Thermal-aware and energy consumption prediction based server bmc co-optimization method and system

Abstract

The application discloses a server BMC cooperative optimization method and system based on heat sensing and energy consumption prediction, relates to the technical field of server hardware management and data center energy saving, and discloses a server BMC cooperative optimization method and system based on heat sensing and energy consumption prediction, which comprises the following steps: a heat-energy coupling model is established through multidimensional data acquisition and fusion analysis; precise sensing is realized by adopting three-dimensional heat distribution calculation and LSTM prediction; heat dissipation and power supply parameters are dynamically adjusted in combination with a multi-objective optimization algorithm; a closed-loop feedback mechanism is formed; the energy efficiency ratio of the server can be improved; in addition, dynamic cooperative optimization of heat dissipation control and energy consumption management, improvement of server heat dissipation efficiency and prolongation of the service life of hardware can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of server hardware management and data center energy saving, and particularly relates to a server BMC collaborative optimization method and system based on heat perception and energy consumption prediction. BACKGROUND

[0002] With the continuous expansion of data centers, the problems of server energy consumption and heat dissipation are increasingly serious. As the core of hardware management, the server BMC is responsible for monitoring hardware status, adjusting heat dissipation equipment (such as fans), and managing power supply, but the traditional BMC has obvious limitations in energy consumption and heat dissipation control. The current data center is generally faced with the dual challenges of high server cluster energy consumption and low heat dissipation efficiency, which not only increases the operating cost, but also affects the hardware reliability and service life. The existing technology mainly manages through preset threshold strategy, and this extensive control method has many defects. In terms of heat dissipation control, the traditional scheme sets a fixed fan speed gear according to the hardware temperature, and the fan speed is passively increased when the temperature exceeds a certain threshold. This single threshold response mechanism cannot perceive the complex heat distribution status inside the server, resulting in uneven allocation of heat dissipation resources: on the one hand, it may cause local hardware overheating, and on the other hand, it may cause the fan in the non-hot spot area to idle and consume energy. In terms of energy consumption management, the existing scheme mostly adopts a static power distribution strategy, such as distributing power according to the rated power of the server, which lacks dynamic response capability to real-time load fluctuations and future energy consumption trends.

[0003] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a server BMC collaborative optimization method and system based on heat perception and energy consumption prediction, aiming to improve the energy efficiency ratio of the server.

[0005] To achieve the above purpose, the present application provides a server BMC collaborative optimization method based on heat perception and energy consumption prediction, which comprises:

[0006] The temperature sensor deployed on the key hardware of the server is used to obtain the original temperature data, the heat flow sensor is used to obtain the original heat flux density data, the baseboard management controller (BMC) is used to obtain the real-time load data, historical load data and task queue information, and the computer room environment monitoring system is used to obtain the environmental parameters; and the original temperature data and the original heat flux density data are preprocessed to obtain the processed temperature data and the processed heat flux density data;

[0007] Based on the processed temperature data, the processed heat flux density data and the real-time load data, a heat perception model is used for analysis and processing to obtain heat bottleneck data and temperature prediction data;

[0008] performing prediction processing by using an energy consumption prediction model based on the real-time load data, historical load data, task queue information and environment parameters, to obtain predicted energy consumption data;

[0009] performing a collaborative optimization control strategy based on the thermal bottleneck data and temperature prediction data, and the predicted energy consumption data, to obtain an optimized fan rotating speed parameter and an optimized power distribution parameter;

[0010] performing the optimized fan rotating speed parameter by the BMC to control the server fan rotating speed, and performing the optimized power distribution parameter to adjust the power supply of each hardware component.

[0011] In an embodiment, the step of performing analysis processing by using a thermal-aware model based on the processed temperature data, processed heat flux density data and real-time load data, to obtain thermal bottleneck data and temperature prediction data comprises:

[0012] performing three-dimensional grid division processing on the processed temperature data, processed heat flux density data and real-time load data, to divide the internal space of the server into a plurality of grid units and map to corresponding hardware components;

[0013] performing thermal conduction calculation processing on each grid unit based on the principle of physical heat conduction, to generate three-dimensional thermal distribution data;

[0014] performing temperature change trend prediction processing based on the three-dimensional thermal distribution data, to obtain temperature prediction data; and performing thermal bottleneck identification processing based on the three-dimensional thermal distribution data and the processed heat flux density data, to obtain thermal bottleneck data.

