A Server Power Prediction Method Based on CPU-GPU Coupled Load

Abstract

This invention discloses a method for predicting the overall power of a server based on CPU-GPU coupled load. First, a CPU-GPU coupled power prediction system is constructed, decomposing the server's overall power into idle power consumption, CPU-independent power consumption, and GPU-independent power consumption. An interpretable CPU-GPU coupled power consumption term is introduced to build an overall power prediction model. An automated parameter calibration mechanism is provided to complement the model. Based on measured overall power data under mixed load conditions, the optimal coupling coefficient is solved using the least squares method. Then, during runtime, real-time load data acquisition and measured overall power feedback are used to monitor prediction errors and detect anomalies. When the deviation meets preset thresholds and duration conditions, parameter updates are triggered to maintain stable prediction accuracy. This method can accurately quantify the interaction overhead between heterogeneous computing components, thereby significantly improving prediction accuracy and maintaining stability in scenarios with high concurrency of CPU and GPU operation and competition for interconnect and memory resources.

Description

Technical Field

[0001] This invention belongs to the field of server power prediction and energy management technology, specifically relating to a method for predicting the overall power of a server based on CPU-GPU coupled load. Background Technology

[0002] In data center environments, server power consumption is a crucial foundation for influencing energy efficiency, power supply capacity planning, data center cooling design, and power capping strategies. CPU-GPU heterogeneous architecture is a mainstream data center computing node design, integrating the logical control capabilities of a general-purpose processor (CPU) with the high-concurrency computing advantages of a graphics processing unit (GPU). It enables efficient data processing in complex tasks such as AI training, image processing, and large-scale scientific computing. Seamless data interaction between devices is achieved through PCIe buses or other high-speed interconnect structures, overcoming the computing bottlenecks of single architectures and supporting heterogeneous collaborative computing. Due to the superior performance and efficiency of heterogeneous computing mechanisms in various tasks, it is widely used in fields such as intelligent analysis and scientific exploration, especially providing an irreplaceable computing foundation in environments facing massive concurrent computing demands. However, with the large-scale deployment of CPU-GPU heterogeneous servers in data centers, the nonlinear coupling problem of overall power consumption characteristics has gradually become a key challenge restricting data center energy efficiency optimization. Due to the complex interaction between CPUs and GPUs in actual operation, accurate prediction of server power consumption faces significant challenges. Especially in environments where both CPU and GPU are under high load, the system not only generates its own basic power consumption, but also incurs additional coupled power consumption due to factors such as data interaction, cache consistency maintenance, memory bandwidth contention, and power management linkage. This complex hardware interaction not only breaks the traditional assumption of a simple linear superposition of CPU and GPU power, but may also lead to a significant increase in the overall power prediction error, resulting in wasted power supply capacity, over-supplied cooling system, and even inaccurate dynamic power capping control decisions.

[0003] Therefore, it is particularly important to conduct in-depth research on the power coupling mechanism and nonlinear characteristics during the operation of CPU-GPU heterogeneous servers, and to explore high-precision whole-machine power prediction methods. This can not only avoid the idle waste of data center power and heat dissipation resources, but also improve the reliability and allocation accuracy of power capping control strategies. Especially with the continuous expansion of heterogeneous computing power in data centers and the increasing complexity of computing tasks, the problem of whole-machine energy consumption management will become more multi-dimensional and dynamic. Only by constructing a high-precision, low-complexity, and real-time executable power prediction mechanism can we ensure the optimal energy efficiency and long-term stable operation of data centers in heterogeneous computing power environments.

[0004] For predicting the overall power consumption of heterogeneous servers, existing technologies typically employ empirical linear models or data-driven regression models based on hardware monitoring metrics. Specifically, conventional implementation methods mainly include:

[0005] (1) Feature fitting: Collect real-time status indicators such as CPU and GPU utilization, operating frequency and temperature, and use multiple linear regression or machine learning algorithms to directly fit the power output curve;

[0006] (2) Component superposition: Independent power consumption models for CPU and GPU are established separately, and the power consumption of the whole machine is calculated by simple linear superposition, while the static idle power of the system is used as a bias term.

[0007] While the aforementioned methods can achieve some effectiveness in certain load ranges, their practical applications suffer from the following significant limitations: Existing models generally treat the CPU and GPU as physically isolated and independent power contributors. Although some schemes attempt to introduce higher-order polynomials to statistically fit the nonlinear changes in power, they lack an explicit characterization of the system-level mechanism of "additional power consumption generated by CPU-GPU collaborative operation." This lack of modeling of the system-level collaborative mechanism significantly limits the prediction accuracy of existing technologies when facing complex computing tasks with high concurrency and high collaboration.

[0008] In summary, the existing technologies have the following shortcomings: (1) When the CPU and GPU run concurrently and exchange data, the simple superposition model is prone to significant underestimation or overestimation, and the error varies with the workload type and platform topology, making it difficult to correct with a single bias; (2) Although the complex regression model driven by pure data can fit local curves, it has weak interpretability and is difficult to migrate across platforms. It needs to be retrained or calibrated extensively after changes in hardware configuration or power management strategy; (3) Existing methods often rely on many monitoring features, resulting in high online computing and deployment costs, which makes it difficult to meet the low-latency application requirements of data centers for real-time dynamic scheduling and energy consumption control. Therefore, a new technology is urgently needed to solve the problem of inaccurate prediction of overall power of heterogeneous servers when the CPU and GPU are running in tandem. Summary of the Invention

[0009] To address the aforementioned technical issues, this invention provides a server power prediction method based on CPU-GPU coupled load. Without relying on complex feature engineering and large-scale training data, it establishes a server power prediction model by introducing an interpretable CPU-GPU coupled power consumption term. Furthermore, it proposes a basic parameter calibration method and runtime prediction process that matches this model, enabling the model to be quickly deployed and maintain stable prediction accuracy under different hardware platforms and topologies.

[0010] The technical solution adopted in this invention is: a method for predicting the overall power of a server based on CPU-GPU coupled load, which includes the following steps:

[0011] S1. Construct a CPU-GPU coupled power prediction system;

[0012] The CPU-GPU coupled power prediction system includes: a load acquisition module, a parameter calibration module, a power calculation module, an anomaly detection module, and an output interface module. Each module interacts with the other and operates in tandem through a standardized data interface.

[0013] The parameter calibration module includes a basic parameter calibration submodule and a coupling coefficient solution submodule. The power calculation module adopts a CPU-GPU coupled whole-system power prediction model that includes idle power terms, independent power consumption terms, and coupled power consumption terms.

