Server management method and device, storage medium and program product

Abstract

The embodiment of the invention provides a server management method and device, a storage medium and a program product, and relates to the technical field of power regulation, and the method comprises the steps: obtaining a first power set of a server in a first time period, the first power set being used for indicating the overall power of the server and the power of each part of the server; inputting the first power set into a power prediction model to obtain a second power set, the second power set being used for indicating the power of the server and each component of the server in a second time period; and adjusting the working state of the server under the condition that the second power set meets a preset condition. By adopting the scheme, the technical problem that the accuracy of the current server management method under the dynamic load is relatively low in the related technology is solved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of power regulation technology, and in particular, to a server management method and device, a storage medium, and a program product. Background Art

[0002] The Power Supply Unit (PSU) is a core component of the server, responsible for converting alternating current (AC) into direct current (DC) required by the server. The PSU must provide stable and reliable power to ensure that the server and its components such as the motherboard, processor, memory, and hard disk can operate normally.

[0003] Predicting the power of a server can achieve efficient energy management. However, in the related art, the power prediction method is not accurate enough when facing the dynamic load fluctuation of the server, and cannot achieve dynamic and accurate power prediction, and thus cannot achieve effective energy consumption management of the server.

[0004] In the related technology, the current server management method has a low accuracy rate under dynamic load, and no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present application provide a server management method and device, a storage medium, and a program product to at least solve the technical problem in the related art that the current server management method has low accuracy under dynamic load.

[0006] According to one embodiment of the present application, a server management method is provided, including: obtaining a first power set of a server in a first time period, wherein the first power set is used to indicate the overall power of the server and the power of each component of the server; inputting the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and each component of the server in a second time period; and adjusting the working state of the server when the second power set meets preset conditions.

[0007] In an exemplary embodiment, the inputting the first power set into the power prediction model to obtain the second power set includes: obtaining a first power feature using the first power set; and inputting the first power feature into the power prediction model to obtain the second power set.

[0008] In an exemplary embodiment, the above-mentioned obtaining the first power characteristic by using the first power set includes at least one of the following: determining the mean of the first power set; determining the standard deviation of the first power set; and determining the frequency domain characteristic of the first power set.

[0009] In an exemplary embodiment, the above-mentioned inputting the first power characteristic into the power prediction model to obtain the second power set includes at least one of the following: inputting the utilization, clock frequency, and temperature of the first component of the server into the power prediction model to obtain the second power set; inputting the access frequency, usage rate, and type of the second component of the server into the power prediction model to obtain the second power set; inputting the read and write frequency, access mode, and type of the third component of the server into the power prediction model to obtain the second power set; and inputting the transmission frequency, bandwidth utilization, and data size of the fourth component of the server into the power prediction model to obtain the second power set.

[0010] In an exemplary embodiment, the above-mentioned inputting the first power set into the power prediction model to obtain the second power set includes: inputting the first power set and the second power set into the first power prediction model to obtain a first prediction result; inputting the first power set and the second power set into the second power prediction model to obtain a second prediction result; and weightedly fusing the first prediction result and the second prediction result to obtain a third power set.

[0011] In an exemplary embodiment, the training process of the above-mentioned power prediction model includes: obtaining a reference power set of the server within a third time period, and determining a reference power characteristic corresponding to the reference power set; obtaining initialization parameters of the power prediction model; using the reference power characteristic to train the power prediction model to update the parameters of the power prediction model, and obtaining a trained power prediction model when the training conditions are met.

[0012] In an exemplary embodiment, the training process of the above-mentioned first power prediction model includes: performing T random sampling on the reference power feature, and in each random sampling process, performing the following operations: collecting n samples from the reference power feature to obtain a sampling set containing n samples, wherein T is an integer greater than 0, and n is the same as the number of features contained in the reference power feature; establishing T power prediction sub-models, and corresponding the T sampling sets to the T power prediction sub-models one by one; using the sampling sets to train the corresponding power prediction sub-models, and the training process of the power prediction sub-models includes: when currently in the first node layer, randomly determining the first feature set from the reference power feature according to the feature number, wherein the number of feature types of the power feature in the feature set is the same as the feature number; establishing candidate nodes of the second node layer according to the first feature set, and evaluating the candidate nodes using the first feature set to obtain evaluation results of the candidate nodes of the second node layer; determining the nodes of the second node layer according to the evaluation results of the candidate nodes of the second node layer; and comparing the sampling set with the nodes of the second node layer according to the node number of the second node layer. Split to obtain the sampling subset corresponding to the nodes of the second node layer; when currently in the p-th node layer, according to the feature number, randomly determine the q-th feature set from the sampling subset corresponding to the nodes of the p-th node layer; establish the candidate nodes of the p+1-th node layer according to the q-th feature set, and use the q-th feature set to evaluate the candidate nodes of the p+1-th node layer to obtain the evaluation results of the candidate nodes of the p+1-th node layer; determine the nodes of the p+1-th node layer according to the evaluation results of the candidate nodes of the p+1-th node layer; The sampling subset corresponding to the nodes of the p-th node layer is split according to the nodes of the p+1-th node layer to obtain the sampling subset corresponding to the nodes of the p+1-th node layer, wherein p is an integer greater than 1, and q is an integer greater than 1; when the number of reference power features contained in the sampling subset corresponding to the nodes of the p+1-th node layer is less than or equal to the preset data number or p is greater than or equal to the preset number of layers, the model composed of the current p+1 node layers is determined as the power prediction sub-model; T power prediction sub-models are integrated to obtain the power prediction model.

[0013] In an exemplary embodiment, the above-mentioned acquisition of the first power set of the server in the first time period includes: using a first communication protocol to obtain the input and output power of the server in the first time period; obtaining current data of current sensors corresponding to components of the server, and determining the power corresponding to the components of the server based on the current data.

[0014] In an exemplary embodiment, the above-mentioned adjusting the working state of the server includes: determining the output voltage and current corresponding to the server and components of the server according to the second power set; and determining the duty cycle of the server according to the second power set.

