Server energy consumption prediction method and device, computing equipment and storage medium

Abstract

The invention discloses a server energy consumption prediction method and device, computing equipment and a storage medium. First. The power of the first server at the multiple sampling moments is predicted according to the rated power and the no-load power of the first server and the load rates of the first server at the multiple sampling moments in the prediction time period, and the load rates of the multiple sampling moments correspond to the power of the multiple sampling moments one to one. And then, predicting the energy consumption of the first server in the prediction time period according to the power at the plurality of sampling moments. According to the scheme of the embodiment of the invention, server energy consumption prediction is performed based on the load rate, and the prediction result is relatively accurate.

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to a server energy consumption prediction method, device, computing equipment and storage medium. Background Art

[0002] During the entire life cycle of a data center, from equipment installation and deployment, software configuration, formal operation to maintenance and upgrades, there are many scenarios that require predictions of energy consumption data for servers and the entire server cluster. Therefore, server energy consumption prediction has become a key technology to support the planning, design, operation and maintenance of data centers.

[0003] The current energy consumption prediction technology based on time series considers the change pattern of energy consumption data over time, and uses historical time series energy consumption data for statistical analysis to infer the energy consumption of the server at a future time point. However, in fact, the current energy consumption development pattern of the server may not be consistent with its historical energy consumption pattern, that is, the correlation between the energy consumption of the server and time is relatively weak, and the prediction results may not be accurate. In addition, the energy consumption prediction technology based on time series relies on historical energy consumption data in both the training stage and the prediction stage. The quantity and quality of the data will directly affect the accuracy and precision of the prediction of the time series prediction model. Summary of the invention

[0004] The present application provides a server energy consumption prediction method, device, computing equipment and storage medium, which can provide a relatively accurate server energy consumption prediction.

[0005] In the first aspect, the present application provides a method for predicting server energy consumption, the method comprising: predicting the power of the first server at the above-mentioned multiple sampling moments according to the rated power, no-load power and the load rate of the first server at multiple sampling moments in the prediction period, and then predicting the energy consumption of the first server in the prediction period according to the power at the above-mentioned multiple sampling moments. Among them, the first server is a certain server / certain servers for which energy consumption prediction is required, and the prediction period is a certain time period to be predicted, which is not specifically limited here. The load rates at the above-mentioned multiple sampling moments correspond one-to-one to the powers at the above-mentioned multiple sampling moments, and the number of the above-mentioned multiple sampling moments and the sampling period / sampling interval (equal interval or unequal interval), etc., are not specifically limited here.

[0006] It should be understood that since the server load rate is a key factor affecting the server energy consumption, this solution can achieve a higher prediction accuracy by predicting the server energy consumption based on the server load rate. Compared with the server energy consumption prediction technology based on time series, this solution does not rely on historical energy consumption data in the prediction stage, but predicts the server power consumption in the prediction period based on the server's rated power, no-load power and the load rate at multiple sampling moments during the server operation.

[0007] Based on the first aspect, in a possible implementation scheme, the load rate of the first server at the first sampling moment includes one or more of the processor usage, disk usage, memory usage, and network bandwidth usage of the first server at the first sampling moment, and the first sampling moment is one of the above-mentioned multiple sampling moments (it can be any one or a specific sampling moment).

[0008] In this solution, the load rate of the server can be evaluated by one or more resources (such as computing, memory, disk, network and other resource dimensions), and can be reasonably selected according to the actual business scenario. For example, the server used in the high-performance computing scenario is usually equipped with multiple central processing units (CPUs), and the main energy-consuming component of the server is the CPU. The CPU usage rate can directly reflect the energy consumption trend / energy consumption level of the server. Therefore, it is more reasonable to use the CPU usage rate as an evaluation indicator of the load rate of the server. For another example, a server used in a database scenario has a high frequency of disk access, and its main energy-consuming components are the CPU and the disk. Therefore, the CPU usage rate and the disk usage rate can be combined to reflect the load rate of the server, such as taking the weighted average of the CPU usage rate and the disk usage rate as the load rate, and the weight can be set and adjusted according to the actual application scenario.

[0009] Based on the first aspect, in a possible implementation scheme, the processor usage rate is determined according to the rated powers of the multiple processors in the first server and the usage rates of the multiple processors at the first sampling moment.

[0010] That is to say, when the first server includes multiple processors, the utilization rate of the first server at the sampling moment can be determined based on the rated power of the multiple servers and their utilization rates at the sampling moment. For example, in addition to the central processing unit, the server in the artificial intelligence (AI) computing scenario is usually also equipped with coprocessors such as a graphics processing unit (GPU), a neural network processing unit (NPU), and a data processing unit (DPU). The utilization rates of the multiple processors in the first server at a certain sampling moment can be used as the weights of the corresponding processors, and then the rated powers of the multiple processors are weighted averaged according to the above weights, and the weighted average obtained is used as the load rate of the first server at the sampling moment.

[0011] Based on the first aspect, in a possible implementation scheme, the multiple processors include one or more of a central processing unit, a graphics processing unit, a neural network processor, and a data processor. The multiple processors in the first server may belong to the same processor type or to multiple different processor types, which is not specifically limited here. In other words, the present invention supports energy consumption prediction for heterogeneous computing servers.

[0012] Based on the first aspect, in a possible implementation scheme, the load rate of the first server at the first sampling moment is determined based on the system average load (load average) of the first server at the first sampling moment, wherein the first sampling moment is one of the above-mentioned multiple sampling moments, which can be any one or a specific sampling moment.

[0013] It should be understood that the system average load indicator can be used to measure the number of threads that are working and waiting to work in the server. In other words, it measures the number of threads that are not completely idle, including the number of threads that are occupying the CPU execution right and the number of threads waiting for the CPU. It reflects the usage of various resources (such as computing, memory, disk, network bandwidth, etc.), and can better reflect the energy consumption level of the server.

[0014] Based on the first aspect, in a possible implementation scheme, the rated power, no-load power and load rate of the first server at the second sampling moment of the first server can be input into the server power prediction model, so as to predict the power of the first server at the second sampling moment. The second sampling moment is one of the above-mentioned multiple sampling moments, and can be any one of the multiple sampling moments or a specific sampling moment, which is not specifically limited here. The server power consumption prediction model includes a curve fitting function, which is used to represent the relationship between the power consumption of the first server and the load rate of the first server.

[0015] It should be understood that according to the statistics and analysis of the server energy consumption data, it can be found that there is a certain change relationship between the energy consumption of the server and its load rate. Therefore, this solution uses a curve fitting function to represent the change trend of the energy consumption of the first server with the load rate. Compared with the server energy consumption prediction technology based on neural networks, the curve fitting function used in this solution requires fewer parameters (neural networks require a large number of parameters, such as network topology, weights and initial values ​​of thresholds), has lower computational complexity, and is faster, thereby supporting rapid energy consumption prediction of servers or even server clusters.