[0014] The CPU-GPU coupled power prediction system is deployed on a CPU-GPU heterogeneous server and runs on the server operating system, underlying management controller, or local power management unit; its prediction results are output to the local power management platform or data center energy consumption control platform. The hardware infrastructure of the server includes: A CPUs, B GPUs, and a power measurement interface.

[0015] Where A=1, B≥1.

[0016] The power measurement interface is used to obtain the actual measured power of the server. The power measurement interface includes: onboard power sensor, BMC-side power reading, rack PDU, external power meter or electricity meter.

[0017] S2. Based on step S1, the CPU-GPU coupled power prediction system acquires CPU / GPU load-related index data and obtains the actual measured power of the entire server, forming standardized input data with timestamps.

[0018] The CPU-GPU coupled power prediction system obtains CPU / GPU load-related indicator data through the server operating system performance statistics interface, the underlying management controller interface, or the GPU driver management interface, and obtains the actual measured power of the server through the power measurement interface. The load acquisition module collects, aligns, and standardizes the load-related indicator data and the actual measured power of the server at a preset sampling period to form standardized input data with timestamps.

[0019] The load-related metrics include: CPU real-time load rate. GPU real-time load rate In a multi-GPU architecture, this also includes the independent real-time load rate of each GPU. The load acquisition module collects load-related metrics through the operating system interface, BMC interface, or GPU management interface, and outputs standardized load data including timestamps.

[0020] S3. Based on the standardized input data formed in step S2, the load-related index data and the actual measured power input parameter calibration module automatically calibrate and update the basic parameters and coupling coefficients of the CPU-GPU coupled whole machine power prediction model, and the power calculation module calls the updated model to calculate the real-time predicted power of the server.

[0021] First, on the target server, the inherent basic power parameters of the server hardware are obtained and calibrated through the basic parameter calibration submodule. During the calibration phase, the server is made to run under a fixed power supply strategy and a fixed heat dissipation strategy, and the background business load is controlled within a preset range.

[0022] The inherent basic power parameters of the server hardware include: server idle power. CPU net power at full load and the net power consumption of the GPU at full load .

[0023] The CPU-GPU coupled system power prediction model includes a basic model applicable to a single CPU + single GPU architecture, and an extended model applicable to a single CPU + N GPU architecture. The power calculation module calls the calibration results from the basic parameter calibration submodule and combines them with the standardized input data output by the load acquisition module, substituting them into the CPU-GPU coupled system power prediction model to calculate the real-time predicted power of the server.

[0024] Among them, the coupled power consumption term of the CPU-GPU coupled whole-system power prediction model is used to characterize the additional power change caused by the coordinated operation of the CPU and GPU.

[0025] Finally, the coupling coefficient calculation submodule uses the least squares method to solve for the optimal coupling coefficient based on measured data. Furthermore, the coupling coefficient is continuously updated using a dynamic optimization mechanism to achieve automated, high-precision calibration and long-term optimization of the coupling coefficient.

[0026] S4. Based on step S3, the anomaly detection module's matching model self-update mechanism monitors the effectiveness of the predicted power results and model bias, identifies anomaly types, and triggers corresponding processing mechanisms based on the anomaly types.

[0027] The anomaly identification includes numerical anomalies and model anomalies. The anomaly detection module takes into account the predicted power and the measured total power as inputs, and identifies model anomalies by calculating the relative error and combining it with continuous period criteria. The numerical anomalies include: abnormal load input, parameter out-of-bounds errors, and sudden changes in calculation results. The model anomalies include: the deviation between the predicted power and the measured power continuously exceeding the threshold.

[0028] S5. Based on step S4, the output interface module outputs the predicted power of the whole machine to the data center management and control platform or the local power management unit of the server, providing data support for energy consumption control, power capping and other strategies.

[0029] Furthermore, step S3 is specifically as follows:

[0030] S31. Initialization preparation and basic power parameter calibration are performed through the basic parameter calibration submodule;

[0031] First, perform initialization calibration on the target server to run it under preset power and cooling policies, and keep background service load within preset ranges. Then, perform basic power parameter calibration, including:

[0032] 1) Idle power calibration;

[0033] The CPU and GPU are kept in an idle state, meaning their respective load rates are close to zero. Server power data is collected continuously for a preset number of times, and the average value is calculated to obtain the server's idle power. .

[0034] set up This indicates the real-time load rate on the corresponding side. This indicates the no-load determination threshold; when the load rate of the corresponding side exceeds the threshold within multiple consecutive sampling periods. All are less than the preset no-load judgment threshold. If the load fluctuates, the load rate is averaged before a determination is made.

[0035] 2) CPU full-load net power calibration;

[0036] Keeping the GPU idle, the CPU load rate is gradually increased from low to full load in preset steps. The total server power consumption under full CPU load is collected, and then the total server power consumption is compared with the idle power consumption. The difference yields the net power of the CPU at full load. .

[0037] 3) GPU full-load net power calibration;

[0038] Keeping the CPU idle, the GPU load rate is gradually increased from low to full load in preset steps. The server's total power consumption under full GPU load is collected, and then the server's total power consumption is compared with the idle power consumption. The difference yields the net power of the GPU at full load. In a multi-GPU architecture, the net power at full load for each GPU is obtained. .

[0039] 4) Parameter storage;

[0040] The calibration results , , or Stored in the parameter library, and the stated , , or These are the inherent basic power parameters of the server hardware.

[0041] S32. Establish a CPU-GPU coupled whole-system power prediction model;

[0042] 1) Basic model of single CPU + single GPU architecture;

[0043] If the server used is a heterogeneous server with a single CPU and a single GPU, then the basic model of the single CPU and single GPU architecture is adopted. The overall power prediction model consists of idle power, CPU independent power consumption, GPU independent power consumption, and coupled power consumption. The mathematical expression of the model is as follows:

[0044] (1);

[0045] in, This represents the real-time predicted power of the entire server, which is the target output of the model. This represents the server's inherent power consumption when the CPU and GPU loads are both 0, indicating the server's idle power. This represents the net power consumption when the GPU load is 0 and the CPU is running at full load. This represents the net power consumption when the CPU load is 0 and the GPU is running at full load. This represents the real-time CPU load rate, with a value range of [0,1]. This represents the real-time GPU load rate, with a value range of [0,1]. This represents the CPU-GPU coupling coefficient. This is the coupling power consumption term.

[0046] 2) Single CPU + N GPU architecture extended model;

[0047] For a high-performance heterogeneous server with a single CPU and N GPUs, the basic model is extended using a single CPU + N GPU architecture. The mathematical expression of the model is as follows:

[0048] (2);

[0049] in, Indicates the first Net power consumption at full load for each GPU; Indicates the first The real-time load rate of each GPU, with a value range of [0,1]; This indicates the number of GPUs installed on the server.