[0015] In an exemplary embodiment, after obtaining the trained power prediction model as mentioned above, it also includes: obtaining a third power set of the server in a fourth time period, and determining a second power characteristic corresponding to the third power set; using the third power set and the second power characteristic to update the parameters of the power prediction model to obtain an updated power prediction model.

[0016] According to another embodiment of the present application, a server management device is provided, including: a first acquisition module, used to acquire a first power set of the server in a first time period, wherein the first power set is used to indicate the overall power of the server and the power of each component of the server; a power prediction module, used to input the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and each component of the server in a second time period; and a state adjustment module, used to adjust the working state of the server when the second power set meets preset conditions.

[0017] According to another embodiment of the present application, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when run.

[0018] According to another embodiment of the present application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0019] Through this application, a first power set of a server in a first time period is obtained, and the first power set may be only the overall power of the server and the power of each component of the server; the first power set is input into a power prediction model to obtain a second power set, and the second power set may indicate the power of the server and each component of the server in the second time period; when the second power aggregation meets the preset conditions, the working state of the server may be adjusted. The overall power of the server and the power of each component in the server obtained in the first time period may be used to predict the overall power of the server and the power of each component in the second time period, and the working state of the server and each component therein may be adjusted according to the predicted power to achieve dynamic load management and energy consumption regulation, thereby enabling the server to adapt to different load conditions. The technical problem in the related art that the current server management method has low accuracy under dynamic load is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a hardware structure block diagram of an optional server management method according to an embodiment of the present application;

[0021] Figure 2 is a flow chart of an optional server management method according to an embodiment of the present application;

[0022] Figure 3 is a schematic diagram of an optional server management method according to an embodiment of the present application;

[0023] Figure 4 is a schematic diagram of another optional server management method according to an embodiment of the present application;

[0024] Figure 5 is a schematic diagram of another optional server management method according to an embodiment of the present application;

[0025] Figure 6 is a schematic diagram of another optional server management method according to an embodiment of the present application;

[0026] Figure 7 It is a structural block diagram of an optional server management device according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0029] According to one aspect of the embodiments of the present application, a server management method is provided. As an optional implementation, the server management method can be applied to, but is not limited to, Figure 1 The server management system in the hardware environment shown in the figure may include, but is not limited to, a power supply unit 102, a baseboard management controller 104, an operating system 106, a central processing unit 108, a memory 110, a hard disk 112, a network card 114, and the like.

[0030] Optionally, in this embodiment, the power supply unit (PSU) 102 is a core component of the server, responsible for converting alternating current (AC) into direct current (DC) required by the server. The power supply unit 102 must provide a stable and reliable power supply to ensure that the server and its components such as the motherboard, processor, memory, and hard disk can operate normally.

[0031] Optionally, in this embodiment, the baseboard management controller (Baseboard Management Controller, BMC) 104 is a dedicated microcontroller in the server for managing and monitoring the hardware status of the server. It can monitor the temperature, voltage and other states of the server's central processor 108, memory 110, hard disk 112, network card 114 and other components to ensure that the system is in a healthy state. The baseboard management controller 104 can communicate with the power supply unit 102 of the server through the SMBus or PMBus interface using the I2C protocol, and read the input and output power consumption data of the power supply unit 102 in real time. PMBus is an extension of SMBus specifically used for power management. It defines a set of commands and data structures for monitoring and controlling power supply devices.

[0032] Optionally, in this embodiment, the operating system 106 can cooperate with the baseboard management controller 104 to use management tools, etc. to manage and monitor the power status of the server and components, and perform power management operations such as power on, power off, and restart. The operating system 106 can use the server power redundancy function and intelligent power distribution strategy to balance the load among multiple power modules, extend the service life of the power supply, and improve the efficiency of power management. The server operating system 106 allows the administrator to automatically adjust the power management strategy according to the load of the server, for example, selecting the performance mode when the load is high and the energy saving mode when the load is low.

[0033] The server's CPU 108, memory 110, hard disk 112, network card 114, etc. may be components of the server. They are key components that form the hardware foundation of the server. The management and optimization of these components by the server operating system (OS) is crucial to ensuring the performance, stability, and efficiency of the server.

[0034] Optionally, in this embodiment, the central processing unit (CPU) 108 is responsible for executing program instructions, processing data, and controlling other hardware components. The operating system 106 can manage the resources of the central processing unit 108 through a scheduling algorithm, assign tasks, handle interrupts, and optimize performance. For example, the operating system 106 can adjust the frequency and voltage of the central processing unit 108 to save energy, or balance the load to prevent overheating.

[0035] Optionally, in this embodiment, the memory 110 is a fast storage medium used by the server to store data and instructions currently in use or to be used. The operating system 106 manages memory allocation to ensure that applications and system processes have sufficient resources to run. It also handles caches and buffers to increase data access speed and uses swap space to manage data in physical memory.

[0036] Optionally, in this embodiment, hard disk 112 is used by the server to permanently store the server's data, operating system, application programs, and user files. Operating system 106 manages data read and write operations, file systems, and disk partitioning and formatting. It is also responsible for disk maintenance tasks such as defragmentation and error checking.

[0037] Optionally, in this embodiment, the network card 114 enables the server to connect to the network for data transmission and communication. The operating system 106 controls network communications by configuring the network protocol stack and managing network drivers. It can also be responsible for allocating IP addresses, processing routing and firewall rules, and monitoring network traffic.

[0038] In an optional implementation, the server management method of this embodiment can be executed in the operating system 106, and the specific process is as follows: obtain a first power set of the server in a first time period, and the first power set indicates the overall power of the server and the power of each component of the server. The overall power of the server can be obtained through the power supply unit 102, and the power of the server's central processor 108, memory 110, hard disk 112, network card 114 and other components can be obtained. The first power set is input into the power prediction model to obtain a second power set, and the second power set can be a predicted value of the power of the server as a whole and each component of the server in the second time period. When the second power set meets the preset conditions, adjust the working state of the server.