[0016] Based on the first aspect, in a possible implementation scheme, the above curve fitting function is constructed based on an exponential function and a sigmoid function. x (a is a constant and a>0, a≠1, x is a variable), the image shape of the S-type function is similar to the letter S, and can be a sigmoid function, an inverse tangent function (i.e., an arctan function), etc., without specific limitation.

[0017] It should be understood that by analyzing the power data of the server, it can be found that the power of the server has the following change characteristics: the power of the server shows a trend similar to an exponential increase as the load rate increases. Assume that the load rate of the server is divided into three stages, namely low load (such as 0-30%), medium load (such as 30%-90%) and high load (such as 90%-100%). In the low load stage, as the load rate increases, the power of the server changes slowly, and the actual power is close to the no-load power. In the medium load stage, as the load rate increases, the server power changes faster, and the actual power rises faster as the load increases. In the high load stage, the power of the server is close to the rated power. Therefore, this application uses exponential functions and S-type functions to construct a server power prediction model, which can better reflect the above-mentioned change law between server power consumption and load rate, and thus can obtain better prediction results, and the model is highly interpretable and the prediction results are highly credible.

[0018] Based on the first aspect, in a possible implementation scheme, the method further includes: obtaining a load rate of the first server at a third sampling moment, selecting a plurality of historical power data according to the load rate at the third sampling moment, and then predicting the power of the first server at the third sampling moment according to the plurality of historical power data. The third sampling moment is a sampling moment different from the second sampling moment among the plurality of sampling moments, and the plurality of historical power data are all historical powers of the first server at the load rate at the third sampling moment.

[0019] In this solution, the power at the third sampling moment is not predicted by the server energy consumption prediction model, but is calculated based on multiple historical power data corresponding to the load rate at the third sampling moment (such as calculating the average value or weighted average value, etc.). For example, during the operation of the first server, the power data of the first server can be continuously collected according to a certain set sampling period, and each power data corresponds to a load rate. This solution can use all power data corresponding to the load rate at the third sampling moment collected within a certain time range (such as the current sliding window) as historical power data, and then use the average value of these historical power data as the predicted power at the third sampling moment.

[0020] Based on the first aspect, in a possible implementation scheme, the first server operates in a first mode during the prediction period, and the no-load power and energy consumption of the first server during the prediction period correspond to the second mode.

[0021] It should be understood that the first mode and the second mode are two different operating modes of the server (such as the energy-saving mode and the high-performance mode of the server). The rated power in the two modes is the same, but the no-load power may be different. When the first server operates in the first mode during the forecast period, the power consumption of the first server in the second mode during the forecast period can be predicted based on the load rate at multiple sampling moments in the forecast period. Based on the above prediction, it is helpful to compare the energy consumption differences between the two modes, evaluate the energy-saving effects of different modes, and assist users in selecting the appropriate server operation mode.

[0022] In a second aspect, the present application also provides a server energy consumption prediction device, including a prediction module. The prediction module is used to predict the power of the first server at the above-mentioned multiple sampling moments according to the rated power, no-load power and the load rate of the first server at multiple sampling moments in the prediction period. Among them, the load rate at the above-mentioned multiple sampling moments corresponds one-to-one to the power at the above-mentioned multiple sampling moments. The prediction module is also used to predict the energy consumption of the first server in the prediction period according to the power at the above-mentioned multiple sampling moments.

[0023] The server energy consumption prediction device in the second aspect may also include more or fewer modules, which are not specifically limited here. The server energy consumption prediction device in the second aspect is used to execute the method steps of any implementation scheme in the first aspect, which can be specifically referred to in the above description and will not be repeated here.

[0024] In a third aspect, the present application further provides a computing device, comprising a processor and a memory, wherein the processor is configured to execute instructions stored in the memory so that the computing device executes a method as in any one of the embodiments in the first aspect.

[0025] In a fourth aspect, the present application further provides a computer-readable storage medium, comprising computer program instructions. When the above-mentioned computer program instructions are executed by a computing device, the computing device executes a method as in any implementation scheme in the first aspect.

[0026] In a fifth aspect, the present application further provides a computer program product comprising instructions, characterized in that when the instructions are executed by a computing device, the computing device executes a method as in any one of the embodiments of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments.

[0028] Figure 1 It is a schematic diagram of the architecture of an energy consumption prediction system provided in an embodiment of the present application;

[0029] Figure 2 It is a flow chart of a server energy consumption prediction method provided in an embodiment of the present application;

[0030] Figure 3 This is a schematic diagram of a trend of server power changes with load rate provided in an embodiment of the present application;

[0031] Figure 4 is the exponential function y=e provided in the embodiment of the present application x Schematic diagram of

[0032] Figure 5 is a schematic diagram of a sigmoid function provided in an embodiment of the present application;

[0033] Figure 6 It is a schematic diagram of sampling the power of a server according to a preset sampling period provided in an embodiment of the present application;

[0034] Figure 7 It is an optimization flow chart of a server power prediction model provided by an embodiment of the present application;

[0035] Figure 8It is a schematic diagram of a data center business scenario provided by an embodiment of the present application;

[0036] Fig. 9 It is a structural schematic diagram of a server energy consumption prediction device 100 provided in an embodiment of the present application;

[0037] Fig.10 This is a schematic diagram of the relationship between a server operation and maintenance platform and a server cluster provided in an embodiment of the present application;

[0038] Fig.11 It is a structural diagram of a computing device 1100 provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to facilitate understanding of the technical solutions of the embodiments of the present application, some terms and concepts involved in the embodiments of the present application are briefly introduced below.

[0040] 1. Rated Power

[0041] In electrical and mechanical engineering, the rated power of a device is the highest input power allowed to flow through a specific device. Depending on the specific field, the power referred to by the term can be electrical or mechanical. The rated power may also refer to average power and maximum power, which may vary depending on the type of device and its application. Rated power limits are usually set by manufacturers as a guideline to protect equipment and simplify the design of larger systems by providing an operating level at which the equipment will not be damaged while allowing a certain safety margin.

[0042] 2. Idle Load Power

[0043] No-load power refers to the electrical energy used by a device when it is not performing any function. No-load means neither open circuit nor short circuit. The device circuit has both input and output, and there is also current (no-load current) in the device circuit. For this circuit, there is no open circuit, let alone short circuit.

[0044] 3. System average load (Load Average)

[0045] The system load average measures the number of threads that are working and waiting for work (CPU, disk, uninterruptible lock) for the entire system. In other words, this method measures the number of threads that are not completely idle, including the number of threads that are occupying the CPU execution right and the number of threads waiting for the CPU, reflecting the demand for different resources.