[0050] S33. The coupling coefficient is solved and dynamically optimized through the coupling coefficient solution submodule, and the real-time predicted power of the server is output by the power calculation module.

[0051] Based on the basic power parameter calibration in step S31 and the overall power prediction model established in step S32, for mixed load scenarios where both the CPU and GPU are under non-zero load, the optimal coupling coefficient is solved based on the measured overall power data. .

[0052] For a single CPU + single GPU architecture, the coupling coefficient calculation expression is obtained by modifying the overall power prediction model as follows:

[0053] (3);

[0054] For a single CPU + N GPU architecture, the coupling coefficient calculation expression is obtained by transforming the overall power prediction extended model as follows:

[0055] (4);

[0056] in, This represents the actual measured power of the server as a whole, obtained through the power measurement interface under the corresponding mixed load scenario.

[0057] Then, the coupling coefficient calculation submodule generates calibration samples based on a preset mixed load combination. The calibration samples include: CPU real-time load rate. GPU real-time load rate or real-time load rate of each GPU Measured power of the server at the corresponding time And timestamp information.

[0058] Based on the calibration samples, the coupling coefficients are fitted using the least squares method to construct an error function. The expression is as follows:

[0059] (5);

[0060] in, Indicates the number of samples for mixed load calibration. Indicates the first Group of samples with coupling coefficient of The predicted power is calculated by the overall power prediction model. Indicates the first The measured power of the whole machine corresponding to the sample group.

[0061] By minimizing the error function The optimal coupling coefficient matching the target server hardware platform and topology is obtained. .when When the error is less than or equal to a preset error threshold, the coupling coefficient calibration is deemed to have passed, and the calibration is performed. Store in the parameter library; otherwise, trigger resampling, expand the mixed load sample points, or adjust the load combination and solve again.

[0062] Obtain the optimal coupling coefficient Then, the power calculation module calls the basic power parameters obtained from step S31 and the optimal coupling coefficient obtained from step S33. The standardized input data output by the load acquisition module is then substituted into the CPU-GPU coupled whole-machine power prediction model described in step S32 to calculate the real-time predicted power of the server. .

[0063] Furthermore, step S4 is specifically as follows:

[0064] S41. Anomaly detection rules;

[0065] The anomaly detection module detects the predicted power of the server, the actual power of the server, and the model's operating status in units of power prediction cycle, identifying two types of problems: numerical anomalies and model anomalies.

[0066] The inputs to the anomaly detection module include: the real-time predicted power of the entire server output in step S3. The actual measured power of the server as a whole, obtained through the power measurement interface. The load acquisition module outputs CPU / GPU load-related metrics and current model parameter status information.

[0067] The numerical anomalies include: abnormal load input, parameter out-of-bounds, and sudden changes in calculation results; the model anomalies include: the deviation between the predicted power and the actual measured power of the whole machine continuously exceeds a preset threshold.

[0068] The specific anomaly detection rules are as follows:

[0069] 1) Numerical anomaly detection;

[0070] If the predicted power If the power exceeds the server's allowed power range, or if there are abnormal changes in the prediction results of adjacent prediction periods, the values ​​are judged as abnormal.

[0071] Numerical anomalies are caused by hardware sensor failures, load acquisition errors, parameter failures, communication anomalies, or computing link anomalies.

[0072] 2) Model anomaly detection;

[0073] When the relative prediction error exceeds the preset error threshold for multiple consecutive prediction periods, the model is deemed abnormal.

[0074] The model anomaly is caused by inaccurate coupling coefficients, hardware aging, changes in heat dissipation environment, changes in power supply characteristics, or changes in server hardware topology.

[0075] S42, Exception handling and model self-updating process;

[0076] For different anomaly types, the system executes differentiated processing procedures. The specific anomaly handling and model self-update procedures are as follows:

[0077] 1) Anomaly detection;

[0078] The anomaly detection module receives the predicted power, the actual power of the whole machine and the model status information in real time, determines whether there is an anomaly according to the rules described in step S41, and outputs the anomaly type determination result.

[0079] 2) Handling numerical anomalies;

[0080] If the value is determined to be abnormal, the predicted power of the whole machine in the most recent effective prediction period or the predicted power after filtering correction is used as the alternative output value for the current period to ensure the continuity of power output; at the same time, the abnormal log is recorded and the status self-check of the load acquisition link, power measurement interface and related sensors is triggered.

[0081] After the numerical anomalies are eliminated, the system resumes using the current prediction model to output the real-time predicted power of the entire server.

[0082] 3) Model anomaly handling;

[0083] If a model is determined to be abnormal, a fast re-solution mechanism for the coupling coefficients is immediately triggered. This fast re-solution mechanism is based on recently collected core samples of mixed loads and re-solutions the coupling coefficients. The model parameters are then refitted and the updated coupling coefficients are written into the parameter library, thus completing the online update of the model parameters.

[0084] After the coupling coefficient is updated, the anomaly detection module continues to monitor the prediction error of the updated model. If the prediction error returns to the preset threshold range, the normal real-time prediction process is restored. If the prediction error continues to exceed the threshold, the full-process parameter recalibration is triggered.

[0085] 4) Full-process parameter recalibration;

[0086] If the model error still does not meet the requirements after rapid re-solution, or if the basic power parameters are found to be inaccurate, the system re-executes steps S31 to S33 to update the basic power parameters, coupling coefficients, and overall power prediction model throughout the entire process, thereby restoring the model prediction accuracy.

[0087] 5) Hardware adaptation and updates;

[0088] When the server detects changes in hardware topology or operating environment, it automatically triggers a full-process parameter recalibration. These changes include: increasing or decreasing the number of GPUs, replacing CPU or GPU hardware, firmware upgrades, power policy adjustments, changes in cooling strategies, or switching system operating modes. By recalibrating the basic power parameters and coupling coefficients, the model is re-adapted to the updated server hardware architecture and operating environment.

[0089] 6) Restore normal forecasting;

[0090] After the anomaly handling is completed and the model status returns to normal, the system resumes the real-time power prediction process in step S3, and the anomaly detection module continues to monitor the model status and prediction results.

[0091] Furthermore, the anomaly detection module continuously monitors the deviation between the predicted power and the actual measured power of the entire machine during server operation. When the average relative error of prediction over multiple consecutive sampling periods exceeds a preset threshold, a fast re-solution mechanism for the coupling coefficients is triggered. This fast re-solution mechanism refits the coupling coefficients based on recently collected core mixed load samples without performing full-process basic parameter calibration, obtaining updated optimal coupling coefficients and overwriting the original parameters in the parameter library. If the updated prediction error still does not return to the threshold range, a full-process parameter recalibration is triggered to achieve long-term stable operation of the model and maintain prediction accuracy.