[0039] Through this application, a first power set of a server in a first time period is obtained, and the first power set may be only the overall power of the server and the power of each component of the server; the first power set is input into a power prediction model to obtain a second power set, and the second power set may indicate the power of the server and each component of the server in the second time period; when the second power aggregation meets the preset conditions, the working state of the server may be adjusted. The overall power of the server and the power of each component in the server obtained in the first time period may be used to predict the overall power of the server and the power of each component in the second time period, and the working state of the server and each component therein may be adjusted according to the predicted power to achieve dynamic load management and energy consumption regulation, thereby enabling the server to adapt to different load conditions. The technical problem in the related art that the current server management method has low accuracy under dynamic load is solved.

[0040] In this embodiment, a server management method is provided. Figure 2 is a flow chart of an optional server management method according to an embodiment of the present application, such as Figure 2 As shown, the server management method includes:

[0041] Step S202, obtaining a first power set of the server in a first time period, wherein the first power set is used to indicate the overall power of the server and the power of each component of the server;

[0042] It should be noted that, as described above, the overall power in the server can be obtained by reading data from the power supply unit of the server to obtain the overall data of the server.

[0043] In an optional embodiment, the above-mentioned acquisition of the first power set of the server in the first time period includes: obtaining the input and output power of the server in the first time period using a first communication protocol; obtaining current data of current sensors corresponding to components of the server, and determining the power corresponding to the components of the server based on the current data.

[0044] Figure 3 is a schematic diagram of an optional server management method according to an embodiment of the present application, such as Figure 3 As shown, in addition to the power supply unit 102, baseboard management controller 104, operating system 106, central processing unit 108, memory 110, hard disk 112, network card 114, etc. included in the server, current sensors ( Figure 3 The current sensor 1-n in the example detects the current consumption through a sampling resistor. The output of the current sensor is an analog signal, which needs to be converted into a digital signal through an analog-to-digital converter 302 (ADC) so that the processor can read and process the data. The converted digital signal can be transmitted via the I2C protocol. The baseboard management controller 104 can collect and process the data from the analog-to-digital converter 302. The baseboard management controller 104 can be an independent management controller with its own processor and memory, and can operate independently of the server's operating system.

[0045] In an optional implementation, after the current sensor measures the current, the power consumption of each component can be monitored, and the power can be calculated by combining the current and voltage information. When the current sensor measures the current of each component, the current flows through the resistor, and a voltage drop is generated on the resistor, and this voltage drop is proportional to the current. The sensor detects this voltage drop and converts it into a digital signal through the ADC.

[0046] Through the above-mentioned implementation mode of the present application, the overall power of the server and the power of each component can be read, so as to realize effective monitoring and management of the power consumption of each component in the server, so as to identify and optimize components or systems with low energy efficiency.

[0047] Step S204, inputting the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and each component of the server in a second time period;

[0048] It should be noted that after determining the first power set of the server as a whole and each component, the first power set can be input into the power prediction model to predict the second power set of the server as a whole and each component in the second time period in the future. The power prediction model can be a prediction model such as a random forest model, or a neural network model can be used for prediction.

[0049] In an optional implementation, the power prediction model may be a plurality of models trained based on different data, or may be a single model. In one example, the power prediction model may be trained using data corresponding to the server as a whole, and the power prediction model corresponding to each component of the server may be trained using data corresponding to each component of the server. In another example, a power prediction model may be trained using data corresponding to the server as a whole and each component to perform power prediction for the server or each component.

[0050] Step S206: When the second power set meets the preset condition, adjust the working state of the server.

[0051] It should be noted that the preset condition may be pre-set, and the preset condition may be set to a change in power of the entire server or each component of the server, so that when the power of the entire server or each component of the server changes, the working state of the server is adjusted according to the second power set.

[0052] Through this application, a first power set of a server in a first time period is obtained, and the first power set may be only the overall power of the server and the power of each component of the server; the first power set is input into a power prediction model to obtain a second power set, and the second power set may indicate the power of the server and each component of the server in the second time period; when the second power aggregation meets the preset conditions, the working state of the server may be adjusted. The overall power of the server and the power of each component in the server obtained in the first time period may be used to predict the overall power of the server and the power of each component in the second time period, and the working state of the server and each component therein may be adjusted according to the predicted power to achieve dynamic load management and energy consumption regulation, thereby enabling the server to adapt to different load conditions. The technical problem in the related art that the current server management method has low accuracy under dynamic load is solved.

[0053] In an optional embodiment, the step of inputting the first power set into a power prediction model to obtain a second power set includes: obtaining a first power characteristic using the first power set; and inputting the first power characteristic into a power prediction model to obtain the second power set.

[0054] In an optional implementation, a trained power prediction model may be obtained, and the first power set may be processed to obtain a first power feature that may be input into the model. The first power feature may be data to be input into the model, such as feature information obtained after processing the first power set.

[0055] In an optional implementation, the first power feature may be obtained by extracting features from the first power set, or by processing data in the first power set. Data corresponding to the components in the first power set may also be input as the first power feature into a power prediction model corresponding to the components.

[0056] In an optional embodiment, the above-mentioned obtaining the first power characteristic using the first power set includes at least one of the following: determining the mean of the first power set; determining the standard deviation of the first power set; and determining the frequency domain characteristic of the first power set.

[0057] In an optional implementation, the first power feature may be obtained after processing the first power set, and the first power set may include power and load data of the server as a whole and each component of the server in the first time period. It should be noted that the power of the server may be the electric energy consumed per unit time. The load may be the usage of resources of each component of the server. The load may be the number of tasks and workload being processed by the server. For example, the CPU usage, memory usage, disk read and write rate, network bandwidth usage, etc. may be used to measure the load of the server.

[0058] In an optional embodiment, the first power set may be Among them, P(t) can represent the power data of the server, and L(t) can represent the load data of the server.

[0059] In an optional implementation, feature extraction may be performed on the first power set, such as obtaining the mean of the first power set. Standard Deviation Or perform Fourier transform on the first power set To obtain the frequency spectrum characteristics of the first power set.

