Energy consumption optimization method, device, equipment and storage medium for server

Abstract

The present invention relates to a method, device, equipment and storage medium for optimizing the energy consumption of a server, aiming to reduce the energy consumption of the server during operation and improve the level of energy efficiency management. The method monitors the energy consumption of the server under different load conditions through super-resolution data acquisition technology, and generates load feature embedding representations in combination with a contrastive learning network. Based on these embedded representations, energy consumption is predicted using a generative adversarial network, and a hardware operation parameter adjustment scheme is generated through a Markov decision process and a particle swarm optimization algorithm. Finally, the scheduling strategy is iteratively optimized through distributed reinforcement learning to achieve adaptive regulation, generate a target energy consumption optimization scheme, and significantly improve the accuracy and dynamic response capability of server energy consumption management.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method, device, computer equipment and storage medium for optimizing energy consumption of a server. Background Art

[0002] In modern data centers and cloud computing environments, as the number of servers and performance requirements continue to increase, energy consumption issues have become particularly prominent. As the core equipment for data processing and storage, the energy consumption of servers accounts for a major part of the entire data center's electricity consumption. Traditional server energy consumption management methods often rely on static adjustment strategies or simple load sensing methods, which cannot fully cope with the dynamically changing load characteristics and complex energy consumption interactions between hardware components in the server operating environment. This leads to the problem of low energy efficiency in servers under both high and low load conditions, and it is impossible to effectively reduce energy consumption while ensuring the stability of computing performance.

[0003] In response to the above-mentioned shortcomings in energy management, researchers have proposed many scheduling methods based on machine learning and intelligent optimization, hoping to improve the energy efficiency of servers through the application of intelligent algorithms. However, most of the current research has problems such as insufficient model generalization ability, low real-time performance, and inability to handle complex hardware configurations. Existing energy consumption prediction models usually cannot accurately reflect the energy consumption changes of servers under different load conditions, which limits the accuracy and effectiveness of dynamic scheduling strategies. In addition, when faced with multi-variable coupled hardware parameters, commonly used optimization algorithms are prone to fall into local optimal solutions, making it difficult to find the global optimal energy consumption optimization solution. These problems seriously restrict the actual application effect of server energy consumption optimization.

[0004] Therefore, how to achieve efficient energy consumption management through more advanced optimization methods during server operation has become one of the current research hotspots. In particular, how to use deep learning technology to model complex load behaviors, improve the accuracy of energy consumption prediction through generative adversarial networks, and combine distributed reinforcement learning to achieve adaptive hardware parameter adjustment, and gradually form a dynamic and adaptive energy consumption optimization strategy. There are still many challenges and problems to be solved in these aspects. Summary of the invention

[0005] The main purpose of the present invention is to provide a method, device, equipment and storage medium for optimizing energy consumption of a server, so as to solve the technical problems of poor energy consumption optimization effect and insufficient dynamic adaptability in the prior art.

[0006] To achieve the above-mentioned purpose, the present invention provides a method for optimizing the energy consumption of a server, comprising the following steps: monitoring the energy consumption of different load states of the server through super-resolution data acquisition technology to obtain an initial energy consumption data set and a load behavior data set; performing feature learning on the initial energy consumption data set and the load behavior data set through a comparative learning network to generate an energy consumption feature embedding vector, and embedding the energy consumption features under similar load conditions in the same vector space to obtain an embedding representation of multiple load categories; based on the embedding representation of the multiple load categories, constructing an energy consumption prediction module using a generative adversarial network, and generating energy consumption prediction data through an adversarial process between a generator and a discriminator. ; Based on the energy consumption prediction data, a dynamic scheduling optimization strategy is established using the Markov decision process, and the hardware operating parameters are adjusted in combination with the particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan; the hardware operating parameter adjustment plan is applied to each hardware component of the server to generate real-time feedback data, and based on the real-time feedback data, the scheduling optimization strategy is iteratively optimized through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy; based on the optimized scheduling strategy, the server operating state is adaptively controlled to generate a target energy consumption optimization plan, which is verified through multi-dimensional performance indicators to obtain a target energy consumption optimization configuration plan for the server.

[0007] The present invention provides an energy consumption optimization device for a server, comprising: a monitoring unit, used to monitor the energy consumption of different load states of the server through super-resolution data acquisition technology, and obtain an initial energy consumption data set and a load behavior data set; a first generating unit, used to perform feature learning on the initial energy consumption data set and the load behavior data set through a contrast learning network, generate an energy consumption feature embedding vector, and embed the energy consumption features under similar load conditions in the same vector space to obtain an embedding representation of multiple load categories; a second generating unit, used to construct an energy consumption prediction module based on the embedding representation of the multiple load categories using a generative adversarial network, and generate energy consumption prediction data through an adversarial process between a generator and a discriminator; an adjusting unit A unit is used to establish a dynamic scheduling optimization strategy based on the energy consumption prediction data by using the Markov decision process, and adjust the hardware operating parameters in combination with the particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan; an optimization unit is used to apply the hardware operating parameter adjustment plan to each hardware component of the server, generate real-time feedback data, and iteratively optimize the scheduling optimization strategy based on the real-time feedback data through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy; an acquisition unit is used to adaptively control the server operating state based on the optimized scheduling strategy, generate a target energy consumption optimization plan, and verify it through multi-dimensional performance indicators to obtain a target energy consumption optimization configuration plan for the server.

[0008] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0009] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0010] The present invention provides a method, device, computer equipment and storage medium for optimizing the energy consumption of a server. Through the application of a multi-stage intelligent optimization strategy, the refined management and adaptive regulation of the server energy consumption are achieved, the energy consumption management efficiency of the server under different load conditions is improved, and the overall energy consumption is reduced. The method realizes accurate monitoring of energy consumption through super-resolution data acquisition technology, and improves the extraction and prediction accuracy of energy consumption features through the combination of comparative learning network and generative adversarial network. Based on the generated energy consumption prediction data, the hardware parameters are dynamically adjusted using the Markov decision process and particle swarm optimization algorithm, and are continuously optimized through the distributed reinforcement learning algorithm, and finally an optimized energy consumption scheduling strategy is generated. This solution can effectively reduce the power consumption of the server during operation, while ensuring the stability of system performance, and solves the problems of poor energy consumption optimization effect and insufficient dynamic adaptability in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the steps of a method for optimizing energy consumption of a server in one embodiment of the present invention;

[0012] Figure 2 is a structural block diagram of an energy consumption optimization device for a server in one embodiment of the present invention;

[0013] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0014] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] Reference Figure 1 The embodiment of the present invention provides a method for optimizing energy consumption of a server, comprising the following steps:

[0017] S1, monitor the energy consumption of the server in different load states through super-resolution data acquisition technology to obtain the initial energy consumption data set and load behavior data set.

[0018] S2, performing feature learning on the initial energy consumption data set and the load behavior data set through a comparative learning network, generating an energy consumption feature embedding vector, and embedding the energy consumption features under similar load conditions in the same vector space to obtain embedded representations of multiple load categories.

[0019] S3, based on the embedded representation of the multiple load categories, an energy consumption prediction module is constructed using a generative adversarial network, and energy consumption prediction data is generated through an adversarial process between a generator and a discriminator.

[0020] S4, based on the energy consumption prediction data, a dynamic scheduling optimization strategy is established by using a Markov decision process, and the hardware operating parameters are adjusted in combination with a particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan.

[0021] S5, applying the hardware operating parameter adjustment scheme to each hardware component of the server to generate real-time feedback data, and based on the real-time feedback data, iteratively optimizing the scheduling optimization strategy through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy.