[0092] The beneficial effects of this invention are as follows: The method of this invention first constructs a CPU-GPU coupled power prediction system, decomposing the server's overall power into idle power consumption, CPU-independent power consumption, and GPU-independent power consumption. Based on this, an interpretable CPU-GPU coupled power consumption term is introduced to characterize the additional power consumption generated by the CPU and GPU operating collaboratively, thereby constructing a deployable overall power prediction model. Simultaneously, an automated parameter calibration mechanism is provided to complement the model, and the optimal coupling coefficient is solved using the least squares method based on measured overall power data under mixed load conditions. Furthermore, during operation, real-time load acquisition and measured overall power feedback are used to monitor prediction errors and detect anomalies. When the deviation meets preset thresholds and duration conditions, parameter updates are triggered, enabling the model to be quickly deployed on different hardware platforms and topologies and maintain stable prediction accuracy over a long period. This invention's method, through explicit introduction of the CPU-GPU coupled power consumption term and mechanistic modeling, allows the system to accurately quantify the additional interaction overhead between heterogeneous computing components, achieving a significant improvement and maintenance of prediction accuracy in scenarios with high concurrency operation of CPUs and GPUs and competition for interconnect and memory resources. The method of this invention is based on the fact that the prediction model parameters have clear physical meanings. It designs a basic parameter calibration method that can be quickly completed with only a small number of experimental data points, and has good cross-hardware platform portability. Through a simplified runtime monitoring feature input design, the method of this invention allows the system to directly output the predicted total power value based solely on real-time CPU and GPU load rates, resulting in low computational overhead and suitability for online scheduling and energy management systems. Attached Figure Description

[0093] Figure 1 This is a flowchart of a server power prediction method based on CPU-GPU coupled load according to the present invention.

[0094] Figure 2 This is a flowchart of parameter calibration and coupling coefficient calculation in an embodiment of the present invention.

[0095] Figure 3 This is a flowchart of the anomaly detection and online update closed-loop process in an embodiment of the present invention. Detailed Implementation

[0096] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0097] like Figure 1 The flowchart of a server power prediction method based on CPU-GPU coupled load according to the present invention is shown below. The specific steps are as follows:

[0098] S1. Construct a CPU-GPU coupled power prediction system;

[0099] like Figure 1As shown, the CPU-GPU coupled power prediction system includes: a load acquisition module, a parameter calibration module, a power calculation module, an anomaly detection module, and an output interface module. Each module achieves data interaction and process linkage through a standardized data interface.

[0100] The parameter calibration module includes a basic parameter calibration submodule and a coupling coefficient solution submodule. The power calculation module adopts a CPU-GPU coupled whole-system power prediction model that includes idle power terms, independent power consumption terms, and coupled power consumption terms.

[0101] The CPU-GPU coupled power prediction system is deployed on a CPU-GPU heterogeneous server and runs on the server operating system, underlying management controller, or local power management unit; its prediction results are output to the local power management platform or data center energy consumption control platform. The hardware infrastructure of the server includes: A CPUs, B GPUs, and a power measurement interface.

[0102] Where A=1, B≥1.

[0103] The power measurement interface is used to obtain the actual measured power of the server. The power measurement interface includes: onboard power sensor, BMC-side power reading, rack PDU, external power meter or electricity meter.

[0104] S2. Based on step S1, the CPU-GPU coupled power prediction system acquires CPU / GPU load-related index data and obtains the actual measured power of the entire server, forming standardized input data with timestamps.

[0105] The CPU-GPU coupled power prediction system obtains CPU / GPU load-related indicator data through the server operating system performance statistics interface, the underlying management controller interface, or the GPU driver management interface, and obtains the actual measured power of the server through the power measurement interface. The load acquisition module collects, aligns, and standardizes the load-related indicator data and the actual measured power of the server at a preset sampling period (which can be milliseconds or higher) to form standardized input data with timestamps, which is used for subsequent model calibration, prediction calculation, result verification, and anomaly detection.

[0106] The load-related metrics include: CPU real-time load rate. GPU real-time load rate In a multi-GPU architecture, this also includes the independent real-time load rate of each GPU. The load acquisition module collects load-related metrics through the operating system interface, BMC interface, or GPU management interface, and outputs standardized load data including timestamps.

[0107] S3. Based on the standardized input data formed in step S2, the load-related index data and the measured power input parameter calibration module automatically calibrate and update the basic parameters and coupling coefficients of the CPU-GPU coupled whole machine power prediction model, and the power calculation module calls the updated model to calculate the real-time predicted power of the server.

[0108] First, on the target server, the inherent basic power parameters of the server hardware are obtained and calibrated through the basic parameter calibration submodule. To ensure the repeatability of the calibration, the server is operated under a fixed power supply strategy and a fixed heat dissipation strategy during the calibration phase, and the background service load is controlled within a preset range to reduce the impact of external disturbances on the parameter calibration results.

[0109] The inherent basic power parameters of the server hardware include: server idle power. CPU net power at full load and the net power consumption of the GPU at full load .

[0110] The CPU-GPU coupled system power prediction model includes a basic model applicable to a single CPU + single GPU architecture, and an extended model applicable to a single CPU + N GPU architecture. The power calculation module calls the calibration results from the basic parameter calibration submodule and combines them with the standardized input data output by the load acquisition module, substituting them into the CPU-GPU coupled system power prediction model to calculate the real-time predicted power of the server.

[0111] Among them, the coupled power consumption term of the CPU-GPU coupled whole machine power prediction model is used to characterize the additional power change caused by the coordinated operation of CPU and GPU, so as to improve the accuracy of server whole machine power prediction; the model only contains four arithmetic operations and simple summation, with a computational complexity of O(1), which can meet the millisecond-level real-time computing requirements of the server.

[0112] Finally, the coupling coefficient calculation submodule uses the least squares method to solve for the optimal coupling coefficient based on measured data. Furthermore, the coupling coefficient is continuously updated using a dynamic optimization mechanism, thereby achieving automated, high-precision calibration and long-term optimization of the coupling coefficient.

[0113] S4. Based on step S3, the anomaly detection module's matching model self-update mechanism monitors the effectiveness of the predicted power results and model bias, identifies anomaly types, and triggers corresponding processing mechanisms based on the anomaly types.

[0114] The anomaly identification includes numerical anomalies and model anomalies. The anomaly detection module takes into account the predicted power and the measured total power as inputs, and identifies model anomalies by calculating the relative error and combining it with continuous period criteria. The numerical anomalies include: abnormal load input, parameter out-of-bounds errors, and sudden changes in calculation results. The model anomalies include: the deviation between the predicted power and the measured power continuously exceeding the threshold.