[0060] Through the above-mentioned implementation mode of the present application, the first power set can be subjected to data preprocessing, and its features can be extracted to be input into the power prediction model. The time characteristics of the first power set of the server can be extracted through time domain analysis technology to provide rich input features for subsequent power prediction.

[0061] In an optional embodiment, the above-mentioned inputting the first power characteristic into the power prediction model to obtain the second power set includes at least one of the following: inputting the utilization, clock frequency, and temperature of the first component of the server into the power prediction model to obtain the second power set; inputting the access frequency, usage rate, and type of the second component of the server into the power prediction model to obtain the second power set; inputting the read and write frequency, access mode, and type of the third component of the server into the power prediction model to obtain the second power set; inputting the transmission frequency, bandwidth utilization, and data size of the fourth component of the server into the power prediction model to obtain the second power set.

[0062] In an optional implementation, power prediction can be performed for each component in the server. As mentioned above, each component can have a corresponding power prediction model, or the same power prediction model can be used. The features input into the power prediction model can be different for different components.

[0063] In an optional embodiment, the power consumption of the CPU of the server presents a nonlinear characteristic. When the power of the CPU is predicted, the utilization, clock frequency, temperature, and instruction set architecture of the CPU can be determined according to the first power set. The CPU utilization is used to indicate the percentage of the CPU used to execute tasks within a specific time. High utilization means that the CPU is busy processing tasks, which usually leads to higher power consumption. The clock frequency, also known as the main frequency, is the reference frequency of the internal operation of the CPU. The higher the clock frequency, the faster the CPU operation speed in theory, but it may also lead to higher power consumption. The temperature of the CPU is an important indicator of its working state. Excessive temperature may cause the CPU to reduce the frequency to prevent overheating, thereby affecting performance and power consumption. The instruction set architecture is the part of the computer architecture related to program design, which defines the set of all instructions that the CPU can execute. Different instruction set architectures, such as CISC (complex instruction set computer) and RISC (reduced instruction set computer), will affect the power consumption and performance of the CPU. For example, the RISC architecture usually achieves low power consumption with a simplified instruction set and efficient execution.

[0064] In an optional embodiment, when predicting the power consumption of the memory of the server, the access frequency, memory usage, idle state, memory type, etc. of the memory can be determined according to the first power set. Memory usage refers to the proportion of memory resources in the server occupied by currently running programs and processes. High memory usage usually means that more memory resources are utilized, which may lead to higher power consumption. For example, when the memory usage is close to 100%, the server may need more energy to process data, resulting in increased power consumption. Access frequency refers to how frequently the memory is accessed. Frequent memory access may lead to higher power consumption because the memory unit needs to be constantly activated and refreshed. Idle time refers to the time when the memory is not accessed within a certain period of time. If the idle time of the memory is long, it may enter a low power mode, thereby reducing energy consumption. The type of memory also has a significant impact on power consumption. For example, DDR5 memory has a lower operating voltage than DDR4 memory while providing higher speed and greater memory density, thereby reducing power consumption.

[0065] In an optional embodiment, when predicting the power consumption of the server's disk, the operating frequency, access mode, idle state, and rotation speed of the disk can be determined based on the first power set. The read and write rate of the hard disk directly affects the power consumption. A high read and write rate usually means that the hard disk requires more energy to perform operations, thereby increasing power consumption. The access mode affects the power consumption of the hard disk. Sequential access usually involves continuous reading and writing of large blocks of data, while random access involves reading and writing of small blocks of data. In random access mode, the power consumption of the hard disk is usually higher because the head needs to move frequently to different locations to read data, increasing energy consumption. The longer the hard disk is idle, the lower its power consumption. When the hard disk is not performing read and write operations, it will enter a low power mode to reduce energy consumption. The type of hard disk also affects power consumption. Traditional mechanical hard disks (HDDs) usually have higher power consumption than solid-state drives (SSDs) due to the presence of mechanical moving parts.

[0066] In an optional embodiment, when the power consumption of the network card of the server is predicted, the network load, transmission rate, energy-saving mode, etc. of the network card device can be determined according to the first power set. Network bandwidth utilization refers to the ratio of the actual data transmission volume to the maximum theoretical transmission volume in the network. High bandwidth utilization usually means that more data is being transmitted through the network device, which may lead to increased power consumption of the device. The size of the data packet affects the transmission efficiency and power consumption of the network. Larger data packets can reduce the proportion of header information and improve transmission efficiency, which may reduce the power consumption per unit data. The transmission rate, also known as the data rate or bit rate, refers to the amount of data transmitted per unit time. An increase in transmission rate is usually accompanied by an increase in power consumption because the device requires more energy to process and transmit more data. Energy-Efficient Ethernet (EEE) is a standard that allows network devices to enter a low-power mode when under low load. When the network load is light, EEE can put the device into an energy-saving state, thereby significantly reducing power consumption.

[0067] In an optional implementation, for different components in the server, the corresponding characteristics of each component can be input into the corresponding power prediction model, or the corresponding characteristics of each component can be input into a universal power prediction model to predict the power of each component to obtain a second power set.

[0068] In an optional implementation, after obtaining the power corresponding to each component of the server, the power prediction results of each component can be combined, such as weighted averaging the power prediction results of each component obtained above, to obtain the predicted power of the server as a whole. The predicted power obtained above and the overall power of the server can be determined as the second power set.

[0069] Through the above implementation of the present application, for different sub-components in the server, power prediction can be performed using their corresponding characteristics, so as to achieve more accurate power prediction of the components in the server and realize dynamic adjustment of the power consumption of the server.

[0070] In an optional embodiment, the above-mentioned inputting the first power set into the power prediction model to obtain the second power set includes: inputting the first power set and the second power set into the first power prediction model to obtain a first prediction result; inputting the first power set and the second power set into the second power prediction model to obtain a second prediction result; and weightedly fusing the first prediction result and the second prediction result to obtain a third power set.

[0071] In an optional embodiment, the training process of the above-mentioned power prediction model includes: obtaining a reference power set of the server within a third time period, and determining a reference power characteristic corresponding to the reference power set; obtaining initialization parameters of the power prediction model; using the reference power characteristic to train the power prediction model to update the parameters of the power prediction model, and obtaining a trained power prediction model when the training conditions are met.

