images