[0115] S5. Based on step S4, the output interface module outputs the predicted power of the whole machine to the data center management and control platform or the local power management unit of the server, providing data support for energy consumption control, power capping and other strategies.

[0116] like Figure 2 As shown, in this embodiment, step S3 is specifically as follows:

[0117] S31. Initialization preparation and basic power parameter calibration are performed through the basic parameter calibration submodule;

[0118] Perform initial calibration preparation on the target server, run the server under the preset power policy and preset heat dissipation policy, and control the background business load within the preset range to reduce the impact of external disturbances on the calibration results.

[0119] The calibration of the basic power parameters includes:

[0120] 1) Idle power calibration;

[0121] The CPU and GPU are both kept in an idle state, meaning their respective load rates are close to zero. Server power data is collected continuously for a preset number of times (50 times in this example), and the average value is calculated to obtain the server's idle power. .

[0122] set up This indicates the real-time load rate on the corresponding side. This indicates the no-load determination threshold; when the load rate of the corresponding side is within multiple consecutive sampling periods (50 sampling periods in this embodiment), All are less than the preset no-load judgment threshold. If the load fluctuates, the load rate is averaged before a determination is made.

[0123] 2) CPU full-load net power calibration;

[0124] Keeping the GPU idle, the CPU load rate is gradually increased from low to high in preset steps until full load. In this embodiment, the CPU load is gradually increased from 0 to 100% (in 10% steps). The total power consumption of the server under full CPU load is collected, and then the total server power consumption is compared with the idle power consumption. The difference yields the net power of the CPU at full load. .

[0125] 3) GPU full-load net power calibration;

[0126] Keeping the CPU idle, the GPU load rate is gradually increased from low to high in preset steps until full load. In this embodiment, the GPU load is gradually increased from 0 to 100% (in 10% steps). The total server power consumption under full GPU load is collected, and then the total server power consumption and idle power consumption are compared. The difference yields the net power of the GPU at full load. In a multi-GPU architecture, the net power at full load for each GPU is obtained. .

[0127] 4) Parameter storage;

[0128] The calibration results , , or Store the data in the parameter library for subsequent calculation of coupling coefficients and prediction of overall power.

[0129] Among them, the , , or These are the inherent basic power parameters of the server hardware.

[0130] S32. Establish a CPU-GPU coupled whole-system power prediction model;

[0131] 1) Basic model of single CPU + single GPU architecture;

[0132] If the server used is a heterogeneous server with a single CPU and a single GPU, then the basic model of the single CPU and single GPU architecture is adopted. The overall power prediction model consists of idle power, CPU independent power consumption, GPU independent power consumption, and coupled power consumption. The mathematical expression of the model is as follows:

[0133] (1);

[0134] in, This represents the real-time predicted power of the entire server (unit: W), which is the target output of the model. This represents the server's idle power, with CPU and GPU loads both at 0 (unit: W). This represents the net power (in W) when the GPU load is 0 and the CPU is running at full load. This represents the net power (in W) of the GPU when it is running at full load with a CPU load of 0. This represents the real-time CPU load rate, with a value range of [0,1]. This represents the real-time GPU load rate, with a value range of [0,1]. represents the CPU-GPU coupling coefficient, a core parameter of the model, used to quantify the degree of additional power consumption impact caused by PCIe bus congestion, memory bandwidth contention, and power management coordination when the CPU and GPU are running in tandem. This coefficient is determined by the inherent properties of the server hardware architecture. In equation (1), The coupling power consumption term compensates for the shortcomings of existing technologies that ignore the CPU-GPU coupling effect, enabling the model to accurately adapt to high-load mixed operation scenarios.

[0135] 2) Single CPU + N GPU architecture extended model;

[0136] For high-performance heterogeneous servers with a single CPU and N GPUs, targeting scenarios such as deep learning training and supercomputing, the basic model is extended by considering load balancing among multiple GPUs and the overall coupling effect between multiple GPUs and the CPU. The extended model adopts a single CPU + N GPU architecture, and the mathematical expression of the model is as follows:

[0137] (2);

[0138] in, Indicates the first Net power of each GPU at full load (in W). Indicates the first The real-time load rate of each GPU, with a value range of [0,1]; This indicates the number of GPUs on the server. The extended model sums the independent power consumption terms of multiple GPUs and combines them with a unified coupling coefficient to characterize the coupled power consumption of multiple GPUs and the CPU. This eliminates the need to solve for the coupling coefficient separately for each GPU, ensuring the simplicity and computational efficiency of the model.

[0139] S33. The coupling coefficient is solved and dynamically optimized through the coupling coefficient solution submodule, and the real-time predicted power of the server is output by the power calculation module.

[0140] Based on the basic power parameter calibration in step S31 and the overall power prediction model established in step S32, for mixed load scenarios where both the CPU and GPU are under non-zero load, the optimal coupling coefficient is solved based on the measured overall power data. .

[0141] For a single CPU + single GPU architecture, the coupling coefficient calculation expression is obtained by modifying the overall power prediction model as follows:

[0142] (3);

[0143] For a single CPU + N GPU architecture, the coupling coefficient calculation expression is obtained by transforming the overall power prediction extended model as follows:

[0144] (4);

[0145] in, This represents the actual measured power of the server as a whole, obtained through the power measurement interface under the corresponding mixed load scenario, and is used to eliminate the acquisition noise of the hardware sensors.

[0146] Then, the coupling coefficient calculation submodule generates calibration samples based on a preset mixed load combination. The calibration samples include: CPU real-time load rate. GPU real-time load rate or real-time load rate of each GPU Measured power of the server at the corresponding time And timestamp information.

[0147] Based on the calibration samples, the coupling coefficients are fitted using the least squares method to construct an error function. The expression is as follows:

[0148] (5);

[0149] in, Indicates the number of samples for mixed load calibration. Indicates the first Group of samples with coupling coefficient of The predicted power is calculated by the overall power prediction model. Indicates the first The measured power of the whole machine corresponding to the sample group.

[0150] Then by minimizing the error function The optimal coupling coefficient matching the target server hardware platform and topology is obtained. .when When the error is less than or equal to a preset error threshold, the coupling coefficient calibration is deemed to have passed, and the calibration is performed. Store in the parameter library; otherwise, trigger resampling, expand the mixed load sample points, or adjust the load combination and solve again.

[0151] Obtain the optimal coupling coefficient Then, the power calculation module calls the basic power parameters obtained from step S31 and the optimal coupling coefficient obtained from step S33. The standardized input data output by the load acquisition module is then substituted into the CPU-GPU coupled whole-machine power prediction model described in step S32 to calculate the real-time predicted power of the server. .