[0046] 4. Maximum Likelihood Estimation (MLE)

[0047] Maximum likelihood estimation is a statistical method based on the maximum likelihood principle and is the application of probability theory in statistics. Maximum likelihood estimation provides a method to evaluate model parameters given observed data, that is, "the model is fixed, but the parameters are unknown". Through several experiments, observing the results, and using the experimental results to obtain a parameter value that maximizes the probability of the sample appearing is called maximum likelihood estimation.

[0048] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of an energy consumption prediction system provided in an embodiment of the present application, including a server energy consumption prediction device 100 and a server cluster 200, and the server cluster 200 includes one or more servers.

[0049] The server energy consumption prediction device 100 is used to predict the energy consumption of servers in the server cluster 200. The server energy consumption prediction device 100 may include one or more computing devices, which may be deployed in a data center in a cloud environment, in an edge environment, or in other environments, which is not specifically limited in the embodiments of the present application.

[0050] Optionally, the server cluster 200 may include one or more types of servers, and different server types may differ in hardware configuration and / or application scenarios.

[0051] For example, servers may include general-purpose servers, computing servers, storage servers, and other types. For servers in high performance computing (HPC) scenarios, multiple central processing units (CPUs) are generally configured. For servers in scenarios such as artificial intelligence (AI) or heterogeneous computing power, they are not only configured with CPUs, but may also be configured with coprocessors such as graphics processing units (GPUs), neural network processing units (NPUs), and data processing units (DPUs). In addition to the above-mentioned server types, there may be other ways of dividing server types, which are not specifically limited in the embodiments of the present application.

[0052] Based on the above content, an embodiment of the server energy consumption prediction method provided by the present application is introduced below.

[0053] See also Figure 2 , Figure 2 It is a flow chart of a server energy consumption prediction method provided in an embodiment of the present application, including steps S201 to S202.

[0054] S201. Predict the power of the first server at multiple sampling moments according to the rated power, no-load power and load rate of the first server at multiple sampling moments in a prediction period, wherein the load rates at the multiple sampling moments correspond one-to-one to the powers at the multiple sampling moments.

[0055] The first server may be a certain server / some servers in the server cluster 200, which is not specifically limited here.

[0056] The rated power of the first server can be directly obtained from the specification information of the first server, or can be directly measured by running the first server in a full-load state. For example, by running a benchmark test program to make the first server run in a full-load state, the highest actual operating performance of the hardware of the first server can be measured, thereby obtaining the rated power (maximum operating power) of the first server, which is not specifically limited in the embodiments of the present application.

[0057] The no-load power of the first server can be directly measured when the first server has no business deployed (no task running), or obtained through other methods, and the embodiments of the present application do not make any specific limitations.

[0058] The number of the above-mentioned multiple sampling moments and the sampling interval / sampling period in the prediction period can be set and adjusted according to the actual application scenario, and the embodiments of the present application do not make specific limitations. For example, assuming that the prediction period is 9:00-10:00 on a certain day, and the sampling period is set to 5s, then based on the sampling period, the load rate of the first server needs to be sampled every 5s in the prediction period, so as to obtain the load rate of multiple sampling moments.

[0059] Optionally, the load rate of the first server at the first sampling moment includes one or more of the processor usage, disk usage, memory usage, and network bandwidth usage of the first server at the first sampling moment. The first sampling moment is a sampling moment among the above multiple sampling moments (can be any one or a specific sampling moment). In other words, the load rate of the first server can be evaluated by one or more resources (such as computing, memory, disk, network and other resource dimensions), and can be reasonably selected according to the actual business scenario.

[0060] For example, servers used in HPC scenarios are usually equipped with multiple CPUs. The main energy-consuming component of this type of server is the CPU, and the CPU usage rate can directly reflect the energy consumption trend / energy consumption level of this type of server. Therefore, when the first server is this type of server, it is more reasonable to use the CPU usage rate as an evaluation indicator of the load rate of the first server.

[0061] For another example, assume that the first server is a server used in a database scenario, and its access frequency to the disk is relatively high. The main energy-consuming components of the server are the CPU and the disk. Therefore, the CPU usage and disk usage can be combined to reflect the load rate of the server. For example, the average / weighted average of the CPU usage and the disk usage can be used as the load rate of the server. The weight can be set and adjusted according to the actual application scenario, and no specific limitation is made here.

[0062] Optionally, the processor usage of the first server at the first sampling moment is determined according to the rated power of multiple processors in the first server and the usage of the multiple processors at the first sampling moment. The multiple processors may be processors of the same type or processors of different types, and may include one or more of a central processing unit, a graphics processing unit, a neural network processor, and a data processor, which is not specifically limited in the embodiments of the present application.

[0063] For example, assuming that the processor usage rate is used as the load rate of the first server, the processor usage rate of the first server at sampling time k can be determined according to the rated power of multiple processors in the first server and their usage rates at sampling time k. The specific calculation formula can refer to the following formula (1):

[0064]

[0065] Wherein, LoadRate represents the load rate, LoadRate(k) is the load rate of the first server at sampling time k, and sampling time k is a sampling time in the prediction period. j∈[1,N], N is the number of processors in the first server, and N is a positive integer. Power(j) is the rated power of the jth processor in the first server, and Usage(k,j) is the usage rate of the jth processor in the first server at sampling time k. Power total It is the sum of the rated powers of the N processors in the first server, that is, the total rated power of the N processors.

[0066] Optionally, the load rate of the first server at the first sampling moment is determined according to the system average load of the first server at the first sampling moment. The first sampling moment is a sampling moment among multiple sampling moments in a preset period (can be any one or a specific sampling moment).

[0067] For example, for general server business scenarios in cloud data centers, the processor is not necessarily the main energy-consuming component of the server. The hard disk, memory, and even the network may account for more than 50% of the server's energy consumption. In this case, the system load average indicator can be used to evaluate the server's load rate. Since the system load average measures the number of threads that are occupying the CPU execution right and the number of threads waiting for the CPU (including threads waiting for network input and output (IO) and disk IO), it reflects the usage of various resources (such as computing, memory, disk, network bandwidth, etc.), and can therefore better reflect the server's energy consumption level. The load rate calculation method in this scenario can refer to the following formula (2):

[0068]

[0069] Wherein, LoadRate(k) is the load rate of the first server at sampling time k, and sampling time k is a sampling time in the prediction period. LoadAverage(k) is the system average load of the first server at sampling time k, that is, the number of threads that are not completely idle on the first server at sampling time k, including the number of threads that are occupying the CPU execution right of the first server at sampling time k and the number of threads waiting for the CPU.

[0070] Optionally, the energy consumption of the first server may be predicted based on a server power prediction model, which is described in detail below.