[0072] In an optional implementation, after obtaining the training data (reference power set), the parameters of the power prediction model can be initialized, and then the power prediction model is trained using the reference power set to update its parameters. Meeting the preset condition can be that the preset indicators such as the accuracy, recall rate, and mean square error of the data prediction model reach a preset threshold range. When the preset conditions are met, the model training is considered to be completed.

[0073] In an optional implementation, the first power prediction model and the second power prediction model may be trained separately, the first power prediction model may be a random forest model, and the second power prediction model may be a neural network model. The training methods of the first power prediction model and the second power prediction model will be described in detail below.

[0074] In an optional embodiment, after obtaining the reference power data, the power characteristics corresponding to each of them can be determined. Then the data is cleaned, and missing value processing can be performed: identifying and processing missing values ​​in the reference power set and the corresponding power characteristics, possibly by deletion, filling (such as using the mean, median, mode or model-based prediction, etc.). Outlier detection can be performed: identifying outliers in the reference power set and the corresponding power characteristics and deciding how to handle them, such as deleting, replacing or transforming. Deduplication: Check and delete duplicate records in the reference power set and the corresponding power characteristics to avoid deviations in the analysis. Error correction: Correct entry errors in the reference power set and the corresponding power characteristics or errors in the data extraction process.

[0075] Optionally, after data cleaning, labeling, and feature selection, the reference power set is subjected to an updated reference power set as a data set for the power prediction model training phase.

[0076] Figure 4 is a schematic diagram of another optional server management method according to an embodiment of the present application; Figure 4 As shown, the reference power feature set 402 (such as Figure 4 D) is randomly divided into M power feature subsets 404 (such as Figure 4 D 1 To D M). M power feature subsets can be traversed, and the following operations can be performed during each traversal: all power feature subsets except the i-th power feature subset in the M power feature subsets are used as training sets to train the power prediction model, and the i-th power feature subset is input into the power prediction model to obtain the i-th evaluation function result of the power prediction model. The evaluation function result obtained in each traversal can be averaged and other operations can be performed to evaluate the performance of the model. The above can be used to test the hyperparameter combination in the first power prediction model, such as the number of features, the number of node layers, the access data threshold and other parameters of the first power prediction model. The role of specific parameters in the training process will be described in detail below and will not be repeated here.

[0077] In an optional embodiment, when the evaluation function result does not meet the preset conditions (such as the mean of the evaluation function result is less than the preset result), the parameters of the first power prediction model can be changed and trained again to determine the above-mentioned hyperparameters of the first power prediction model.

[0078] Through the above implementation of the present application, the reference power feature set can be randomly split into a training set and a validation set, and the performance of the first power prediction model can be trained and evaluated multiple times to determine the hyperparameters of the first power prediction model, and the optimal parameter configuration can be selected to improve the performance of the first power prediction model. In this way, the prediction results of the model are more accurate during the application process, and the performance of the system is improved.

[0079] In an optional embodiment, the training process of the above-mentioned first power prediction model includes: performing T random sampling on the reference power feature, and in each random sampling process, performing the following operations: collecting n samples from the reference power feature to obtain a sampling set containing n samples, wherein T is an integer greater than 0, and n is the same as the number of features contained in the reference power feature; establishing T power prediction sub-models, and corresponding the T sampling sets to the T power prediction sub-models one by one; using the sampling sets to train the corresponding power prediction sub-models, and the training process of the power prediction sub-models includes: when currently in the first node layer, randomly determining the first feature set from the reference power feature according to the feature number, wherein the number of feature types of the power features in the feature set is the same as the feature number; establishing candidate nodes of the second node layer according to the first feature set, and evaluating the candidate nodes using the first feature set to obtain evaluation results of the candidate nodes of the second node layer; determining the nodes of the second node layer according to the evaluation results of the candidate nodes of the second node layer; splitting the sampling set into nodes according to the nodes of the second node layer points, and obtains the sampling subset corresponding to the nodes of the second node layer; when currently in the p-th node layer, according to the feature number, randomly determines the q-th feature set from the sampling subset corresponding to the nodes of the p-th node layer; establishes the candidate nodes of the p+1-th node layer according to the q-th feature set, and uses the q-th feature set to evaluate the candidate nodes of the p+1-th node layer to obtain the evaluation results of the candidate nodes of the p+1-th node layer; determines the nodes of the p+1-th node layer according to the evaluation results of the candidate nodes of the p+1-th node layer; The sampling subset corresponding to the nodes of the p-th node layer is split according to the nodes of the p+1-th node layer to obtain the sampling subset corresponding to the nodes of the p+1-th node layer, wherein p is an integer greater than 1, and q is an integer greater than 1; when the number of reference power features contained in the sampling subset corresponding to the nodes of the p+1-th node layer is less than or equal to the preset data number or p is greater than or equal to the preset number of layers, the model composed of the current p+1 node layers is determined as the power prediction sub-model; T power prediction sub-models are integrated to obtain the power prediction model.

[0080] In an optional implementation, the reference power feature set can be used as a training set for training, and the training set can be randomly sampled T times. In each random sampling process, the following operations are performed: the training set is sampled n times to obtain a sampling set containing n samples. In the process of determining the sampling set, a sample can be randomly taken out and put into the sampling set, and then the sample is put back into the initial data set, so that the sample may still be selected in the next sampling.

[0081] In an optional embodiment, a power prediction sub-model may be established. Figure 5 is a schematic diagram of another optional server management method according to an embodiment of the present application; Figure 5 As shown, after obtaining T sampling sets using the reference power feature set, the sampling sets can be used to train the corresponding power prediction sub-models. After the T power prediction sub-models are trained, they are integrated to obtain the first trained power prediction model.