[0152] like Figure 3 As shown, in this embodiment, step S4 is specifically as follows:

[0153] S41. Anomaly detection rules;

[0154] The anomaly detection module detects the predicted power of the server, the actual power of the server, and the model's operating status in units of power prediction cycle, identifying two types of problems: numerical anomalies and model anomalies.

[0155] The inputs to the anomaly detection module include: the real-time predicted power of the entire server output in step S3. The actual measured power of the server as a whole, obtained through the power measurement interface. The load acquisition module outputs CPU / GPU load-related metrics and current model parameter status information.

[0156] The numerical anomalies include: abnormal load input, parameter out-of-bounds, and sudden changes in calculation results; the model anomalies include: the deviation between the predicted power and the actual measured power of the whole machine continuously exceeds a preset threshold.

[0157] The specific anomaly detection rules are as follows:

[0158] 1) Numerical anomaly detection;

[0159] If the predicted power If the power exceeds the server's allowed power range, or if there are abnormal changes in the prediction results of adjacent prediction periods, the values ​​are judged as abnormal.

[0160] Numerical anomalies are caused by hardware sensor failures, load acquisition errors, parameter failures, communication anomalies, or computing link anomalies.

[0161] In practice, the allowable power range of the server can be determined based on the server's rated power, idle power, and operating power boundaries; abnormal changes can be determined based on the matching relationship between the predicted power change and the corresponding load change in adjacent periods.

[0162] 3) Model anomaly detection;

[0163] When the relative error of prediction (the relative difference between predicted power and measured power) is greater than the preset error threshold in multiple consecutive prediction periods, the model is judged to be abnormal.

[0164] The model anomaly is caused by inaccurate coupling coefficients, hardware aging, changes in heat dissipation environment, changes in power supply characteristics, or changes in server hardware topology.

[0165] In practice, the server's allowable power range, number of consecutive cycles, and error threshold can all be preset according to the target server platform and application scenario.

[0166] To address the issue of coupling effect changes caused by factors such as hardware aging, power management strategy adjustments, and driver version updates during long-term server operation, this embodiment designs a fast re-solution mechanism for coupling coefficients, triggered by the anomaly detection module. This eliminates the need for full-process parameter calibration, balancing optimization efficiency and prediction accuracy.

[0167] S42, Exception handling and model self-updating process;

[0168] The system executes differentiated processing procedures for different anomaly types to ensure prediction continuity and adaptive updating of model parameters under abnormal conditions.

[0169] The specific procedures for exception handling and model self-updating are as follows:

[0170] 1) Anomaly detection;

[0171] The anomaly detection module receives the predicted power, the actual power of the whole machine and the model status information in real time, determines whether there is an anomaly according to the rules described in step S41, and outputs the anomaly type determination result.

[0172] 2) Handling numerical anomalies;

[0173] If the value is determined to be abnormal, the predicted power of the whole machine in the most recent valid prediction period or the predicted power after filtering correction is used as the alternative output value for the current period to ensure the continuity of power output; at the same time, the abnormal log is recorded and the status self-check of the load acquisition link, power measurement interface and related sensors is triggered.

[0174] After the numerical anomalies are eliminated, the system resumes using the current prediction model to output the real-time predicted power of the entire server.

[0175] 3) Model anomaly handling;

[0176] If a model is determined to be abnormal, a fast re-solution mechanism for the coupling coefficients is immediately triggered. This fast re-solution mechanism is based on recently collected core samples of mixed loads and re-solutions the coupling coefficients. The model parameters are then refitted and the updated coupling coefficients are written into the parameter library, thus completing the online update of the model parameters.

[0177] After the coupling coefficient is updated, the anomaly detection module continues to monitor the prediction error of the updated model. If the prediction error returns to the preset threshold range, the normal real-time prediction process is restored. If the prediction error continues to exceed the threshold, the full-process parameter recalibration is triggered.

[0178] 4) Full-process parameter recalibration;

[0179] If the model error still does not meet the requirements after rapid re-solution, or if the basic power parameters are found to be inaccurate, the system re-executes steps S31 to S33 to update the basic power parameters, coupling coefficients, and overall power prediction model throughout the entire process to restore the model prediction accuracy.

[0180] 5) Hardware adaptation and updates;

[0181] When the server detects changes in hardware topology or operating environment, it automatically triggers a full-process parameter recalibration. These changes include: increasing or decreasing the number of GPUs, replacing CPU or GPU hardware, firmware upgrades, power policy adjustments, changes in cooling strategies, or switching system operating modes. By recalibrating the basic power parameters and coupling coefficients, the model is re-adapted to the updated server hardware architecture and operating environment.

[0182] 6) Restore normal forecasting;

[0183] After the anomaly handling is completed and the model status returns to normal, the system resumes the real-time power prediction process in step S3, and the anomaly detection module continues to monitor the model status and prediction results.

[0184] In this embodiment, the anomaly detection module continuously monitors the deviation between the predicted power and the actual measured power of the entire machine during server operation. When the average relative error of prediction over multiple consecutive sampling periods exceeds a preset threshold, a fast re-solution mechanism for the coupling coefficients is triggered. This fast re-solution mechanism refits the coupling coefficients based on recently collected core mixed load samples without performing full-process basic parameter calibration, obtaining updated optimal coupling coefficients and overwriting the original parameters in the parameter library. If the updated prediction error still does not return to the threshold range, a full-process parameter recalibration is triggered to achieve long-term stable operation of the model and maintain prediction accuracy.

[0185] Without departing from the basic idea of ​​the present invention—"characterizing the additional power consumption of CPU-GPU collaboration with coupling terms"—alternative implementations are also possible. For example, the coupling term can be extended from a product form to a function form that includes interconnect bandwidth utilization, PCIe transaction rate, or memory bandwidth usage, to further improve accuracy in data-driven scenarios; the coupling coefficient can be extended to a piecewise function related to temperature, frequency, or power limit states to adapt to different power management strategies; and the overall power measurement can be replaced by an external power meter with a BMC sensor and combined with a sensor calibration curve to reduce deployment costs.

[0186] In summary, compared to independent superposition models that ignore interaction terms, the method of this invention explicitly introduces a CPU-GPU coupled power consumption term, which enables stable prediction accuracy even when the CPU and GPU are running concurrently and there is interconnection and memory contention. Compared to purely data-driven black-box models, the parameters of the method of this invention have clear physical meanings, can be quickly calibrated with a small number of experimental points, and have good cross-platform portability. Compared to regression methods that require a large number of monitoring features, the method of this invention only needs CPU load rate and GPU load rate to output the predicted value of the whole machine power during runtime, with low computational overhead, and is suitable for online scheduling and energy management systems.