[0071] See also Figure 3 , Figure 3 This is a trend chart of server power changing with load rate provided by the embodiment of the present application. The chart is generated by collecting power data of a server. The load rate is in the interval [0,1], and the server power is between [no-load power, rated power]. Assuming that the server load rate is divided into three stages: low load (such as 0-30%), medium load (such as 30%-90%), and high load (such as 90%-100%), then Figure 3 The server power has the following characteristics:

[0072] 1) Low load: When the load is less than 30%, the server power changes slowly and the actual power is close to the no-load power;

[0073] 2) Medium load: When 30%<load<90%, the server power changes faster, and the actual power increases faster as the load increases;

[0074] 3) High load: When the load is greater than 90%, the actual power of the server is close to the rated power.

[0075] Similarly, by conducting statistical analysis on the power data of a large number of servers, it can be found that the power of the servers basically conforms to the following change characteristics: the server power shows an exponential upward trend with the increase of the load rate; when the server load rate is 0, the server is in a no-load state, and the no-load power is the smallest; when the server load rate is below 50%, the server power changes relatively slowly, but as the load rate continues to increase, the rate of increase of the server power gradually accelerates; when the server load rate reaches 100%, the server is in a full load state, and the full load power is close to or equal to the rated power.

[0076] Based on the above analysis, a curve fitting function can be used to represent the changing trend of the energy consumption of the first server with the load rate, and then the curve fitting function can be used as a server power prediction model to predict the power consumption of the first server. For example, the rated power, no-load power and load rate of the first server at the second sampling moment can be input into the server power prediction model to predict the power of the first server at the second sampling moment. Among them, the second sampling moment is a sampling moment in the prediction period (can be any one). It should be understood that compared with the server energy consumption prediction technology based on neural networks, the curve fitting function used in this solution has a computational complexity of O(1), a lower computational complexity, and a faster calculation speed, thereby being able to support rapid energy consumption prediction of servers or even server clusters. Due to the low computational complexity, the computing power requirements for the server energy consumption prediction device 100 are relatively low, and the applicable scenarios are wider.

[0077] Optionally, the server power prediction model can be constructed based on the exponential function and the S-shaped function to better reflect the variation law between the server power and its load rate (similar to Figure 3 ), thereby obtaining a better server power prediction effect. Among them, the exponential function is a x (a is a constant and a>0, a≠1, x is a variable). The shape of the S-shaped function is similar to the letter "S", which can be a sigmoid function (i.e. x is an independent variable), inverse tangent function (i.e., arctan function), etc., and the embodiments of the present application do not make specific limitations.

[0078] For example, see Figure 4 , Figure 4 is the exponential function y=e provided in the embodiment of the present application x Schematic diagram of , where e is a natural constant. Figure 5 , Figure 5 is a schematic diagram of the sigmoid function provided in the embodiment of the present application, and the function diagram is similar to the letter "S". Based on the above two functions, a server power prediction model of formula (3) can be constructed:

[0079]

[0080] Among them, Power rated is the rated power of the server. idle is the no-load power of the server. LoadRate is the load rate of the server, which can be calculated by the previous formula (1), formula (2) or other methods. pred It is the server power under the load of LoadRate predicted by the model.

[0081] Among them, α and β are two adjustment factors of the model in formula (3). α is the full-load adjustment factor (α>0), which is used to adjust the speed of approaching the rated power. The larger the value of α, the faster the power approaches the rated power. β is the no-load adjustment factor (0<β≤1), which is used to adjust the speed of approaching the no-load power. The larger the value of β, the faster the power rises under low load. It should be understood that since the parameters of this model are fewer than those of the neural network, it can be quickly put into use after configuring the model parameters (including no-load power, rated power, adjustment factors α and β, which are easy to obtain). There is no need to pre-train the model based on a large-scale data set (the data complexity is low), and the computational complexity is low, so it can support fast server energy consumption prediction and the model is more versatile.

[0082] Optionally, the adjustment factor α in the model of formula (3) can be configured as 1 by default, and β can be configured as 0.5 by default, so that the server power prediction function can be provided based on the default adjustment factor parameters. However, the above α and β may not be optimal for the server power prediction model of the first server, so the adjustment factors α and β in the model can be optimized by collecting the power data of the first server, so that the power prediction of the model is more accurate. Taking the maximum likelihood estimation method as an example, the server power prediction model is optimized by the maximum likelihood estimation method.

[0083] Assuming that the prediction error of the model in equation (3) conforms to the Gaussian distribution, the least squares method and the maximum likelihood estimation method are equivalent. Here, different load rates are used as sampling points, and the load rate range is [0,1]. Each sampling point may correspond to one or more power data in the sliding window. The average / weighted average of all power data corresponding to the sampling point i in the sliding window can be used as the power corresponding to the sampling point i.

[0084] For example, Figure 6This is a schematic diagram of sampling the power of a server according to a preset sampling period provided by an embodiment of the present application. There are 7 sampling moments in the current sliding window, corresponding to 7 power data, the load rate of the first two sampling moments is 0.5, and the load rate of the next 5 sampling moments is 0.8.

[0085] Take the load rate of 0.5 as a sampling point, which corresponds to the first two sampling moments in the sliding window. These two sampling moments correspond to two power data. The average value of these two power data can be used as the power corresponding to the load rate of 0.5. Similarly, take the load rate of 0.8 as another sampling point, which corresponds to the last five sampling moments in the sliding window. These five sampling moments correspond to five power data one by one. The average value of these five power data can be used as the power corresponding to the load rate of 0.8.

[0086] For n sampling points (representing n different load rates), the prediction error diff of the server power prediction model of formula (3) can be expressed by formula (4):

[0087]

[0088] Among them, i belongs to [1,n], and n is a positive integer. is the server power at the load rate corresponding to sampling point i predicted by formula (3). It is the power corresponding to sampling point i, that is, the average value of one or more power data actually sampled corresponding to sampling point i.

[0089] The power of n sampling points is substituted into formula (4) for calculation, and then a set of adjustment factors α and β that minimize diff are selected as the optimized adjustment factors, so as to modify and optimize the model of formula (3), which can be used for server power prediction in the next stage.

[0090] The embodiments of the present application do not specifically limit the optimization timing, optimization frequency, etc. of the server power prediction model.

[0091] For example, by setting an appropriate sliding window size, the number of sampling points in the sliding window will not be too small, thereby providing sufficient data input for model optimization. Alternatively, when the number of sampling points in the sliding window is less than the preset value, the model will not be optimized; when the number of sampling points in the sliding window is greater than or equal to the preset value, the model will be optimized.