[0082] In an optional embodiment, there may be one node (root node) in the first node layer, and the nodes in the second node layer may be obtained by splitting the root node. Determining that the nodes in the second node layer may be split using the type of power features. As previously described, the number of features of the model may be determined, which defines the number of features to be considered when looking for the best split. It may be an integer or a proportion of the total number of features. If the number of features is set to an integer, then at each splitting node, the number of features specified by this integer will be randomly selected for evaluation. If the number of features is set to a floating point number, then it indicates that a certain proportion of the total number of original features will be considered. For example, if the data set has 100 features, an eigenvalue of 0.3 means that 30 features will be considered at each split point.

[0083] In an optional implementation, in the process of determining the splitting node, the determination method can be changed according to the form of the data feature. For example, for continuous features, it is necessary to traverse all unique values ​​and calculate the evaluation value of each possible splitting point; for discrete features, each different category value is a potential splitting point. For each splitting point, the evaluation result after splitting is calculated.

[0084] In an optional implementation, the above process can be repeated for each node until the number of node layers reaches a preset number of layers, or the number of reference power features contained in the sampling subset corresponding to the node is less than or equal to the preset number of data, and the power prediction sub-model training can be completed. The preset number of layers and the preset number of data can be determined in advance as described above.

[0085] Through the above implementation of the present application, the first power prediction model can improve the prediction performance by integrating multiple sub-models, and each sub-model can be regarded as a nonlinear division of the data space of the power feature, which can capture the nonlinear relationship between the server power and the load. And the power consumption of the server is usually not linear, so the first power prediction model is more in line with the server power prediction to improve the accuracy of the power prediction.

[0086] In an optional implementation, the second power prediction model may be a neural network model, the first power prediction model may be run in parallel, the first power prediction model and the second power prediction model may be fused, the same power characteristics may be input into the first power prediction model and the second power prediction model to obtain two power output results, and the two power output results may be weighted averaged to obtain the final power prediction value. The formula is as follows:

[0087]

[0088] Among them, αα, β are fusion weights, and the sum is 1. is the output result of the first power prediction model, is the output result of the second power prediction model.

[0089] In an optional implementation, the first power prediction model and the second power prediction model can be run in series, and the output result of the first power prediction model can be input into the second power prediction model so that the output result of the first power prediction model is used as a preliminary result, and the preliminary result is input into the second power prediction model to obtain the output result of the second power prediction model. The output result of the second power prediction model can be used as the final power prediction result, or the preliminary result and the output result of the second power prediction model can be weighted averaged to obtain the final power prediction result.

[0090] In an optional embodiment, the second power prediction model may be a neural network model, which may be a long short-term memory network, and processes power characteristic data with time characteristics. It may include a long short-term memory network layer to capture the time dependency in the power characteristic data, and may process time-varying patterns in the input data, such as the CPU utilization, memory usage, disk I / O, etc. of the server at different time points. The output of each time step contains not only the information of the current time step, but also the historical information of all previous time steps.

[0091] In an optional embodiment, the second power prediction model may further include a fully connected layer, which may further process the output result of the long short-term memory network layer, such as combining, extracting and converting. An activation function may also be used to introduce nonlinearity to output a prediction result.

[0092] In an optional implementation, the second power prediction model training process may use the minimization of mean square error as a loss function for training. The formula may be as follows:

[0093]

[0094] Among them, P i is the actual power, is the output power of the second power prediction model.

[0095] In an optional implementation, after the power prediction model is trained, different indicators can be used to evaluate the power prediction model, such as mean square error, mean absolute error, determination coefficient, etc. The determination coefficient can indicate the degree to which the power prediction model explains the variability of the data. The calculation formula of the determination coefficient can be as follows:

[0096]

[0097] in, It can represent the average value of the actual power value. It should be noted that the determination coefficient approaching 1 can indicate that the power prediction model has a better fit.

[0098] In an optional embodiment, after the above-mentioned trained power prediction model is obtained, it also includes: obtaining a third power set of the server in a fourth time period, and determining a second power characteristic corresponding to the third power set; using the third power set and the second power characteristic to update the parameters of the power prediction model to obtain an updated power prediction model.

[0099] In an optional implementation, the power prediction model may be evaluated regularly. When the evaluation index is poor, the third power set for the fourth time period may be obtained, and the model may be updated regularly to maintain accuracy.

[0100] In an optional implementation, the above-mentioned adjusting the working state of the server includes: determining output voltages and currents corresponding to the server and components of the server according to the second power set; and determining a duty cycle of the server according to the second power set.

[0101] In an optional implementation, after predicting the power of the server and each component, the relationship between power and load can be determined, and the working state of the server can be adjusted. It should be noted that the working state of the server can be adjusted by adjusting the VRM power point, the BUCK duty cycle, the hard disk energy saving mode, the network card energy saving mode, etc.

[0102] In an optional embodiment, the VRM (Voltage Regulator Module) is a component on the server motherboard that is responsible for providing a stable voltage to the CPU. By dynamically adjusting the power point of the VRM, the power supply can be adjusted according to the real-time workload of the server to optimize the power conversion efficiency. The power prediction results can guide the adjustment strategy of the VRM. For example, if the prediction shows that the server load will decrease in the future, the VRM can reduce the supply voltage in advance, thereby reducing energy consumption.

[0103] In an optional embodiment, a BUCK converter is a DC-DC converter used to convert a higher input voltage to a lower output voltage. The duty cycle refers to the proportion of time that the switch is on in each cycle. By adjusting the duty cycle of the BUCK converter, the power output can be finely controlled to adapt to the current load requirements. When the load is low, reducing the duty cycle can reduce the switching frequency of the power module, thereby reducing energy loss.

[0104] In an optional embodiment, the energy-saving mode of the hard disk includes a sleep mode and a low power mode. These modes can reduce the power consumption of the hard disk when it is not frequently used. By analyzing the server load status using a power prediction model, the hard disk usage demand can be predicted and the hard disk working mode can be dynamically adjusted. For example, when the server load is low and the hard disk access frequency is not high, it can be switched to a sleep mode or a low power mode.

[0105] In an optional implementation paradigm, the network card (network interface card) can also reduce power consumption through energy-saving modes. These modes usually include shutting down unused network functions, reducing operating frequency, etc. Combined with the prediction results of the power prediction model, the working state of the network card can be dynamically adjusted according to the network load. For example, when the network traffic is low, the operating frequency of the network card can be reduced or it can be placed in energy-saving mode to reduce energy consumption.