[0022] S6, based on the optimized scheduling strategy, adaptively control the server operation status, generate a target energy consumption optimization plan, and verify it through multi-dimensional performance indicators to obtain the server's target energy consumption optimization configuration plan.

[0023] That is, the energy consumption optimization method of the server provided by the embodiment of the present invention aims to achieve refined management of server energy consumption through a multi-stage optimization strategy, and improve the energy efficiency performance of the server under different load conditions. First, the energy consumption of the server under different load conditions is monitored through super-resolution data acquisition technology. This process can accurately capture the energy consumption changes of the server under different working conditions, thereby generating an initial energy consumption data set and a load behavior data set. In order to ensure the accuracy of energy consumption data, super-resolution data acquisition technology includes high-frequency data sampling and multi-sensor data fusion, which can reflect the power consumption characteristics of server hardware components in real time and accurately.

[0024] After obtaining the initial energy consumption data set and load behavior data set, feature learning is performed on them through a contrastive learning network to generate an energy consumption feature embedding vector. This process uses the contrastive learning method to extract and analyze the features of the initial energy consumption data and load behavior data, and embeds the energy consumption features under similar load conditions in the same vector space to form an embedded representation of multiple load categories. Through this feature learning and embedding mapping process, the energy consumption characteristics of servers under different loads can be effectively distinguished, thereby providing a basis for subsequent energy consumption prediction. Then, based on the embedded representation of these load categories, the energy consumption prediction module is constructed using a generative adversarial network (GAN). The adversarial process between the generator and the discriminator enables the model to generate energy consumption prediction data that is closer to the actual situation. The generative adversarial network generates prediction data through the generator, and at the same time evaluates the difference between the generated data and the actual energy consumption data through the discriminator. This competitive training mechanism can continuously improve the authenticity of the prediction data and make the energy consumption prediction results more accurate. This energy consumption prediction data provides an important basis for adjusting the operating parameters of the server hardware.

[0025] Based on the above energy consumption prediction data, the Markov decision process is used to establish a dynamic scheduling optimization strategy, and then the hardware operating parameters are adjusted in combination with the particle swarm optimization algorithm to obtain the optimal hardware operating parameter adjustment solution. In this process, the Markov decision process is used to describe the changing characteristics of the hardware operating parameters of the server under different load conditions, and defines the state set, scheduling action set and state transition probability, thereby forming an optimization model for dynamic scheduling. The particle swarm optimization algorithm adjusts the hardware parameters through global optimization to ensure that the optimal solution is found globally to minimize energy consumption.

[0026] After obtaining the hardware operation parameter adjustment scheme, it is applied to each hardware component of the server, and real-time feedback data is generated during the adjustment process. These feedback data include the power consumption parameters, temperature parameters, and load parameters of the server at different time intervals. These real-time feedback data are iteratively optimized through the distributed reinforcement learning algorithm, and finally the optimized scheduling strategy is obtained. The distributed reinforcement learning algorithm uses a multi-agent architecture, and each agent independently evaluates and improves the scheduling strategy, and improves the overall optimization capability by sharing experience, thereby generating a more optimized scheduling scheme.

[0027] Based on the optimized scheduling strategy, the system adaptively regulates the running state of the server and finally generates a target energy consumption optimization scheme. The scheme collects the running data of each hardware component of the server under different load conditions in real time and normalizes these data to eliminate the dimensional differences between the hardware parameters. The normalized data is reduced in dimension by principal component analysis (PCA) to extract the main running feature dimensions, further improving the effect of adaptive regulation. Finally, the optimization scheme is verified by multi-dimensional performance indicators to ensure the effectiveness of the optimization configuration scheme in practical applications. These performance indicators include system response time, power consumption, temperature change rate and throughput. Through the verification of multi-dimensional performance indicators, the running state of the server can be comprehensively evaluated to ensure that the target energy consumption optimization configuration scheme finally selected achieves the best balance between energy consumption and performance. Through this series of steps, the present invention realizes efficient management of server energy consumption, reduces overall energy consumption and ensures the stability of system performance.

[0028] In one example, energy consumption of a server under different load states is monitored by super-resolution data acquisition technology to obtain an initial energy consumption data set and a load behavior data set, including: using super-resolution data acquisition technology to perform fine-grained monitoring of the energy consumption of the server under different load states, wherein the super-resolution data acquisition technology includes high-frequency sampling and multi-sensor data fusion to capture real-time power consumption changes of server hardware components; performing data preprocessing on the collected energy consumption data, removing outliers and interpolating missing values ​​to generate an initial energy consumption data set; and monitoring the working behavior of the server under different load states, collecting multiple load indicators, wherein the load indicators include at least CPU utilization, memory occupancy, and I / O operation frequency, and performing data preprocessing on the load indicators, removing outliers and interpolating missing values ​​to generate a load behavior data set; and associating the initial energy consumption data set with the load behavior data set to characterize the energy consumption behavior of the target server under different load modes.

[0029] In this example, the energy consumption of the server under different load conditions is monitored by super-resolution data acquisition technology, thereby generating an initial energy consumption data set and a load behavior data set, aiming to provide a detailed and accurate data basis for subsequent energy consumption optimization. Specifically, this process involves several key steps, the first of which is to conduct refined monitoring of the energy consumption of the server under different load conditions. In order to achieve high-precision energy consumption monitoring, super-resolution data acquisition technology is applied to the data acquisition process, including high-frequency sampling and multi-sensor data fusion. High-frequency sampling means that the frequency of collecting server hardware power consumption data is greatly increased, and the real-time power consumption fluctuations of server hardware components under different load changes can be captured more timely. The application of multi-sensor data fusion can integrate the power consumption data collected by multiple sensors, build the overall energy consumption characteristics of the server from the perspective of different hardware components, and ensure the comprehensiveness and accuracy of the data.

[0030] Based on data collection, the next step is to preprocess the collected energy consumption data to generate an initial energy consumption data set. Data preprocessing is one of the key steps in data analysis. It ensures the accuracy and consistency of the data by removing outliers and interpolating missing values. Outliers may be caused by instantaneous interference or hardware problems during data collection. These data deviate from the normal range and need to be removed to avoid misleading subsequent analysis and model training. The missing values ​​are processed by interpolation and completion. The missing values ​​are completed by fitting the adjacent data or other prediction methods, so that the continuity of the data set in space and time can be maintained, thereby better describing the changing law of server energy consumption.

[0031] At the same time, in addition to collecting energy consumption data, it is also necessary to monitor the working behavior of the server under different load conditions and collect multiple load indicators. These load indicators include at least key parameters such as CPU utilization, memory occupancy, and I / O operation frequency to reflect the current working status and load of the server. The data collection of these load indicators also requires data preprocessing to remove possible outliers and complete missing data to generate a load behavior data set. CPU utilization reflects the load on the server's processor during work, memory occupancy shows the memory resources used by the server when running applications, and I / O operation frequency can reveal the working status of the server's hard disk and other peripheral devices. These load indicators are very important for understanding the server's load status and the relationship between energy consumption and load.

[0032] After completing the collection and preprocessing of energy consumption data and load behavior data, the initial energy consumption data set needs to be associated with the load behavior data set. This association process aims to establish a corresponding relationship between energy consumption and load, and characterize the energy consumption behavior of the server under different load modes. Through such an association, the energy consumption characteristics of the server under different working conditions can be revealed, and it can be identified which load conditions have the highest energy consumption of the server and which hardware components are the main contributors to energy consumption under specific loads. This can provide a basis for subsequent energy consumption optimization steps. For example, energy consumption can be reduced by adjusting hardware operating parameters or optimizing scheduling strategies under high energy consumption load conditions, thereby improving overall energy efficiency.