[0092] For example, Figure 7 As shown, Figure 7It is an optimization flow chart of a server power prediction model provided by an embodiment of the present application. First, power data is collected based on a sliding window, and then a new set of adjustment factors (α and β) is determined based on the maximum likelihood estimation method. If the degree of improvement of the accuracy of the server power prediction model by this new set of adjustment factors is greater than or equal to the configuration threshold, that is, the accuracy of the model based on the new adjustment factors is higher than the accuracy of the original model by more than the configuration threshold, then the model is modified with this new set of adjustment factors, and the modified model can be used to predict the power of the server subsequently, thereby obtaining a higher prediction accuracy. If the degree of improvement of the accuracy of the server power prediction model by this new set of adjustment factors is less than the configuration threshold, the model retains the old adjustment factors without modification. Subsequently, the next round of data collection can be started for a new round of model optimization.

[0093] The adjustment factor parameters of the server power prediction model can be dynamically optimized during the use of the model, which can dynamically improve the prediction accuracy. The model can also be optimized after the energy consumption prediction task is completed or in the interval between energy consumption prediction tasks, which will not affect the use of the model. Model optimization can be performed through a separate task process, and the process can also be suspended to ensure the priority execution of the energy consumption prediction task.

[0094] Optionally, since the adjustment factors α and β in the server power prediction model have a great influence on the accuracy of energy consumption prediction, different server types (or server models, etc.) can be distinguished, and a set of default optimization adjustment factors can be preset for each server type, so that more accurate adjustment factor parameters can be provided in the model parameter configuration stage, thereby achieving more accurate energy consumption prediction.

[0095] For example, the optimization adjustment factor parameters can be given synchronously with the hardware indicator specifications of the server. When it is necessary to build a corresponding server power prediction model for a server, the optimization adjustment factor parameters can be directly obtained from the hardware indicator specifications of the server for configuring the model. After the model configuration is completed, it can be used to predict the power of the server. At this time, a relatively good prediction effect can be obtained. Subsequently, the adjustment factor parameters of the model can be further optimized according to the collected power data of the server to improve the prediction effect of the model, which will not be elaborated here.

[0096] For another example, the optimization adjustment factors corresponding to various types of servers can be given as default parameters in the form of a configuration table, and the configuration table is used to record a set of optimization adjustment factor parameters corresponding to various server types. When it is necessary to build a corresponding server power prediction model for a certain server, a corresponding set of optimization adjustment factors can be obtained from the configuration table according to the type of the server, and then the model can be configured using this set of optimization adjustment factors, and then power prediction can be performed based on the configured model. Subsequently, the adjustment factor parameters of the model can be further optimized according to the actual sampled power data to improve the prediction ability of the model, which will not be elaborated here.

[0097] Optionally, the optimization adjustment factor parameters corresponding to each server type may be determined based on historical optimization parameters corresponding to the server type.

[0098] For example, for a certain server, assuming that there is no optimization adjustment factor parameter corresponding to the server type of the server in the configuration table, the server power prediction model of the server will initially select the default adjustment factor (i.e., α=1, β=0.5) for configuration, and the configured model can be used for power prediction. Subsequently, the adjustment factor parameters of the model can be further optimized according to the actual sampled power data, and the model will obtain a new set of adjustment factor parameters after optimization, which can be written in the configuration table as a set of optimization adjustment factors corresponding to the server type of the server.

[0099] When another server of the same type needs to be predicted, the optimization adjustment factor parameters corresponding to the server type can be directly obtained from the configuration table, and then the server power prediction model corresponding to the server can be configured so that the model can obtain a more accurate prediction effect from the beginning.

[0100] Optionally, the power at the third sampling moment in the prediction period may not be predicted by the server power prediction model, but may be predicted based on the collected historical power data. The third sampling moment is a sampling moment in the prediction period, but the third sampling moment is different from the previous second sampling moment (the power at the second sampling moment is obtained through the server power prediction model). Specifically, the load rate of the first server at the third sampling moment may be obtained first, and then a plurality of historical power data may be selected according to the load rate at the third sampling moment. The plurality of historical power data are all the historical powers of the first server at the load rate at the third sampling moment, that is, the load rates corresponding to the plurality of power data are the same as the load rates at the third sampling moment. Subsequently, the power of the first server at the third sampling moment is predicted based on the plurality of historical power data.

[0101] For example, during the operation of a server, the power data and load rate of the server can be collected according to a certain sampling period. Figure 6 As shown in the figure, the current sliding window contains 7 sampling moments, the load rate corresponding to the first 2 sampling moments is 0.5, and the load rate corresponding to the last 5 sampling moments is 0.8. Assuming that the load rate of a sampling moment in the prediction period is 0.8, the power data corresponding to the last 5 sampling moments can be selected from the above sliding window as the historical power data, and then the average value (or weighted average value) of these historical power data can be used as the predicted value of the power at the sampling moment.

[0102] S202: predict the energy consumption of the first server in the prediction period according to the power at the above-mentioned multiple sampling moments.

[0103] Specifically, after the power at multiple sampling moments in the prediction period is predicted through step S201, the energy consumption of the first server in the prediction period can be predicted based on formula (5):

[0104]

[0105] Wherein, t1 and t2 are the start time and end time of the prediction period, respectively. It can be seen that formula (5) integrates the predicted server power according to time in the prediction period, thereby obtaining the energy consumption of the server in the prediction period.

[0106] Optionally, when the sampling period of the above-mentioned multiple sampling moments is T, formula (5) can be further rewritten as formula (6), so that the energy consumption of the server can be predicted based on formula (6).

[0107] Power Consumption pred =∑ k T·Power pred (k) (6)

[0108] Where k∈[1,M], M is the total number of sampling moments in the preset time period, and M is determined by the length of the prediction period and the sampling period T. pred (k) is the power of the server at sampling time k within the preset time period, which can be predicted according to any of the methods described above and will not be repeated here.

[0109] Optionally, the first server can operate in the first mode during the prediction period, and the load rates at multiple sampling moments in the prediction period are all the load rates of the first server in the first mode. It should be understood that the no-load power of the first server operating in different modes (such as energy-saving mode, high-performance mode, etc.) may be different. Assuming that the no-load power used in step S201 corresponds to the second mode, the power predicted at multiple sampling moments is the power in the second mode. Then in step S202, the energy consumption of the first server operating in the second mode during the prediction period can be further predicted based on the power at these multiple sampling moments. In other words, the method of the embodiment of the present application can use the load rate of the server operating in the first mode to predict the energy consumption of the server under the second model (the server does not need to actually operate in the second mode), which helps to compare the energy consumption differences of the server in different modes, and then evaluate the energy-saving effects of different modes, and can also assist users in selecting a suitable server operation mode.

[0110] Next, combined with formula (3), Figure 2 The server energy consumption prediction method is illustrated with a specific example.

[0111] Taking a server used for AI model training as an example, assume that the AI ​​server is configured with 2 CPUs (respectively denoted as CPU1 and CPU2) and 4 GPUs (respectively denoted as GPU1, GPU2, GPU3, and GPU4).