[0106] Through the above-mentioned implementation mode of the present application, after determining the predicted power of the server as a whole and each component, the working state of each component can be adjusted according to the pre-stored results to dynamically adjust the working mode of each sub-component of the server, optimize the overall energy consumption of the server, and improve the performance of the server.

[0107] Figure 6 is a schematic diagram of another optional server management method according to an embodiment of the present application; Figure 6 As shown, the power data of the server and its components can be obtained, and the power prediction model can be trained using the obtained power data. The power prediction model can include a random forest model (first power prediction model) and a neural network model (second power prediction model). The power data can be feature extracted to obtain power features, and the model can be trained using the obtained features to obtain a trained power prediction model. After the model training is completed, the real-time power data of the server and its components can be obtained, and the data can be preprocessed. The processed data is input into the trained power prediction model to predict the future power of the server and its components, and the working state of the server is adjusted according to the obtained predicted power results to optimize the overall energy consumption of the server.

[0108] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, or by hardware, but in many cases the former is a better implementation method. Based on this understanding, the computer software product is stored in a storage medium (such as ROM / RAM, disk, CD), including several instructions for enabling a terminal device (which can be a computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0109] In this embodiment, a server management device is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0110] Figure 7 is a structural block diagram of an optional server management device according to an embodiment of the present application, such as Figure 7 As shown, the device comprises:

[0111] A first acquisition module 702 is used to acquire a first power set of the server in a first time period, wherein the first power set is used to indicate the overall power of the server and the power of each component of the server;

[0112] A power prediction module 704, configured to input the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and each component of the server in a second time period;

[0113] The state adjustment module 706 is used to adjust the working state of the server when the second power set meets the preset condition.

[0114] Through this application, a first power set of a server in a first time period is obtained, and the first power set may be only the overall power of the server and the power of each component of the server; the first power set is input into a power prediction model to obtain a second power set, and the second power set may indicate the power of the server and each component of the server in the second time period; when the second power aggregation meets the preset conditions, the working state of the server may be adjusted. The overall power of the server and the power of each component in the server obtained in the first time period may be used to predict the overall power of the server and the power of each component in the second time period, and the working state of the server and each component therein may be adjusted according to the predicted power to achieve dynamic load management and energy consumption regulation, thereby enabling the server to adapt to different load conditions. The technical problem in the related art that the current server management method has low accuracy under dynamic load is solved.

[0115] Optionally, the power prediction module 704 uses the first power set to obtain a first power feature; and inputs the first power feature into a power prediction model to obtain a second power set.

[0116] Optionally, the above-mentioned obtaining the first power characteristic by using the first power set includes at least one of the following: determining the mean of the first power set; determining the standard deviation of the first power set; and determining the frequency domain characteristic of the first power set.

[0117] Optionally, the power prediction module 704 is also used to: input the utilization rate, clock frequency, and temperature of the first component of the server into the power prediction model to obtain a second power set; input the access frequency, usage rate, and type of the second component of the server into the power prediction model to obtain a second power set; input the read and write frequency, access mode, and type of the third component of the server into the power prediction model to obtain a second power set; input the transmission frequency, bandwidth utilization, and data size of the fourth component of the server into the power prediction model to obtain a second power set.

[0118] Optionally, the above-mentioned power prediction module 704 is also used to: input the first power set and the second power set into the first power prediction model to obtain a first prediction result; input the first power set and the second power set into the second power prediction model to obtain a second prediction result; and perform weighted fusion on the first prediction result and the second prediction result to obtain a third power set.

[0119] Optionally, the training process of the above-mentioned power prediction model includes: obtaining a reference power set of the server within a third time period, and determining a reference power characteristic corresponding to the reference power set; obtaining initialization parameters of the power prediction model; using the reference power characteristics to train the power prediction model to update the parameters of the power prediction model, and obtaining a trained power prediction model when the training conditions are met.

[0120] Optionally, the training process of the above-mentioned first power prediction model includes: performing T random sampling on the reference power feature, and in each random sampling process, performing the following operations: collecting n samples from the reference power feature to obtain a sampling set containing n samples, wherein T is an integer greater than 0, and n is the same as the number of features contained in the reference power feature; establishing T power prediction sub-models, and corresponding the T sampling sets to the T power prediction sub-models one by one; using the sampling sets to train the corresponding power prediction sub-models, and the training process of the power prediction sub-models includes: when currently in the first node layer, randomly determining the first feature set from the reference power feature according to the feature number, wherein the number of feature types of the power features in the feature set is the same as the feature number; establishing candidate nodes of the second node layer according to the first feature set, and using the first feature set to evaluate the candidate nodes to obtain evaluation results of the candidate nodes of the second node layer; determining the nodes of the second node layer according to the evaluation results of the candidate nodes of the second node layer; splitting the sampling set according to the nodes of the second node layer, and obtaining to the sampling subset corresponding to the nodes of the second node layer; when currently in the p-th node layer, according to the feature number, randomly determine the q-th feature set from the sampling subset corresponding to the nodes of the p-th node layer; establish the candidate nodes of the p+1-th node layer according to the q-th feature set, and use the q-th feature set to evaluate the candidate nodes of the p+1-th node layer to obtain the evaluation results of the candidate nodes of the p+1-th node layer; determine the nodes of the p+1-th node layer according to the evaluation results of the candidate nodes of the p+1-th node layer; The sampling subsets corresponding to the nodes of the p node layers are split according to the nodes of the p+1 node layer to obtain the sampling subsets corresponding to the nodes of the p+1 node layer, wherein p is an integer greater than 1, and q is an integer greater than 1; when the number of reference power features contained in the sampling subsets corresponding to the nodes of the p+1 node layer is less than or equal to the preset data number or p is greater than or equal to the preset number of layers, the model composed of the current p+1 node layers is determined as the power prediction sub-model; T power prediction sub-models are integrated to obtain the power prediction model.