[0033] In summary, super-resolution data acquisition technology combined with multi-sensor data fusion and high-frequency sampling methods can achieve refined monitoring of server energy consumption under different load conditions. Subsequently, the data is preprocessed and the relationship between load and energy consumption is established to build an initial data foundation for describing the energy consumption characteristics of the server. This process provides detailed and reliable data support for subsequent feature extraction, energy consumption prediction and optimization, and is an important prerequisite for achieving server energy consumption optimization.

[0034] In one example, feature learning is performed on the initial energy consumption data set and the load behavior data set through a contrastive learning network to generate an energy consumption feature embedding vector, and the energy consumption features under similar load conditions are embedded in the same vector space to obtain embedded representations of multiple load categories, including: preprocessing the association results of the initial energy consumption data set and the load behavior data set, standardizing the energy consumption data and behavior data under each load state, respectively, to generate standardized energy consumption data and standardized load behavior data; inputting the standardized energy consumption data and standardized load behavior data into the contrastive learning network, constructing a similarity measurement model through the contrastive learning network, and extracting features from the standardized energy consumption data and standardized load behavior data under different load states to obtain an initial energy consumption feature vector and an initial load behavior feature vector; embedding mapping is performed on the initial energy consumption feature vector and the initial load behavior feature vector to map them to the same high-dimensional vector space to obtain a feature embedding vector; and clustering analysis is performed on the feature embedding vector, energy consumption features of similar loads are clustered based on a clustering algorithm to obtain embedded representations of multiple load categories, each embedded representation corresponding to a specific load category.

[0035] In this example, the contrastive learning network is used to perform feature learning on the initial energy consumption dataset and the load behavior dataset to generate an energy consumption feature embedding vector, and the energy consumption features with similar load characteristics are embedded in the same vector space, and finally the embedding representation of multiple load categories is obtained. First, the association results of the initial energy consumption dataset and the load behavior dataset need to be preprocessed. This process is an important step to ensure data consistency and model training effect. In data preprocessing, the energy consumption data and load behavior data under each load state are standardized to generate standardized energy consumption data and standardized load behavior data. The standardization process is mainly to eliminate the dimensional differences between different features so that the data has the same scale, thereby avoiding the influence of large differences in data values ​​on the model learning process.

[0036] After obtaining the standardized data, these standardized energy consumption data and standardized load behavior data are input into the contrastive learning network, and the similarity measurement model is constructed through the contrastive learning network. The core of the contrastive learning network is to maximize the similarity of data under similar load conditions and minimize the difference of data under different load conditions through the process of feature extraction. In this process, the contrastive learning network extracts features from the standardized energy consumption data and load behavior data to obtain the initial energy consumption feature vector and the initial load behavior feature vector. These feature vectors contain abstract representations of the energy consumption characteristics and load behavior of the server under different load conditions, and can effectively describe the implicit rules in the data.

[0037] Next, these initial energy consumption feature vectors and initial load behavior feature vectors are embedded and mapped. The purpose of the embedding mapping process is to map these feature vectors into the same high-dimensional vector space to obtain feature embedding vectors. This mapping process not only unifies different feature vectors into the same coordinate system, but also makes similar load features closer in the high-dimensional space. In other words, under the same or similar load conditions, the energy consumption characteristics of the server have a smaller distance in the high-dimensional vector space, while under different load conditions, these feature vectors have a larger spatial distance. By mapping these feature vectors into the same vector space, the intrinsic relationship between load and energy consumption can be better revealed, providing a basis for subsequent optimization and scheduling strategies.

[0038] After the feature embedding is completed, cluster analysis is performed on these feature embedding vectors. The purpose of cluster analysis is to classify loads with similar characteristics, so as to obtain embedded representations of multiple load categories. Specifically, clustering the feature embedding vectors based on the clustering algorithm can bring together the energy consumption characteristics of similar loads to obtain a set of embedded representations representing different load categories. Each embedded representation corresponds to a type of energy consumption characteristics of the server under a specific load state. This classification result can be used for subsequent energy consumption prediction and optimization strategy formulation. For example, by understanding the energy consumption characteristics corresponding to a certain type of load, the energy consumption of the load category in the future can be predicted, so that corresponding measures can be taken to regulate and control it to achieve the purpose of energy consumption optimization.

[0039] Through this whole process, the contrastive learning network effectively extracts and abstracts the associated features of the initial energy consumption and load behavior data, and obtains the results of classifying and describing the energy consumption characteristics of the server under different load conditions through embedding mapping and clustering analysis. This processing not only improves the understanding of server energy consumption and load behavior, but also provides a reliable foundation for subsequent energy consumption prediction, hardware regulation, and dynamic scheduling strategies. By embedding features into a high-dimensional vector space and clustering them, we can better understand the complex relationship between server energy consumption and load, thereby achieving more accurate energy consumption optimization in practical applications.

[0040] In one example, based on the embedded representations of the multiple load categories, an energy consumption prediction module is constructed using a generative adversarial network, and energy consumption prediction data is generated through an adversarial process between a generator and a discriminator, including: based on the embedded representations of the multiple load categories, a generative adversarial network model is established, the generative adversarial network includes a generator and a discriminator, the generator is used to generate energy consumption prediction data according to the input load category embedded representation, and the discriminator is used to judge the difference between the generated data and the actual energy consumption data; the embedded representation of the load category is input into the generator, the generator performs feature conversion on the input data through a multi-layer fully connected neural network to obtain energy consumption prediction data, and the generator continuously optimizes its parameters during the training process to make the generated energy consumption data closer to the actual data distribution; the energy consumption prediction data is generated by the generator. The energy consumption prediction data and the real energy consumption data set are simultaneously input into the discriminator, and the discriminator extracts and classifies the input data through a multi-layer convolutional neural network to determine whether it is real data and evaluate the generation ability of the generator, wherein the real energy consumption data is the initial energy consumption data set; in the training process of the generative adversarial network, through the adversarial process between the generator and the discriminator, the generator gradually improves the authenticity of the generated data, and the discriminator gradually enhances the ability to distinguish between the generated data and the real data, until the similarity between the energy consumption prediction data generated by the generator and the real energy consumption data meets the preset conditions, and a trained generative adversarial network model is obtained; and based on the trained generative adversarial network model, the embedded representations of multiple load categories are input to generate corresponding energy consumption prediction data.

[0041] In this example, an energy consumption prediction module is constructed by using a generative adversarial network (GAN) to accurately predict the energy consumption of the server under different load conditions. This process is based on the embedded representations of multiple load categories generated previously, and a generative adversarial network model is established. The model consists of two main parts: the generator and the discriminator. The generator is responsible for generating corresponding energy consumption prediction data based on the input load category embedding representation, while the discriminator is responsible for judging whether the generated data is close to the actual energy consumption data.

[0042] Specifically, we first establish a generative adversarial network model based on the embedding representation of multiple load categories. In this network, the task of the generator is to use the input load category embedding representation to generate prediction data similar to the actual energy consumption distribution. The generator transforms the input data features through a multi-layer fully connected neural network and gradually generates energy consumption prediction data. As the training progresses, the generator continuously adjusts its internal parameters so that the data it generates can be closer to the actual energy consumption distribution, thereby improving the authenticity of the generated data.

[0043] At the same time, the task of the discriminator is to classify the input data to determine whether the data is the predicted data from the generator or the actual energy consumption data actually collected. To this end, the discriminator usually uses a multi-layer convolutional neural network to extract features from the input data and finally output a classification result. The predicted data generated by the generator and the real energy consumption data set (that is, the energy consumption data set initially collected) are input into the discriminator at the same time. The discriminator continuously improves its ability to distinguish between generated data and real data in this process. It provides feedback to the generator to help it improve by continuously evaluating the difference between the energy consumption data generated by the generator and the real data.