[0112] As mentioned above, the server power prediction model of formula (3) is based on the load rate. Before using the model for prediction, the benchmark parameter configuration needs to be completed. The benchmark parameters include no-load power, full-load power, full-load adjustment factor α and no-load adjustment factor β. It is assumed here that the two adjustment factors use the default values, that is, α=1 and β=0.5. The rated power and no-load power of the server can be obtained by reading the server hardware specifications or measuring through the program. It is assumed here that the rated power is 3000w, and the no-load power of the server in high-performance mode is 300w. Substituting the above four parameters into formula (3), the server power prediction model of the server in high-performance mode can be obtained, as shown in the following formula (7):

[0113]

[0114] Among them, LoadRate(k) is the load rate of the AI ​​server at sampling time k, k∈[1,M], M is the number of sampling times in the prediction period, which can be determined by the length of the prediction period and the sampling period T.

[0115] Next, we need to determine the calculation method of the load rate in formula (7). It should be understood that since the server is used for AI model training and is equipped with multiple processors (including CPU and GPU), these processors are the main energy-consuming components of the server, so the load rate of the server can be determined based on the usage rate and rated power of these processors.

[0116] The rated power of each processor can be directly obtained from the hardware specifications of the AI ​​server. Here, it is assumed that in high-performance mode, the rated power of each CPU of the AI ​​server is 300W, and the rated power of each GPU is 500W. Substituting the rated power of each processor and the actual sampled utilization rate of each processor into the formula (1) in the previous text, the server load rate calculation formula (8) can be obtained:

[0117]

[0118] Among them, U CPU1 (k) and U CPU2 (k) are the utilization rates of CPU1 and CPU2 at sampling time k, U GPU1 (k) U GPU2 (k) U GPU3 (k) and U GPU4 (k) are the utilization rates of GPU1, GPU2, GPU3 and GPU4 at sampling time k. LoadRate(k) is the power of the server at sampling time k predicted by formula (8).

[0119] At this point, the server power prediction model of the server has been configured, and the calculation method of the load rate has been determined. Next, the energy consumption of the server in the prediction period can be predicted based on formulas (7) and (8).

[0120] Assume that the server is currently running in energy-saving mode, and you want to predict the energy consumption of the server running in high-performance mode for one hour. Assume that the sampling period T is 5s, and sample the usage rate of each processor of the server within one hour according to the sampling period, so that the usage rate data of each processor at 720 different sampling times can be obtained. Substitute the usage rate of each processor at each sampling time into formula (8) to obtain the load rate of the server at each sampling time. Then, substitute the load rate at each sampling time into the server power prediction model of formula (7), so that the power of the server at each sampling time can be predicted.

[0121] Subsequently, referring to formula (5) or (6) in the previous text, the power at the above 720 sampling moments can be integrated according to time, so as to predict the energy consumption of the AI ​​server running in high-performance mode for one hour. The specific calculation formula can be found in formula (9):

[0122]

[0123] Similarly, energy consumption prediction can be performed for each server in the server cluster to obtain the energy consumption prediction result of the entire server cluster, thereby helping users make reasonable designs and decisions on the planning, design and operation of the data center.

[0124] In summary, since the load rate of the server is a key factor affecting the energy consumption of the server, the embodiment of the present application predicts the energy consumption of the server based on the load rate of the server, and can obtain a higher prediction accuracy. Compared with the server energy consumption prediction technology based on time series, the embodiment of the present application can not rely on historical energy consumption data in the prediction stage, but predict the power consumption of the server in the prediction period based on the rated power, no-load power and the load rate of the server at multiple sampling times. Compared with the server energy consumption prediction technology based on neural networks, the server energy consumption prediction model of the embodiment of the present application is implemented based on a curve fitting function, and the model only has the variable of load rate (single factor), so the calculation complexity is lower, and the computing power requirement of the server energy consumption prediction device 100 is lower, thereby supporting rapid energy consumption prediction for a large number of servers (server clusters), and having a wider range of applications.

[0125] Figure 2 The method of the embodiment can be applied to various business scenarios that require prediction of server energy consumption. Figure 8 As shown, Figure 8 This is a schematic diagram of a data center business scenario provided by an embodiment of the present application, which may include business scenario types such as a supercomputing center, an artificial intelligence computing center (i.e., an intelligent computing center), and a cloud data center. The server energy consumption prediction method of the embodiment of the present application can be applied in the entire life cycle of a data center, from planning and design, system configuration, and operation analysis.

[0126] During the planning and design phase of a data center, the power supply system design of the data center, computer room, cabinet, etc. needs to be calculated based on the power consumption (energy consumption) under different loads. Figure 2 The method of the embodiment predicts the energy consumption of the server (cluster) in the data center to help design the power supply and distribution system of the data center and plan a reasonable simultaneous coefficient.

[0127] During the system configuration phase in the data center, Figure 2 The method of the embodiment predicts the energy consumption of servers (clusters) in a data center, thereby optimizing the configuration of various power consumption capping parameters according to the energy consumption prediction results, including cabinet power consumption capping and computer room power consumption capping configuration, etc., which not only helps to ensure the power supply security of the data center, but also can reasonably coordinate electricity use.

[0128] In the operational analysis phase of the data center, Figure 2 The method of the embodiment predicts server energy consumption, thereby supporting direct comparison of server energy consumption differences between energy-saving mode and high-performance mode, helping users predict energy-saving effects in different modes and evaluate power safety. Figure 2 The method of the embodiment is used to predict server energy consumption, thereby helping users make decisions on the following issues: whether the data center power consumption planning needs to be expanded after system expansion, server upgrade or changes in software and hardware systems, whether the power consumption capping parameters need to be changed, etc.

[0129] It should be understood that in addition to the data center field, Figure 2 The server energy consumption prediction method can also be applied to embedded systems, the Internet of Things (IoT) and other fields, and is not specifically limited in the embodiments of the present application.

[0130] See also Fig. 9 , Fig. 9 1 is a schematic diagram of the structure of a server energy consumption prediction device 100 provided in an embodiment of the present application, including a prediction module 101.

[0131] The prediction module 101 is used to predict the power of the first server at multiple sampling moments according to the rated power, no-load power and load rate of the first server at multiple sampling moments in the prediction period, and to predict the energy consumption of the first server in the prediction period according to the power at the multiple sampling moments. The load rates at the multiple sampling moments correspond to the powers at the multiple sampling moments.

[0132] Optionally, the load rate of the first server at the first sampling moment includes one or more of processor usage, disk usage, memory usage, and network bandwidth usage of the first server at the first sampling moment, and the first sampling moment is one sampling moment among multiple sampling moments.