[0187] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for predicting the overall power consumption of a server based on CPU-GPU coupled load, comprising the following steps: S1. Construct a CPU-GPU coupled power prediction system; The CPU-GPU coupled power prediction system includes: Load acquisition module, parameter calibration module, power calculation module, anomaly detection module, and output interface module; Each module achieves data interaction and process linkage through standardized data interfaces; The parameter calibration module includes a basic parameter calibration submodule and a coupling coefficient solution submodule; the power calculation module adopts a CPU-GPU coupled whole-system power prediction model that includes idle power terms, independent power consumption terms and coupled power consumption terms. The CPU-GPU coupled power prediction system is deployed on a CPU-GPU heterogeneous server and runs on the server operating system, underlying management controller, or local power management unit; its prediction results are output to the local power management platform or data center energy consumption control platform; the hardware foundation of the server includes: A CPUs, B GPUs, and a power measurement interface; Where A=1, B≥1; The power measurement interface is used to obtain the actual measured power of the server. The power measurement interface includes: onboard power sensor, BMC-side power reading, rack PDU, external power meter or electricity meter. S2. Based on step S1, the CPU-GPU coupled power prediction system acquires CPU / GPU load-related index data and obtains the actual measured power of the entire server, forming standardized input data with timestamps. The CPU-GPU coupled power prediction system obtains CPU / GPU load-related indicator data through the server operating system performance statistics interface, the underlying management controller interface, or the GPU driver management interface, and obtains the actual measured power of the server through the power measurement interface; the load acquisition module collects, aligns, and standardizes the load-related indicator data and the actual measured power of the server at a preset sampling period to form standardized input data with timestamps; The load-related index data includes: a CPU real-time load rate , a GPU real-time load rate , and, in a multi-GPU architecture, an independent real-time load rate of each GPU ; the load collection module collects load-related indexes through an operating system interface, a BMC interface or a GPU management interface, and outputs standardized load data containing a time stamp; S3. Based on the standardized input data formed in step S2, the load-related index data and the actual measured power input parameter calibration module automatically calibrate and update the basic parameters and coupling coefficients of the CPU-GPU coupled whole machine power prediction model, and the power calculation module calls the updated model to calculate the real-time predicted power of the server. First, on the target server, the inherent basic power parameters of the server hardware are obtained and calibrated through the basic parameter calibration submodule. During the calibration phase, the server is made to run under a fixed power supply policy and a fixed heat dissipation policy, and the background business load is controlled within a preset range. The server hardware inherent base power parameter includes: server idle power , CPU full load net power , and GPU full load net power ; The CPU-GPU coupled whole-machine power prediction model includes a basic model applicable to a single CPU + single GPU architecture, and an extended model applicable to a single CPU + N GPU architecture. The power calculation module calls the calibration results of the basic parameter calibration submodule and combines them with the standardized input data output by the load acquisition module. The data is then substituted into the CPU-GPU coupled whole-machine power prediction model to calculate the real-time predicted power of the server. Among them, the coupled power consumption term in the CPU-GPU coupled whole-system power prediction model is used to characterize the additional power change caused by the coordinated operation of the CPU and GPU; Finally, the coupling coefficient solving submodule solves the optimal coupling coefficient based on the measured data using the least square method and continuously updates the coupling coefficient in combination with a dynamic optimization mechanism to realize automated, high-precision calibration and long-term optimization of the coupling coefficient. S4. Based on step S3, the anomaly detection module's matching model self-update mechanism monitors the effectiveness of the predicted power results and model bias, identifies anomaly types, and triggers corresponding processing mechanisms based on the anomaly types. The anomaly identification types include numerical anomalies and model anomalies; the anomaly detection module inputs the predicted power and the measured total power, and identifies model anomalies by calculating the relative error and combining it with continuous period criteria; the numerical anomalies include: load input anomalies, parameter out-of-bounds errors, and sudden changes in calculation results; the model anomalies include: the deviation between the predicted power and the measured power continuously exceeding the threshold. S5. Based on step S4, the output interface module outputs the predicted power of the whole machine to the data center management and control platform or the local power management unit of the server, providing data support for energy consumption control, power capping and other strategies.

2. The CPU-GPU coupled load based server whole machine power prediction method of claim 1, wherein, Step S3 is as follows: S31. Initialization preparation and basic power parameter calibration are performed through the basic parameter calibration submodule; First, perform initialization calibration on the target server to run it under preset power and cooling policies, and keep background service load within preset ranges; then, perform basic power parameter calibration, including: 1) Idle power calibration; The control CPU and the GPU are in an idle running state, i.e. the corresponding load rate is close to zero, the server whole machine power data is continuously collected for a preset number of times, and an average value is obtained, to obtain the server idle power ; Setting represents the real-time load rate of the corresponding side, represents the idle determination threshold; when the load rate of the corresponding side is less than the preset idle determination threshold for a plurality of consecutive sampling periods , it is determined that the side is in an idle state; if there is fluctuation in the load, the load rate is determined after using a moving average ; 2) CPU full-load net power calibration; The GPU is kept in an idle state, the CPU load rate is gradually increased from low to high by a preset step, the server whole machine power under the CPU full load state is collected, and then the CPU full load net power is obtained according to the difference between the server whole machine power and the idle power . ​ 3) GPU full-load net power calibration; Keeping the CPU idle, the GPU load rate is gradually increased from low to full load in preset steps. The server's total power consumption under full GPU load is collected, and then the server's total power consumption is compared with the idle power consumption. The difference yields the net power of the GPU at full load. In a multi-GPU architecture, the net power at full load for each GPU is obtained. ; 4) Parameter storage; The calibration results , , or Stored in the parameter library, and the stated , , or These are the inherent basic power parameters of the server hardware; S32. Establish a CPU-GPU coupled whole-system power prediction model; 1) Basic model of single CPU + single GPU architecture; If the server used is a heterogeneous server with a single CPU and a single GPU, then the basic model of the single CPU and single GPU architecture is adopted. The overall power prediction model consists of idle power, CPU independent power consumption, GPU independent power consumption, and coupled power consumption. The mathematical expression of the model is as follows: （1）； in, This represents the real-time predicted power of the entire server, which is the target output of the model. This represents the server's inherent power consumption when the CPU and GPU loads are both 0, indicating the server's idle power. This represents the net power consumption when the GPU load is 0 and the CPU is running at full load. This represents the net power consumption when the CPU load is 0 and the GPU is running at full load. This represents the real-time CPU load rate, with a value range of [0,1]. This represents the real-time GPU load rate, with a value range of [0,1]. Indicates the CPU-GPU coupling coefficient; This is the coupling power consumption term; 2) Single CPU + N GPU architecture extended model; For a high-performance heterogeneous server with a single CPU and N GPUs, the basic model is extended using a single CPU + N GPU architecture. The mathematical expression of the model is as follows: （2）； in, Indicates the first Net power consumption at full load for each GPU; Indicates the first The real-time load rate of each GPU, with a value range of [0,1]; Indicates the number of GPUs on the server; S33. The coupling coefficient is solved and dynamically optimized through the coupling coefficient solution submodule, and the real-time predicted power of the server is output by the power calculation module. Based on the basic power parameter calibration in step S31 and the overall power prediction model established in step S32, for mixed load scenarios where both the CPU and GPU are under non-zero load, the optimal coupling coefficient is solved based on the measured overall power data. ; For a single CPU + single GPU architecture, the coupling coefficient calculation expression is obtained by modifying the overall power prediction model as follows: （3）； For a single CPU + N GPU architecture, the coupling coefficient calculation expression is obtained by transforming the overall power prediction extended model as follows: （4）； in, This indicates the actual measured power of the server as a whole, obtained through the power measurement interface under the corresponding mixed load scenario. Then, the coupling coefficient calculation submodule generates calibration samples based on a preset mixed load combination. The calibration samples include: CPU real-time load rate. GPU real-time load rate or real-time load rate of each GPU Measured power of the server at the corresponding time And timestamp information; Based on the calibration samples, the coupling coefficients are fitted using the least squares method to construct an error function. The expression is as follows: （5）； in, Indicates the number of samples for mixed load calibration. Indicates the first Group of samples with coupling coefficient of The predicted power is calculated by the overall power prediction model. Indicates the first The measured power of the entire machine corresponding to each sample group; By minimizing the error function The optimal coupling coefficient matching the target server hardware platform and topology is obtained. ;when When the error is less than or equal to a preset error threshold, the coupling coefficient calibration is deemed to have passed, and the calibration is performed. Store in the parameter library; otherwise, trigger resampling, expand the mixed load sample points, or adjust the load combination and solve again. Obtain the optimal coupling coefficient Then, the power calculation module calls the basic power parameters obtained from step S31 and the optimal coupling coefficient obtained from step S33. The standardized input data output by the load acquisition module is then substituted into the CPU-GPU coupled whole-machine power prediction model described in step S32 to calculate the real-time predicted power of the server. .