[0044] During the training process of the generative adversarial network, an adversarial relationship is formed between the generator and the discriminator, that is, the generator tries to generate data as close to the real data as possible, while the discriminator tries to accurately distinguish between the real data and the generated data. This adversarial training mechanism enables the generator to continuously improve the authenticity of its generated data through the feedback of the discriminator, and the similarity between the energy consumption prediction data and the real energy consumption data can finally meet the pre-set conditions. This means that the training process of the generative adversarial network has reached a balance point, and it is difficult for the discriminator to distinguish between the generated data and the real data, indicating that the generator's generation ability is already very close to the actual energy consumption situation.

[0045] Once the generative adversarial network model is trained, it can be used to predict energy consumption. Specifically, the embedded representations of multiple load categories are input into the trained generative adversarial network model, and the generator generates corresponding energy consumption prediction data based on these inputs. These prediction data reflect the possible energy consumption of the server under a specific load state, and can provide an important reference for subsequent scheduling optimization strategies.

[0046] In summary, by building and training a generative adversarial network model, we can use the load category embedding representation of the server under different load conditions to accurately predict the energy consumption data of the server. In this process, the adversarial relationship between the generator and the discriminator drives the model to continuously improve the accuracy of the generated data, and finally forms a generative adversarial network model that can generate high-quality energy consumption prediction data. Compared with traditional energy consumption prediction models, this energy consumption prediction method can better capture the complex relationship between load characteristics and energy consumption, thereby improving the accuracy and reliability of the prediction and providing a solid data foundation for energy consumption optimization.

[0047] In one example, based on the energy consumption prediction data, a dynamic scheduling optimization strategy is established using a Markov decision process, and the hardware operating parameters are adjusted in combination with a particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan, including: based on the energy consumption prediction data, defining a state set of server hardware operating states, a set of scheduling actions that can be taken, and corresponding state transition probabilities, and constructing a Markov decision process model to describe the changing characteristics of the hardware operating parameters of the server under different load states; obtaining the current load and energy consumption state of the server, and calculating the reward value of each scheduling action in different states based on the state transition in the Markov decision process and the current load and energy consumption state of the server, wherein the reward value is used to measure the contribution of different scheduling actions to server energy consumption optimization and performance improvement; based on the state transition probability data and the scheduling reward value, solving the scheduling actions under different load conditions to obtain a dynamic scheduling optimization strategy, and using the dynamic scheduling optimization strategy as input, and combining the particle swarm optimization algorithm to globally optimize the hardware operating parameters of the server to obtain a hardware operating parameter adjustment plan.

[0048] In this example, by further analyzing the server's energy consumption prediction data, a dynamic scheduling optimization strategy is established using the Markov decision process (MDP), and the particle swarm optimization algorithm (PSO) is combined to adjust the hardware operating parameters to achieve server energy consumption optimization and performance improvement. The entire process aims to design an efficient hardware parameter adjustment solution based on the load and energy consumption changes of the server to maximize energy efficiency.

[0049] First, based on the energy consumption prediction data, the state set, scheduling action set and state transition probability of the server hardware operating state are defined, and a Markov decision process model is constructed. Specifically, the state set describes the hardware operation status of the server at different time points, including CPU utilization, memory usage, I / O operation status, etc., while the scheduling action set includes the hardware parameter adjustment measures that the server may take, such as adjusting the CPU frequency, reducing memory power consumption, etc. The state transition probability describes the possibility of the server state changing from one state to another after executing a certain scheduling action. These state transition probabilities can be estimated by analyzing historical energy consumption and load data, thereby forming a description of the changing rules of server hardware parameters under different load conditions.

[0050] After obtaining the current server load and energy consumption status, the reward value of each scheduling action in different states is calculated based on the state transition in the Markov decision process and the current load and energy consumption status. The reward value is used to measure the contribution of each scheduling action to the energy consumption optimization and performance improvement of the server. For example, under a certain load state, adjusting the CPU frequency may significantly reduce energy consumption while ensuring a certain performance level, so the reward value of this scheduling action is higher. If a scheduling action can reduce energy consumption but significantly affects the performance of the server, its reward value will be lower. In this way, each possible scheduling action can be scored, so as to select the scheduling measure that is most conducive to energy consumption optimization and performance balance.

[0051] Next, based on the state transition probability data and the scheduling reward value, the scheduling actions under different load conditions are solved to obtain a dynamic scheduling optimization strategy. This optimization strategy describes the best scheduling actions that the server should take under different load and energy consumption conditions to maximize energy efficiency. In other words, it provides an optimal hardware parameter adjustment solution for each possible server state, so as to minimize energy consumption while maintaining system performance.

[0052] After obtaining the dynamic scheduling optimization strategy, the particle swarm optimization algorithm (PSO) is further combined to globally optimize the server's hardware operating parameters to obtain the final adjustment plan for the hardware operating parameters. The particle swarm optimization algorithm is an optimization algorithm based on swarm intelligence. It simulates the foraging behavior of a flock of birds and gradually approaches the optimal solution through the collaboration and information sharing of a group of "particles" in the solution space. Specifically, in the particle swarm optimization process, each particle represents a possible hardware parameter setting plan, and through multiple iterations, each particle continuously adjusts its position based on its own historical experience and the exploration results of other particles, and finally finds the optimal hardware operating parameter configuration.

[0053] Combined with the global search capability of the particle swarm optimization algorithm, it can effectively make up for the deficiency that the Markov decision process may fall into the local optimal solution, and ensure that the adjustment scheme of the hardware parameters is optimal in the global scope. In this way, the hardware parameters of the server can be adjusted more finely, which can not only effectively reduce energy consumption, but also ensure that the server maintains efficient and stable operation under different load conditions.

[0054] In summary, this example designs a dynamic scheduling optimization strategy and hardware operation parameter adjustment scheme by combining the Markov decision process and the particle swarm optimization algorithm. First, the Markov decision process describes the change law of the server's hardware parameters under different load conditions through the state set, the scheduling action set, and the state transition probability, and then calculates the reward value of the scheduling action and obtains the best scheduling strategy; then, the particle swarm optimization algorithm is combined to globally optimize the hardware parameters to ensure that the configuration of the hardware parameters can achieve the best balance between energy consumption and performance under different load conditions. This method can help the server intelligently adjust the hardware parameters under different working conditions, thereby effectively reducing energy consumption and improving energy efficiency, while maintaining high-performance operation of the system.

[0055] In one example, the hardware operating parameter adjustment scheme is applied to each hardware component of the server to generate real-time feedback data, and based on the real-time feedback data, the scheduling optimization strategy is iteratively optimized through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy, including: applying the hardware operating parameter adjustment scheme to each hardware component of the server, adjusting the operating parameters of the hardware components, and obtaining the adjusted operating status of the hardware components; based on the adjusted operating status of the hardware components, real-time sampling of the operating parameters of the server at different time intervals to generate multi-dimensional real-time feedback data, wherein the operating parameters include at least: power consumption parameters, temperature parameters and load parameters; feature extraction and dimensionality reduction processing are performed on the real-time feedback data to obtain feedback features. The method comprises the following steps: generating a set of feature vectors, and performing abnormal detection on the hardware operation state based on the feedback feature vector set, and marking the abnormal feedback feature vector; inputting the feedback feature vector set into the multi-agent architecture in the distributed reinforcement learning algorithm, and each agent independently performs strategy evaluation and strategy improvement on the current scheduling optimization strategy to generate its own scheduling improvement plan; sharing the experience of the scheduling improvement plan of each agent through the multi-agent experience pool, and using a genetic algorithm to optimize the scheduling improvement plan in the experience pool to obtain several candidate optimization strategies; based on the candidate optimization strategy, using the policy gradient method in the distributed reinforcement learning algorithm to iteratively solve the scheduling optimization strategy, and combining the best solution in the candidate optimization strategy to perform parameter fusion to obtain the final optimized scheduling strategy.