[0133] Optionally, the server energy consumption prediction device 100 further includes a determination module 102, and the determination module 102 is used to determine the processor usage rate according to the rated power of multiple processors in the first server and the usage rate of the multiple processors at the first sampling moment.

[0134] Optionally, the above-mentioned multiple processors include one or more of CPU, GPU, NPU, and DPU.

[0135] Optionally, the server energy consumption prediction device 100 also includes a determination module 102, which is used to determine the load rate of the first server at the first sampling moment based on the system average load of the first server at the first sampling moment, wherein the first sampling moment is one of the above-mentioned multiple sampling moments.

[0136] Optionally, the prediction module 101 is further used to input the rated power, no-load power and the load rate at the second sampling moment into the server power prediction model, so as to predict the power of the first server at the second sampling moment. The second sampling moment is one of the above-mentioned multiple sampling moments, and the server power consumption prediction model includes a curve fitting function, which is used to represent the relationship between the power consumption of the first server and the load rate of the first server.

[0137] Optionally, the above curve fitting function is constructed based on an exponential function and a S-type function.

[0138] Optionally, the server energy consumption prediction device 100 also includes an acquisition module 103. The acquisition module 103 is used to obtain the load rate of the first server at the third sampling moment, and is also used to select multiple historical power data according to the load rate at the third sampling moment, wherein the multiple historical power data are all historical powers of the first server at the load rate at the third sampling moment. The third sampling moment is a sampling moment among the multiple sampling moments that is different from the second sampling moment. The prediction module 101 is also used to: predict the power of the first server at the third sampling moment according to the multiple historical power data.

[0139] Optionally, the first server operates in a first mode during the prediction period, and the no-load power and energy consumption of the first server during the prediction period correspond to the second mode.

[0140] It should be noted that Fig. 9 The server energy consumption prediction device 100 can be used to perform Figure 2 The method steps of the embodiment can be specifically referred to the above description and will not be repeated here. Fig. 9 The server energy consumption prediction device 100 is only described by taking the above functional module division method as an example. In actual applications, the server energy consumption prediction device 100 can be divided into other functional modules as needed to complete all or part of the functions described above. The above functional modules can be implemented in the form of hardware, software functional modules, or software + hardware, which is not specifically limited in the embodiments of this application.

[0141] Optional, Fig. 9The functions of the server energy consumption prediction device 100 can be implemented as an independent tool software, or can be implemented in the form of a service component in an operation and maintenance management platform (software).

[0142] For example, Fig.10 As shown, Fig.10 This is a schematic diagram of the relationship between a server operation and maintenance platform and a server cluster provided in an embodiment of the present application. The server cluster includes multiple servers (the figure takes server 1 to server N as an example), these servers can belong to one or more server types, and different types of servers have different hardware configurations. Fig.10 This is an example of a server type that is equipped with a CPU, GPU, NPU, DPU and other coprocessors. This type of server can also support load balancing. For other information about server clusters, please refer to Figure 1 The relevant introduction of the server cluster 200 is not repeated here.

[0143] The functions of the server energy consumption prediction device 100 can be implemented in the form of energy consumption management services, such as Fig.10 As shown in the figure, the energy consumption management service has multiple functional modules, including benchmark configuration, load assessment, data collection, energy consumption prediction, model optimization, energy consumption visualization, etc. The energy consumption management service is deployed as a service component on the operation and maintenance management platform of the server cluster, thereby supporting the server cluster to provide data collection, energy consumption prediction, energy consumption data visualization and other capabilities.

[0144] Specifically, the server operation and maintenance platform can uniformly manage the servers in the server cluster. After management, the benchmark configuration module in the energy consumption management service provided by the server operation and maintenance platform can measure the rated power and no-load power of the server before the server deploys the service. Then, the load evaluation module can determine the evaluation method of the load rate according to the type of server and the business scenario, such as weighting according to the rated power and usage rate of multiple processors or calculating the load rate according to the average load of the system. In the operation stage after the server deploys the service, the energy consumption prediction module can predict the energy consumption of the server according to the load rate, and even predict the energy consumption of the entire server cluster, so as to support the evaluation of the power quota, power safety and energy-saving effect of the server cluster, and facilitate the operation and maintenance administrator to manage the operation configuration of the server cluster and give configuration suggestions. The energy consumption prediction module can also notify the alarm system to perform prompt operations when the predicted energy consumption exceeds the threshold, so that the operation and maintenance administrator can formulate adjustment plans in time. The data acquisition module can obtain the power data of the server based on the baseboard management controller (BMC) on the server, and then the model optimization module can optimize the parameters of the server energy consumption prediction model according to the actual power data (the optimization record can be reflected in the log system), thereby improving the prediction accuracy of the model. The energy consumption visualization module in the energy consumption management service also supports the visualization of energy consumption data, and the presentation method is not specifically limited (such as pictures, tables, text, etc.), so that the operation and maintenance administrator can intuitively understand the energy consumption level of one or more servers or even the entire server cluster. The energy consumption visualization module can also intuitively compare the energy consumption prediction results with the historical energy consumption data stored in the database / actually collected energy consumption data.

[0145] Fig.11 It is a structural diagram of a computing device 1100 provided in an embodiment of the present application.

[0146] The computing device 1100 includes a processor 1101, a memory 1102, and a communication interface 1103. The computing device 1100 is specifically used to implement Figure 2Server energy consumption prediction method of an embodiment. Among them, the processor 1101, the memory 1102 and the communication interface 1103 can be interconnected through the internal bus 1104, and can also communicate through other means such as wireless transmission. The embodiment of the present application takes the connection through the bus 1104 as an example. The bus 1104 can be a peripheral component interconnect (PCI) bus, an extended industry standard architecture (EISA) bus, a unified bus (Ubus or UB), a computer express link (CXL) bus, a cache coherent interconnect for accelerators (CCIX) bus, etc. The bus 1104 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.11 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0147] The processor 1101 may be composed of at least one general-purpose processor, such as a central processing unit (CPU), or a combination of a CPU and a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. The processor 1101 executes various types of digital storage instructions, such as software or firmware programs stored in the memory 1102, which enables the computing device 1100 to provide a variety of services.

[0148] The memory 1102 is used to store program codes, and the execution is controlled by the processor 1101. The memory 1102 may include a volatile memory, such as a random access memory (RAM); the memory 1102 may also include a non-volatile memory, such as a read-only memory (ROM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); the memory 1102 may also include a combination of the above types. The memory 1102 may store program codes, specifically including a memory for executing Figure 2 The program code of the server energy consumption prediction method of the embodiment will not be described in detail here.

[0149] The communication interface 1103 may be a wired interface (e.g., an Ethernet interface) or a wireless interface (e.g., a cellular network interface or a wireless LAN interface), or may be an internal interface (e.g., a high-speed serial computer expansion bus (Peripheral Component Interconnect Express, PCIe) bus interface). The wired interface or wireless interface is used to communicate with other devices or modules.