3. The server power prediction method based on CPU-GPU coupled load according to claim 2, characterized in that, Step S4 is as follows: S41. Anomaly detection rules; The anomaly detection module detects the predicted power of the server, the actual power of the server, and the model running status in units of power prediction cycle, and identifies two types of problems: numerical anomalies and model anomalies. The inputs to the anomaly detection module include: the real-time predicted power of the entire server output in step S3. The actual measured power of the server as a whole, obtained through the power measurement interface. The load acquisition module outputs CPU / GPU load-related metrics and current model parameter status information; The numerical anomalies include: abnormal load input, parameter out-of-bounds errors, and sudden changes in calculation results; the model anomalies include: the deviation between the predicted power and the actual measured power of the whole machine continuously exceeds a preset threshold. The specific anomaly detection rules are as follows: 1) Numerical anomaly detection; If the predicted power If the power exceeds the server's allowable range, or if there is an abnormal change in the prediction results of adjacent prediction periods, it is judged as a numerical anomaly. Numerical anomalies are caused by hardware sensor failures, load acquisition errors, parameter failures, communication anomalies, or computing link anomalies. 2) Model anomaly detection; When the relative prediction error in multiple consecutive prediction periods is greater than the preset error threshold, the model is judged to be abnormal. The model anomaly is caused by inaccurate coupling coefficients, hardware aging, changes in heat dissipation environment, changes in power supply characteristics, or changes in server hardware topology. S42, Exception handling and model self-updating process; For different anomaly types, the system executes differentiated processing procedures. The specific anomaly handling and model self-update procedures are as follows: 1) Anomaly detection; The anomaly detection module receives the predicted power, the actual power of the whole machine and the model status information in real time, determines whether there is an anomaly according to the rules described in step S41, and outputs the anomaly type determination result. 2) Handling numerical anomalies; If the value is determined to be abnormal, the predicted power of the whole machine in the most recent valid prediction period or the predicted power after filtering correction is used as the alternative output value for the current period to ensure the continuity of power output; at the same time, the abnormal log is recorded and the status self-check of the load acquisition link, power measurement interface and related sensors is triggered. After the numerical anomalies are eliminated, the system resumes using the current prediction model to output the real-time predicted power of the entire server. 3) Model anomaly handling; If the model is determined to be abnormal, a fast re-solution mechanism for the coupling coefficients is immediately triggered; this fast re-solution mechanism is based on recently collected core samples of mixed loads and re-solutions the coupling coefficients. The model parameters are updated online by refitting and solving the problem, and the updated coupling coefficients are written into the parameter library. After the coupling coefficient is updated, the anomaly detection module continues to monitor the prediction error of the updated model. If the prediction error returns to the preset threshold range, the normal real-time prediction process is restored. If the prediction error continues to exceed the threshold, the full-process parameter recalibration is triggered. 4) Full-process parameter recalibration; If the model error still does not meet the requirements after rapid re-solution, or if the basic power parameters are found to be inaccurate, the system re-executes steps S31 to S33 to update the basic power parameters, coupling coefficients and whole machine power prediction model throughout the entire process to restore the model prediction accuracy. 5) Hardware adaptation and updates; When the server detects changes in hardware topology or operating environment, it automatically triggers full-process parameter recalibration. These changes include: increasing or decreasing the number of GPUs, replacing CPU or GPU hardware, firmware upgrades, adjusting power policies, changing heat dissipation policies, or switching system operating modes. By recalibrating the basic power parameters and coupling coefficients, the model is re-adapted to the updated server hardware architecture and operating environment. 6) Restore normal forecasting; After the anomaly handling is completed and the model status returns to normal, the system resumes the real-time power prediction process in step S3, and the anomaly detection module continues to monitor the model status and prediction results.

4. The server power prediction method based on CPU-GPU coupled load according to claim 3, characterized in that, The anomaly detection module continuously monitors the deviation between the predicted power and the actual measured power of the whole machine during server operation. When the average relative error of prediction for multiple consecutive sampling periods exceeds a preset threshold, a fast re-solution mechanism for coupling coefficients is triggered. The fast re-solution mechanism is to refit the coupling coefficients based on the recently collected core mixed load samples without performing full-process basic parameter calibration, obtain the updated optimal coupling coefficients, and overwrite the original parameters in the parameter library. If the prediction error still does not return to the threshold range after the update, full-process parameter recalibration is triggered.

Images