[0056] In this example, the hardware operating parameter adjustment scheme is applied to each hardware component of the server, and the server's operating status in different time periods is sampled and analyzed in real time. The scheduling strategy is gradually optimized by combining the distributed reinforcement learning algorithm, and finally the optimized scheduling strategy is obtained. The whole process ensures continuous improvement and intelligent optimization of server energy consumption management through closed-loop feedback and multi-agent collaboration.

[0057] First, apply the hardware operating parameter adjustment scheme to each hardware component of the server. Hardware components include CPU, memory, storage devices, etc. By adjusting the operating parameters of these hardware (such as CPU frequency, memory power consumption mode, etc.), the server enters a new operating state. These adjusted states will affect the overall energy consumption and performance of the server. In order to evaluate the effect of these adjustments, it is necessary to sample the operating state of the adjusted hardware components in real time. By collecting the operating parameters of the server at different time intervals, such as power consumption parameters, temperature parameters, and load parameters, multi-dimensional real-time feedback data can be generated. These data are used to evaluate the actual effect of the hardware parameter adjustment.

[0058] For these real-time feedback data, feature extraction and dimensionality reduction processing are first required to obtain a set of feedback feature vectors. These feedback data may contain operating information of various dimensions (such as power consumption, temperature, load, etc.). In order to better use them in learning algorithms, it is necessary to extract the most representative features and reduce the dimensions to reduce data complexity and highlight the main features. In this step, it is also necessary to perform anomaly detection on the hardware operating status. By analyzing the feedback feature vector set, those abnormal feedback feature vectors are marked. These abnormal feedback may represent the failure of hardware components or some undesirable operating conditions, which require special attention during the optimization process.

[0059] Next, the feedback feature vector set is input into the multi-agent architecture in the distributed reinforcement learning algorithm. The multi-agent architecture in distributed reinforcement learning refers to a system composed of multiple independent agents, each of which independently evaluates and improves the current scheduling optimization strategy, and then generates its own scheduling improvement plan. Each agent can be regarded as an independent decision maker, and they optimize the scheduling strategy based on their own feedback data. Through such a design, the learning ability of each agent can be fully utilized to explore a variety of different optimization paths, thereby increasing the chance of finding the global optimal solution.

[0060] In order to further improve the optimization effect, a multi-agent experience pool is used to share the experience of the scheduling improvement plans generated by each agent. The experience pool is a repository that records the scheduling actions and their effects performed by each agent in different situations, so that each agent can learn from the experience of other agents. Next, a genetic algorithm is used to optimize the scheduling improvement plans in the experience pool and select the one with better performance. Genetic algorithm is an optimization method based on biological evolution. It simulates the process of natural selection and generates better optimization plans through crossover and mutation operations. Under the action of genetic algorithm, several candidate optimization strategies can be selected from multiple scheduling improvement plans, providing a higher quality reference for subsequent reinforcement learning iterations.

[0061] Based on the selected candidate optimization strategies, the policy gradient method in the distributed reinforcement learning algorithm is used to iteratively solve the scheduling optimization strategy. The policy gradient method is an optimization method that improves the strategy by calculating the gradient of the strategy to the cumulative reward. Its purpose is to find a scheduling strategy that can maximize long-term benefits. Through the policy gradient method, combined with the best performing solution among the candidate optimization strategies for parameter fusion, the performance of the scheduling strategy can be effectively improved, thereby obtaining an optimized scheduling strategy.

[0062] Finally, after multiple iterations of solving and improving strategies, an optimized scheduling strategy is obtained. This strategy can provide the best operating parameter configuration for each hardware component of the server under different load conditions, which can not only effectively reduce energy consumption, but also ensure that the performance of the server reaches the expected level. Through this distributed reinforcement learning process based on real-time feedback and multi-agent collaboration, the server's scheduling optimization strategy can be continuously improved to adapt to dynamically changing load and energy consumption states, and achieve efficient resource management and control. This method has a high degree of flexibility and adaptability, allowing the server to maintain the best energy efficiency in a complex and changing environment.

[0063] In one example, based on the optimized scheduling strategy, the server operating state is adaptively regulated to generate a target energy consumption optimization plan, which is verified by multi-dimensional performance indicators to obtain the target energy consumption optimization configuration plan of the server, including: based on the optimized scheduling strategy, real-time collection of operating data of each hardware component of the server under different load conditions to generate a server operating state data set, wherein the operating data includes at least any of the following: CPU frequency, memory usage, I / O operation speed and temperature data; normalization of the server operating state data set to eliminate dimensional differences between different hardware parameters, and feature dimension reduction of the normalized data through principal component analysis (PCA), extraction of main operating feature dimensions, and generation of a reduced operating feature set; based on the operating feature set, real-time regulation of the hardware operating state of the server is performed through an adaptive regulation module, wherein the adaptive regulation module uses the optimized scheduling strategy, combined with the operating feature set The real-time changes of each dimension in the combination are combined, and the operating parameters of the server are gradually adjusted to generate a target energy consumption optimization plan; the target energy consumption optimization plan is applied to the hardware components of the server, and the hardware operating parameters are adjusted to cope with different load conditions. At the same time, the operating status data after each adjustment is recorded to generate an operating status data set after the optimization application; based on the operating status data set after the optimization application, the operating effect of the server under the target energy consumption optimization plan is verified to generate a verification result set, wherein the verification process is performed by calculating multi-dimensional performance indicators, and the multi-dimensional performance indicators include at least: system response time, power consumption, temperature change rate and throughput indicators; a multi-dimensional data comprehensive evaluation is performed on the verification result set, and the evaluation results of each performance indicator are weighted and integrated using a fuzzy comprehensive evaluation method to generate an energy consumption optimization evaluation score for the target server, and the target energy consumption optimization plan with the highest score is selected as the final target server energy consumption optimization configuration plan.

[0064] In this example, the server operating status is adaptively regulated based on the optimized scheduling strategy to generate a target energy consumption optimization plan, which is verified through multi-dimensional performance indicators to finally obtain the target server energy consumption optimization configuration plan. This process aims to achieve the optimal balance between energy consumption and performance by adaptively adjusting the server's hardware parameters, and select the best configuration plan through systematic verification and evaluation.

[0065] First, based on the optimized scheduling strategy, the operating data of each server hardware component under different load conditions is collected in real time to form a server operating status data set. This data set includes operating parameters such as CPU frequency, memory usage, I / O operation speed and temperature data, which reflect the specific status of the server under different working conditions. These data provide a basis for subsequent regulation and optimization.

[0066] After obtaining the operating status data, the data set needs to be normalized to eliminate the dimensional differences between different hardware parameters and ensure that the comparison between different parameters is reasonable and consistent. For example, the data units of CPU frequency and temperature are different, and the data that has not been normalized cannot be directly compared. Through normalization, these data can have the same scale, thus providing conditions for subsequent analysis and processing. After completing the normalization process, the principal component analysis (PCA) is used to reduce the dimension of these data. The role of PCA is to extract the main feature dimensions, simplify the complexity of the data, reduce the number of features while retaining the main information of the data. This dimensionality reduction process not only improves the efficiency of data processing, but also helps to highlight the features that are most important for energy consumption optimization.