[0150] It should be noted that this embodiment can be implemented by a general physical server, such as an ARM server or an X86 server, or by a virtual machine based on a general physical server combined with NFV technology. A virtual machine refers to a complete computer system with complete hardware system functions simulated by software and running in a completely isolated environment. This application does not make specific limitations. It should be understood that Fig.11 The computing device 1100 shown may be a host computer, a laptop computer, a desktop computer, or the like, or may be a computing device cluster consisting of at least one computing device, which is not specifically limited in the present application.

[0151] It also needs to be explained. Fig.11 This is only a possible implementation of the embodiment of the present application. In actual applications, the computing device 1100 may also include more or fewer components, and the present application does not make specific limitations. For content not shown or described in the embodiment of the present application, please refer to Figure 2 The relevant description of the embodiments will not be repeated here.

[0152] The present application also provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is executed on a computing device (such as the computing device 1100 described above), Figure 2The method of the embodiment is realized. Wherein, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0153] The present application also provides a computer program product. When the computer program product is run on a computing device (such as the computing device 1100 described above), Figure 2 The method of the embodiment is realized.

[0154] What is disclosed above is only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application. It should also be noted that the terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application and the attached claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in this article refers to any or all possible combinations of one or more associated listed items.

Claims

1. A server energy consumption prediction method, characterized in that: The method comprises: Predicting the power of the first server at multiple sampling moments according to the rated power and no-load power of the first server and the load rates of the first server at multiple sampling moments in a prediction period, wherein the load rates at the multiple sampling moments correspond one-to-one to the powers at the multiple sampling moments; The energy consumption of the first server in the prediction period is predicted according to the power at the multiple sampling moments.

2. The method according to claim 1, characterized in that The load rate of the first server at the first sampling moment includes one or more of the processor usage, disk usage, memory usage, and network bandwidth usage of the first server at the first sampling moment, and the first sampling moment is one of the multiple sampling moments.

3. The method according to claim 2, characterized in that The processor usage rate is determined according to rated powers of multiple processors in the first server and usage rates of the multiple processors at the first sampling moment.

4. The method according to claim 3, characterized in that The multiple processors include one or more of a central processing unit, a graphics processing unit, a neural network processor, and a data processor.

5. The method according to claim 1, characterized in that The load rate of the first server at the first sampling moment is determined according to the system average load of the first server at the first sampling moment, wherein the first sampling moment is one sampling moment among the multiple sampling moments.

6. The method according to any one of claims 1 to 5, characterized in that The predicting the power of the first server at the multiple sampling moments according to the rated power, no-load power and load rate of the first server at the multiple sampling moments in the prediction period includes: The rated power, the no-load power and the load rate at the second sampling moment are input into a server power prediction model to predict the power of the first server at the second sampling moment, wherein the second sampling moment is one of the multiple sampling moments, and the server power consumption prediction model includes a curve fitting function, which is used to represent the relationship between the power consumption of the first server and the load rate of the first server.

7. The method according to claim 6, characterized in that The curve fitting function is constructed based on an exponential function and a sigmoid function.

8. The method according to claim 6 or 7, characterized in that: The method further comprises: Obtaining a load rate of the first server at a third sampling time, wherein the third sampling time is a sampling time among the multiple sampling times that is different from the second sampling time; Selecting a plurality of historical power data according to the load rate at the third sampling moment, wherein the plurality of historical power data are all historical powers of the first server at the load rate at the third sampling moment; The power of the first server at the third sampling moment is predicted according to the multiple historical power data.

9. The method according to any one of claims 1 to 8, characterized in that The first server operates in a first mode during the prediction period, and the no-load power and the energy consumption of the first server during the prediction period correspond to a second mode.

10. A server energy consumption prediction device, characterized in that: Including prediction module, The prediction module is used to predict the power of the first server at multiple sampling moments according to the rated power and no-load power of the first server and the load rate of the first server at multiple sampling moments in the prediction period, wherein the load rates at the multiple sampling moments correspond to the powers at the multiple sampling moments one by one; The prediction module is further used to predict the energy consumption of the first server in the prediction period according to the power at the multiple sampling moments.

11. The device according to claim 10, characterized in that The load rate of the first server at the first sampling moment includes one or more of the processor usage, disk usage, memory usage, and network bandwidth usage of the first server at the first sampling moment, and the first sampling moment is one of the multiple sampling moments.

12. The device according to claim 11, characterized in that The device also includes a determination module, which is used to determine the processor usage rate according to the rated power of multiple processors in the first server and the usage rate of the multiple processors at the first sampling time.

13. The device according to claim 12, characterized in that The multiple processors include one or more of a central processing unit, a graphics processing unit, a neural network processor, and a data processor.

14. The device according to claim 10, characterized in that The device also includes a determination module, which is used to determine the load rate of the first server at the first sampling moment according to the system average load of the first server at the first sampling moment, wherein the first sampling moment is one of the multiple sampling moments.

15. The device according to any one of claims 10 to 14, characterized in that The prediction module is also used to: The rated power, the no-load power and the load rate at the second sampling moment are input into a server power prediction model to predict the power of the first server at the second sampling moment, wherein the second sampling moment is one of the multiple sampling moments, and the server power consumption prediction model includes a curve fitting function, which is used to represent the relationship between the power consumption of the first server and the load rate of the first server.

16. The device according to claim 15, characterized in that The curve fitting function is constructed based on an exponential function and a sigmoid function.

17. The device according to claim 15 or 16, characterized in that The device also includes an acquisition module, The acquisition module is used to acquire the load rate of the first server at a third sampling time, wherein the third sampling time is a sampling time different from the second sampling time among the multiple sampling times; The acquisition module is further used to select a plurality of historical power data according to the load rate at the third sampling moment, wherein the plurality of historical power data are all historical powers of the first server at the load rate at the third sampling moment; The prediction module is further used to predict the power of the first server at the third sampling moment according to the multiple historical power data.

18. The device according to any one of claims 10 to 17, characterized in that The first server operates in a first mode during the prediction period, and the no-load power and the energy consumption of the first server during the prediction period correspond to a second mode.

19. A computing device, characterized in that The device comprises a processor and a memory, wherein the processor is configured to execute instructions stored in the memory so that the computing device executes the method according to any one of claims 1 to 9.

20. A computer-readable storage medium, characterized in that: The method comprises computer program instructions, and when the computer program instructions are executed by a computing device, the computing device performs the method according to any one of claims 1 to 9.

21. A computer program product comprising instructions, characterized in that When the instructions are executed by a computing device, the computing device is caused to perform the method according to any one of claims 1 to 9.

Images