[0121] Optionally, the first acquisition module 702 is further used to: acquire input and output power of the server in a first time period using a first communication protocol; acquire current data of current sensors corresponding to components of the server, and determine the power corresponding to the components of the server based on the current data.

[0122] Optionally, the state adjustment module 706 is further configured to: determine output voltages and currents corresponding to the server and components of the server according to the second power set; and determine a duty cycle of the server according to the second power set.

[0123] Optionally, after obtaining the trained power prediction model, the method further includes: obtaining a third power set of the server in a fourth time period, and determining a second power characteristic corresponding to the third power set; updating parameters of the power prediction model using the third power set and the second power characteristic to obtain an updated power prediction model.

[0124] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0125] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when run.

[0126] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0127] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0128] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0129] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0130] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0131] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A server management method, characterized in that: The method comprises: Acquire a first power set of the server in a first time period, wherein the first power set is used to indicate the overall power of the server and the power of each component of the server; Inputting the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and each component of the server in a second time period; When the second power set meets a preset condition, the working state of the server is adjusted.

2. The method according to claim 1, characterized in that The step of inputting the first power set into a power prediction model to obtain a second power set includes: Obtaining a first power characteristic using the first power set; The first power characteristic is input into the power prediction model to obtain the second power set.

3. The method according to claim 2, characterized in that The obtaining the first power characteristic by using the first power set includes at least one of the following: determining a mean value of the first power set; determining a standard deviation of the first set of powers; A frequency domain characteristic of the first power set is determined.

4. The method according to claim 2, characterized in that: The step of inputting the first power feature into the power prediction model to obtain the second power set includes at least one of the following: Inputting the utilization rate, clock frequency, and temperature of the first component of the server into the power prediction model to obtain the second power set; Inputting the access frequency, usage rate, and type of the second component of the server into the power prediction model to obtain the second power set; Inputting the read / write frequency, access mode, and type of the third component of the server into the power prediction model to obtain the second power set; The transmission frequency, bandwidth utilization, and data size of the fourth component of the server are input into the power prediction model to obtain the second power set.

5. The method according to any one of claims 1 to 4, characterized in that: The step of inputting the first power set into a power prediction model to obtain a second power set includes: Inputting the first power set and the second power set into a first power prediction model to obtain a first prediction result; Inputting the first power set and the second power set into a second power prediction model to obtain a second prediction result; The first prediction result and the second prediction result are weightedly fused to obtain a third power set.

6. The method according to claim 5, characterized in that The training process of the power prediction model includes: Acquire a reference power set of the server in a third time period, and determine a reference power feature corresponding to the reference power set; Obtaining initialization parameters of the power prediction model; The power prediction model is trained using the reference power feature to update the parameters of the power prediction model, and when a training condition is met, the trained power prediction model is obtained.

7. The method according to claim 6, characterized in that The training process of the first power prediction model includes: The reference power feature is randomly sampled T times, and in each random sampling process, the following operations are performed: sampling the reference power feature n times to obtain a sampling set containing n samples, wherein T is an integer greater than 0, and n is the same as the number of features contained in the reference power feature; Establish T power prediction sub-models, and correspond the T sampling sets to the T power prediction sub-models one by one; use the sampling sets to train the corresponding power prediction sub-models, and the training process of the power prediction sub-models includes: In the case of being currently in the first node layer, according to the feature number, randomly determine the first feature set from the reference power features, wherein the number of feature types of the power features in the feature set is the same as the feature number; establish candidate nodes of the second node layer according to the first feature set, and evaluate the candidate nodes using the first feature set to obtain evaluation results of the candidate nodes of the second node layer; determine the nodes of the second node layer according to the evaluation results of the candidate nodes of the second node layer; split the sampling set according to the nodes of the second node layer to obtain sampling subsets corresponding to the nodes of the second node layer; In the case of being currently in the p-th node layer, according to the feature number, randomly determine the q-th feature set from the sampling subsets corresponding to the nodes of the p-th node layer; establish the candidate nodes of the p+1-th node layer according to the q-th feature set, and use the q-th feature set to evaluate the candidate nodes of the p+1-th node layer to obtain the evaluation results of the candidate nodes of the p+1-th node layer; determine the nodes of the p+1-th node layer according to the evaluation results of the candidate nodes of the p+1-th node layer; split the sampling subsets corresponding to the nodes of the p-th node layer according to the nodes of the p+1-th node layer to obtain the sampling subsets corresponding to the nodes of the p+1-th node layer, wherein p is an integer greater than 1, and q is an integer greater than 1; When the number of the reference power features contained in the sampling subset corresponding to the nodes of the p+1th node layer is less than or equal to the preset number of data or p is greater than or equal to the preset number of layers, the model composed of the current p+1 node layer is determined as the power prediction sub-model; The T power prediction sub-models are integrated to obtain the power prediction model.

8. The method according to claim 1, characterized in that: The obtaining a first power set of the server in a first time period includes: Obtaining input and output power of the server during the first time period using a first communication protocol; Current data of a current sensor corresponding to a component of the server is acquired, and power corresponding to the component of the server is determined according to the current data.

9. The method according to claim 1, characterized in that: The adjusting the working state of the server includes: Determining output voltages and currents corresponding to the server and components of the server according to the second power set; A duty cycle of the server is determined according to the second power set.

10. The method according to claim 6, characterized in that After obtaining the trained power prediction model, the method further includes: Acquire a third power set of the server in a fourth time period, and determine a second power feature corresponding to the third power set; The third power set and the second power feature are used to update parameters of the power prediction model to obtain an updated power prediction model.

11. A server management device, characterized in that: The device comprises: A first acquisition module, configured to acquire a first power set of a server in a first time period, wherein the first power set is used to indicate an overall power of the server and a power of each component of the server; A power prediction module is used to input the first power set into a power prediction model to obtain a second power set, wherein the second power set is used to indicate the power of the server and its components in a second time period; a state adjustment module is used to adjust the working state of the server when the second power set meets preset conditions.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the method according to any one of claims 1 to 10 when executed by a processor.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 10 are implemented.

Images