[0067] Based on the reduced-dimensional operation feature set, the adaptive control module controls the server's hardware operation status in real time. In this process, the adaptive control module uses the optimized scheduling strategy and combines the real-time monitored operation feature set to gradually adjust the server's operation parameters and generate a target energy consumption optimization plan. The control module dynamically adjusts the hardware parameters according to the server's current load and operation status to ensure that the server can maintain the optimal balance between energy consumption and performance under different load conditions.

[0068] Apply the generated target energy consumption optimization solution to the server's hardware components, adjust the hardware operating parameters to cope with different load conditions, and record the current operating status after each adjustment to generate an operating status data set after the optimization application. This data set contains the actual operating performance of each hardware component of the server after the target energy consumption optimization solution is adopted, which provides a basis for subsequent verification steps.

[0069] Next, based on the running status data set after the optimization application, the running effect of the server under the target energy consumption optimization scheme is verified to generate a set of verification results. The verification process is completed by calculating multi-dimensional performance indicators, which include at least: system response time, power consumption, temperature change rate and throughput indicators. The system response time reflects the response speed of the server to the request, the power consumption reflects the overall energy consumption, the temperature change rate indicates the heat dissipation performance of the hardware, and the throughput indicator measures the amount of tasks that the server can handle per unit time. The comprehensive evaluation of these performance indicators can help determine the effectiveness of the optimization scheme.

[0070] Finally, a multi-dimensional comprehensive evaluation is performed on the verification result set, and the evaluation results of each performance indicator are weighted and integrated using the fuzzy comprehensive evaluation method to generate the energy consumption optimization evaluation score of the target server. The fuzzy comprehensive evaluation method is a tool for processing multi-indicator and multi-factor comprehensive decision-making. It can weight different indicators according to the relative importance of each performance indicator to calculate the overall evaluation result. In this process, the weights and evaluation scores of each indicator are comprehensively considered to obtain the final score of each target energy consumption optimization solution.

[0071] According to the evaluation results, the target energy consumption optimization solution with the highest score is selected as the final target server energy consumption optimization configuration solution. This optimization configuration solution represents the optimal hardware parameter configuration that can maximize server performance while minimizing energy consumption under different load conditions. Through such a series of adaptive control, verification and evaluation steps, efficient management of server energy consumption is finally achieved, ensuring the best balance between system performance and energy efficiency.

[0072] In order to better understand the technical solution of the present application, a specific embodiment is now provided, which is as follows:

[0073] On the servers in the data center, we first use super-resolution data acquisition technology to finely monitor energy consumption under different load conditions. In different time periods, for example, between 9 a.m. and 9 p.m., the server load is mainly concentrated on the response of streaming services and online applications, while from 9 p.m. to the early hours of the next morning, it is more about background batch processing and offline data analysis tasks. Therefore, high-frequency sampling and multi-sensor data fusion are needed to capture the real-time power consumption changes of server hardware components.

[0074] Specifically, during the energy consumption data collection process, the sampling frequency was set to 10,000 times per second to ensure that changes in power consumption can be accurately recorded when the server load fluctuates. At the same time, multiple sensors were used to collect power consumption data for the CPU, power supply, memory, and fan, and the data was fused. After data preprocessing, outliers were removed, such as spike data caused by hardware failures, and missing values ​​were interpolated to obtain the initial energy consumption data set and load behavior data set.

[0075] The load behavior data set includes multiple indicators such as CPU utilization, memory usage, I / O operation frequency, etc. For example, during a peak period, the collected data shows that the CPU utilization is above 85%, the memory usage is 70%, and the I / O operation frequency is above 800 times per second. By standardizing these data, standardized energy consumption data and standardized load behavior data are generated.

[0076] After obtaining the initial energy consumption data set and load behavior data set, the contrastive learning network is used to perform feature learning on these data to generate energy consumption feature embedding vectors. In this process, the collected data is first preprocessed to ensure that the energy consumption data and behavior data under each load state are standardized. The contrastive learning network extracts energy consumption feature vectors under different load states by constructing a similarity measurement model. For example, for video on demand applications, high CPU and memory utilization are the main features, while for offline batch processing tasks, the frequency of I / O operations is the key feature.

[0077] Subsequently, the feature vectors are mapped into the same high-dimensional vector space to form feature embedding vectors. These embedding vectors represent the energy consumption characteristics of the server under similar load conditions. Through clustering analysis, similar load categories can be grouped into one category. For example, video on demand and online games have similar CPU and memory usage patterns, so they are clustered into the same load category, while data analysis tasks and batch processing tasks are grouped into another category.

[0078] Based on the embedding representation of multiple load categories, an energy consumption prediction module was constructed using a generative adversarial network (GAN). In GAN, the generator generates energy consumption prediction data based on the input load category embedding representation, while the discriminator determines whether the generated data is close to the actual energy consumption data. During the training process, the generator and the discriminator constantly compete with each other to improve the authenticity of the generated energy consumption data, and finally a trained generative adversarial network model is obtained.

[0079] Based on the generated energy consumption prediction data, the Markov decision process (MDP) is used to establish a dynamic scheduling optimization strategy, and the particle swarm optimization (PSO) algorithm is combined to adjust the server hardware operating parameters. For example, for a certain type of video streaming service load, the MDP model defines a set of server states, including the current CPU utilization, memory occupancy, etc., as well as a set of scheduling actions that can be taken, such as adjusting the CPU frequency, adjusting the fan speed, etc. The reward value of each scheduling action in different states represents the contribution of the action to energy consumption optimization and performance improvement. For example, when the CPU utilization is too high, reducing the CPU frequency may reduce power consumption, but it may also cause video playback to freeze, so the reward value of this action is lower.

[0080] Based on the scheduling optimization strategy obtained by solving the MDP, the PSO algorithm is combined to globally optimize the server's hardware parameters. PSO simulates the process of a group of particles searching for the optimal solution in the solution space, and each particle represents a possible hardware parameter configuration. After multiple iterations, a set of hardware parameter adjustment solutions with the lowest energy consumption and the best performance were finally found.

[0081] Next, the hardware operating parameter adjustment scheme is applied to each hardware component of the server, and the operating status of the server in different time periods is sampled in real time. For example, after applying the optimization scheme, the power consumption of a server during the high-load period at night was reduced by 15%, the temperature dropped by 2°C, and the system response time was shortened by 10%. These operating parameters are collected in real time to generate multi-dimensional feedback data, including power consumption parameters, temperature parameters, and load parameters.

[0082] These real-time feedback data are subjected to feature extraction and dimensionality reduction to obtain a set of feedback feature vectors. Subsequently, the set of feedback feature vectors is input into the multi-agent architecture in the distributed reinforcement learning algorithm, and each agent independently evaluates and improves the current scheduling optimization strategy. For example, agent A is responsible for evaluating the changes in CPU utilization, agent B evaluates the impact of memory occupancy on power consumption, and agent C is responsible for the effect of fan speed on temperature control.

[0083] These agents learn from each other's experience through a shared experience pool, and use genetic algorithms to select the best scheduling solutions in the experience pool to obtain several candidate optimization strategies. The scheduling optimization strategy is iteratively solved through the policy gradient method, and the parameters are fused by combining the best solution in the candidate optimization strategy to finally obtain the optimized scheduling strategy.

[0084] After obtaining the optimized scheduling strategy, the operating data of each server hardware component under different load conditions is collected in real time to form a server operating status data set. For this data, it is necessary to first perform normalization processing to eliminate the dimensional differences between different hardware parameters. Then, the main operating feature dimensions are extracted through principal component analysis (PCA), such as merging features such as CPU frequency, memory usage, and I / O operation frequency into several main dimensions to generate a reduced dimension operating feature set.

[0085] The adaptive control module is used to control the hardware operation status of the server in real time. The module uses the optimized scheduling strategy and the real-time changing operation feature set to gradually adjust the server's operation parameters and generate a target energy consumption optimization plan. For example, for a server, when the CPU utilization rate is detected to be above 90%, the adaptive control module will reduce the CPU frequency by 10% and increase the fan speed to reduce the temperature and power consumption.

[0086] Finally, the target energy consumption optimization solution is applied to the hardware components of the server, the operating status is monitored and a data set of the operating status after the optimization application is generated. Based on this data set, the operating effect of the server is verified through multi-dimensional performance indicators. The verification process includes calculating indicators such as system response time, power consumption, temperature change rate and throughput. For example, through the optimized scheduling strategy, the system response time of a server is shortened by 15%, the power consumption is reduced by 12%, and the temperature change rate is more stable.

[0087] A multi-dimensional comprehensive evaluation was conducted on these verification result sets, and the fuzzy comprehensive evaluation method was used to weight and fuse the performance indicators to generate the energy consumption optimization evaluation score of the target server. The evaluation results showed that a certain solution had the highest score, which means that it performed best in all test scenarios. Therefore, this solution was selected as the final target server energy consumption optimization configuration solution.

[0088] In summary: In this complex scenario, efficient management of server energy consumption is achieved through a multi-stage optimization strategy. Super-resolution data acquisition technology provides high-precision energy consumption and load data, contrastive learning networks and generative adversarial networks are used to extract load characteristics and predict energy consumption, Markov decision processes and particle swarm optimization algorithms help to formulate optimal scheduling strategies, and distributed reinforcement learning further optimizes scheduling plans through real-time feedback. Finally, through adaptive regulation and multi-dimensional performance verification, a configuration solution that achieves the optimal balance between energy efficiency and performance is selected. This complex and sophisticated energy consumption optimization management method provides an effective energy consumption optimization solution for server operations in data centers, which reduces overall energy consumption while ensuring that system performance is not affected, significantly improving the operational efficiency of data centers.

[0089] Reference Figure 2 , an embodiment of the present invention provides an energy consumption optimization device for a server, comprising:

[0090] The monitoring unit 1 is used to monitor the energy consumption of the server under different load states by using super-resolution data acquisition technology to obtain an initial energy consumption data set and a load behavior data set.

[0091] The first generation unit 2 is used to perform feature learning on the initial energy consumption data set and the load behavior data set through a comparative learning network, generate an energy consumption feature embedding vector, and embed the energy consumption features under similar load conditions in the same vector space to obtain embedded representations of multiple load categories.

[0092] The second generation unit 3 is used to construct an energy consumption prediction module based on the embedded representation of the multiple load categories using a generative adversarial network, and generate energy consumption prediction data through an adversarial process between a generator and a discriminator.

[0093] The adjustment unit 4 is used to establish a dynamic scheduling optimization strategy based on the energy consumption prediction data by using a Markov decision process, and adjust the hardware operating parameters in combination with a particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan.

[0094] The optimization unit 5 is used to apply the hardware operating parameter adjustment scheme to each hardware component of the server, generate real-time feedback data, and iteratively optimize the scheduling optimization strategy based on the real-time feedback data through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy.

[0095] The acquisition unit 6 is used to adaptively control the server operation status based on the optimized scheduling strategy, generate a target energy consumption optimization plan, and verify it through multi-dimensional performance indicators to obtain the server's target energy consumption optimization configuration plan.

[0096] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0097] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes 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 operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0098] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0099] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0100] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0101] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0102] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for optimizing energy consumption of a server, characterized in that: The following steps are involved: The energy consumption of the server under different load states is monitored by super-resolution data acquisition technology to obtain the initial energy consumption data set and load behavior data set; Preprocessing the association results of the initial energy consumption data set and the load behavior data set, standardizing the energy consumption data and behavior data under each load state, generating standardized energy consumption data and standardized load behavior data; inputting the standardized energy consumption data and standardized load behavior data into a contrastive learning network, constructing a similarity measurement model through the contrastive learning network, performing feature extraction on the standardized energy consumption data and standardized load behavior data under different load states, and obtaining an initial energy consumption feature vector and an initial load behavior feature vector; performing embedding mapping processing on the initial energy consumption feature vector and the initial load behavior feature vector, mapping them to the same high-dimensional vector space, and obtaining a feature embedding vector; and, performing cluster analysis on the feature embedding vectors, clustering the energy consumption characteristics of similar loads based on a clustering algorithm, and obtaining embedding representations of multiple load categories, each embedding representation corresponding to a specific load category; Based on the embedding representations of the multiple load categories, a generative adversarial network model is established, wherein the generative adversarial network includes a generator and a discriminator, wherein the generator is used to generate energy consumption prediction data according to the input load category embedding representation, and the discriminator is used to judge the difference between the generated data and the real energy consumption data; the embedding representation of the load category is input into the generator, and the generator performs feature conversion on the input data through a multi-layer fully connected neural network to obtain energy consumption prediction data, and the generator continuously optimizes its parameters during the training process to make the generated energy consumption data closer to the real data distribution; the energy consumption prediction data and the real energy consumption data set are simultaneously input into the discriminator, and the discriminator performs feature conversion on the input data through a multi-layer fully connected neural network to obtain energy consumption prediction data. Extracting features and classifying input data to determine whether it is real data, and evaluating the generation capability of the generator, wherein the real energy consumption data is the initial energy consumption data set; in the training process of the generative adversarial network, through the adversarial process between the generator and the discriminator, the generator gradually improves the authenticity of the generated data, and the discriminator gradually enhances the ability to distinguish between the generated data and the real data, until the similarity between the energy consumption prediction data generated by the generator and the real energy consumption data meets the preset conditions, and a trained generative adversarial network model is obtained; and based on the trained generative adversarial network model, the embedded representations of multiple load categories are input to generate corresponding energy consumption prediction data; Based on the energy consumption prediction data, a dynamic scheduling optimization strategy is established by using a Markov decision process, and the hardware operating parameters are adjusted in combination with a particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan; Applying the hardware operating parameter adjustment scheme to each hardware component of the server to generate real-time feedback data, and iteratively optimizing the scheduling optimization strategy through a distributed reinforcement learning algorithm based on the real-time feedback data to obtain an optimized scheduling strategy; Based on the optimized scheduling strategy, the server operating status is adaptively controlled to generate a target energy consumption optimization plan, which is verified through multi-dimensional performance indicators to obtain a target energy consumption optimization configuration plan for the server.

2. The energy consumption optimization method according to claim 1, characterized in that: The energy consumption of the server under different load conditions is monitored by super-resolution data acquisition technology to obtain the initial energy consumption data set and load behavior data set, including: Using super-resolution data acquisition technology to monitor the energy consumption of the server under different load conditions in a refined manner, wherein the super-resolution data acquisition technology includes high-frequency sampling and multi-sensor data fusion to capture the real-time power consumption changes of server hardware components; Preprocessing the collected energy consumption data, removing abnormal values ​​and interpolating missing values ​​to generate an initial energy consumption data set; and Monitor the working behavior of the server under different load conditions, collect multiple load indicators, the load indicators at least include CPU utilization, memory occupancy, I / O operation frequency, and perform data preprocessing on the load indicators, remove outliers and interpolate missing values ​​to generate a load behavior data set; The initial energy consumption data set is associated with the load behavior data set to characterize the energy consumption behavior of the target server under different load modes.

3. The energy consumption optimization method according to claim 1, characterized in that: Based on the energy consumption prediction data, a dynamic scheduling optimization strategy is established using the Markov decision process, and the hardware operating parameters are adjusted in combination with the particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan, including: Based on the energy consumption prediction data, the state set of the server hardware operating state, the set of scheduling actions that can be taken, and the corresponding state transition probability are defined, and a Markov decision process model is constructed to describe the changing characteristics of the server hardware operating parameters under different load conditions; Obtain the current load and energy consumption state of the server, and calculate the reward value of each scheduling action in different states according to the state transition in the Markov decision process and the current load and energy consumption state of the server, wherein the reward value is used to measure the contribution of different scheduling actions to the energy consumption optimization and performance improvement of the server; Based on the state transition probability data and the scheduling reward value, the scheduling actions under different load conditions are solved to obtain a dynamic scheduling optimization strategy. The dynamic scheduling optimization strategy is used as input, and the hardware operating parameters of the server are globally optimized in combination with the particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan.

4. The energy consumption optimization method according to claim 1, characterized in that: The hardware operation parameter adjustment scheme is applied to each hardware component of the server to generate real-time feedback data, and based on the real-time feedback data, the scheduling optimization strategy is iteratively optimized through a distributed reinforcement learning algorithm to obtain an optimized scheduling strategy, including: Applying the hardware operating parameter adjustment scheme to each hardware component of the server, adjusting the operating parameters of the hardware component, and obtaining the adjusted operating status of the hardware component; Based on the adjusted operating status of the hardware components, real-time sampling is performed on the operating parameters of the server at different time intervals to generate multi-dimensional real-time feedback data, wherein the operating parameters at least include: power consumption parameters, temperature parameters and load parameters; Performing feature extraction and dimensionality reduction processing on the real-time feedback data to obtain a feedback feature vector set, and performing abnormality detection on the hardware operation state based on the feedback feature vector set, and marking abnormal feedback feature vectors; The feedback feature vector set is input into the multi-agent architecture in the distributed reinforcement learning algorithm, and each agent independently evaluates and improves the current scheduling optimization strategy to generate its own scheduling improvement plan; Sharing the experience of each agent's scheduling improvement plan through a multi-agent experience pool, and optimizing the scheduling improvement plans in the experience pool using a genetic algorithm to obtain several candidate optimization strategies; Based on the candidate optimization strategy, the policy gradient method in the distributed reinforcement learning algorithm is used to iteratively solve the scheduling optimization strategy, and the parameters are integrated with the best solution in the candidate optimization strategy to obtain the final optimized scheduling strategy.

5. The energy consumption optimization method according to claim 1, characterized in that: Based on the optimized scheduling strategy, the server operation status is adaptively regulated to generate a target energy consumption optimization plan, which is verified by multi-dimensional performance indicators to obtain a target energy consumption optimization configuration plan for the server, including: Based on the optimized scheduling strategy, the operating data of each hardware component of the server under different load conditions is collected in real time to generate a server operating status data set, wherein the operating data includes at least any one of the following: CPU frequency, memory usage, I / O operation speed and temperature data; Normalizing the server operation status data set to eliminate the dimensional differences between different hardware parameters, and performing feature dimension reduction on the normalized data through principal component analysis to extract the main operation feature dimensions and generate a reduced-dimensional operation feature set; Based on the operation feature set, the hardware operation state of the server is controlled in real time by an adaptive control module. The adaptive control module uses the optimized scheduling strategy and combines the real-time changes of each dimension in the operation feature set to gradually adjust the operation parameters of the server and generate a target energy consumption optimization plan; Applying the target energy consumption optimization scheme to the hardware components of the server, adjusting the hardware operating parameters to cope with different load conditions, and recording the operating status data after each adjustment to generate an operating status data set after the optimization application; Based on the running status data set after the optimization application, the running effect of the server under the target energy consumption optimization scheme is verified to generate a verification result set, wherein the verification process is performed by calculating multi-dimensional performance indicators, and the multi-dimensional performance indicators at least include: system response time, power consumption, temperature change rate and throughput indicators; A multi-dimensional data comprehensive evaluation is performed on the verification result set, and the evaluation results of each performance indicator are weighted and integrated using a fuzzy comprehensive evaluation method to generate an energy consumption optimization evaluation score for the target server, and the target energy consumption optimization plan with the highest score is selected as the final target server energy consumption optimization configuration plan.

6. A server energy consumption optimization device, characterized in that: include: A monitoring unit, used to monitor the energy consumption of the server under different load states by using super-resolution data acquisition technology to obtain an initial energy consumption data set and a load behavior data set; A first generating unit is used to pre-process the association result of the initial energy consumption data set and the load behavior data set, standardize the energy consumption data and behavior data under each load state, and generate standardized energy consumption data and standardized load behavior data; input the standardized energy consumption data and standardized load behavior data into a contrastive learning network, construct a similarity measurement model through the contrastive learning network, perform feature extraction on the standardized energy consumption data and the standardized load behavior data under different load states, and obtain an initial energy consumption feature vector and an initial load behavior feature vector; perform embedding mapping processing on the initial energy consumption feature vector and the initial load behavior feature vector, map them to the same high-dimensional vector space, and obtain a feature embedding vector; and, performing cluster analysis on the feature embedding vectors, clustering the energy consumption characteristics of similar loads based on a clustering algorithm, and obtaining embedding representations of multiple load categories, each embedding representation corresponding to a specific load category; The second generation unit is used to establish a generative adversarial network model based on the embedded representation of the multiple load categories, and the generative adversarial network includes a generator and a discriminator. The generator is used to generate energy consumption prediction data according to the input load category embedded representation, and the discriminator is used to judge the difference between the generated data and the actual energy consumption data; the embedded representation of the load category is input into the generator, and the generator performs feature conversion on the input data through a multi-layer fully connected neural network to obtain energy consumption prediction data. The generator continuously optimizes its parameters during the training process to make the generated energy consumption data closer to the actual data distribution; the energy consumption prediction data and the actual energy consumption data set are simultaneously input into the discriminator, and the discriminator performs multi-layer convolution to obtain energy consumption prediction data. The neural network extracts and classifies the input data to determine whether it is real data and evaluates the generation capability of the generator, wherein the real energy consumption data is the initial energy consumption data set; in the training process of the generative adversarial network, through the adversarial process between the generator and the discriminator, the generator gradually improves the authenticity of the generated data, and the discriminator gradually enhances the ability to distinguish between the generated data and the real data, until the similarity between the energy consumption prediction data generated by the generator and the real energy consumption data meets the preset conditions, and a trained generative adversarial network model is obtained; and based on the trained generative adversarial network model, the embedded representations of multiple load categories are input to generate corresponding energy consumption prediction data; An adjustment unit, configured to establish a dynamic scheduling optimization strategy based on the energy consumption prediction data by using a Markov decision process, and to adjust the hardware operating parameters in combination with a particle swarm optimization algorithm to obtain a hardware operating parameter adjustment plan; An optimization unit, configured to apply the hardware operation parameter adjustment scheme to each hardware component of the server, generate real-time feedback data, and iteratively optimize the scheduling optimization strategy through a distributed reinforcement learning algorithm based on the real-time feedback data to obtain an optimized scheduling strategy; The acquisition unit is used to adaptively control the server operation status based on the optimized scheduling strategy, generate a target energy consumption optimization plan, and verify it through multi-dimensional performance indicators to obtain the target energy consumption optimization configuration plan of the server.

7. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images