[0015] In an embodiment, the step of performing thermal bottleneck identification processing based on the three-dimensional thermal distribution data and the processed heat flux density data, to obtain thermal bottleneck data comprises:

[0016] calculating thermal load data of each hardware component based on the three-dimensional thermal distribution data;

[0017] identifying grid units with temperature exceeding a hardware safety threshold based on the thermal load data, to obtain hotspot position data;

[0018] based on the processed heat flux density data, if the hotspot position data has not been relieved for a plurality of time steps and the heat flux density exceeds a preset threshold, determining the hotspot position data as the thermal bottleneck data.

[0019] In an embodiment, the energy consumption prediction model comprises a pre-trained long short-term memory network model; the step of performing prediction processing on the real-time load data, historical load data, task queue information and environmental parameters by using the energy consumption prediction model to obtain predicted energy consumption data comprises:

[0020] performing time sequence alignment processing and normalization processing on the real-time load data, historical load data, task queue information and environmental parameters to obtain processed time sequence feature data;

[0021] inputting the processed time sequence feature data into the pre-trained long short-term memory network model to perform time sequence analysis processing to output a predicted energy consumption sequence of a future time period.

[0022] In an embodiment, the method further comprises:

[0023] calculating confidence data of the predicted energy consumption sequence based on historical prediction error data recorded by the pre-trained long short-term memory network model during a training stage;

[0024] if the confidence data is lower than a preset value, calculating a thermal influence quantitative value according to the thermal bottleneck data;

[0025] inputting the thermal influence quantitative value as a dynamic correction factor into the energy consumption prediction model to regenerate a corrected predicted energy consumption sequence.

[0026] In an embodiment, the step of performing collaborative optimization control strategy based on the thermal bottleneck data and temperature prediction data, and the predicted energy consumption data to obtain optimized fan rotation speed parameters and optimized power distribution parameters comprises:

[0027] generating fan speed adjustment suggestion data based on the thermal bottleneck data and temperature prediction data;

[0028] generating power distribution suggestion data based on the predicted energy consumption data;

[0029] performing parameter optimization processing on the fan speed adjustment suggestion data and power distribution suggestion data by using a multi-objective optimization algorithm to obtain optimized fan rotation speed parameters and optimized power distribution parameters.

[0030] In an embodiment, the step of generating fan speed adjustment suggestion data based on the thermal bottleneck data and temperature prediction data comprises:

[0031] determining a hardware region to be adjusted preferentially based on the thermal bottleneck data;

[0032] combining the temperature prediction data to generate fan rotation speed adjustment suggestions in advance to obtain fan speed adjustment suggestion data.

[0033] In an embodiment, the step of generating power distribution suggestion data based on the predicted energy consumption data comprises:

[0034] analyzing future load trends of each hardware component based on the predicted energy consumption data;

[0035] generating suggestions of dynamically adjusting power supply voltage and current to obtain power distribution suggestion data.

[0036] In an embodiment, after the step of performing the optimized fan speed parameter by the BMC to control the server fan speed and performing the optimized power distribution parameter to adjust power supply of each hardware component, the method further comprises:

[0037] monitoring temperature data and energy consumption data after execution;

[0038] feeding back the temperature data and energy consumption data to the thermal-aware model and the energy consumption prediction model to update model parameters of the thermal-aware model and the energy consumption prediction model.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes a server BMC collaborative optimization system based on thermal awareness and energy consumption prediction, which comprises a memory, a processor and a server BMC collaborative optimization program based on thermal awareness and energy consumption prediction stored on the memory and executable on the processor, and the server BMC collaborative optimization program based on thermal awareness and energy consumption prediction is configured to implement the steps of the server BMC collaborative optimization method based on thermal awareness and energy consumption prediction.

[0040] The server BMC collaborative optimization method and system based on thermal awareness and energy consumption prediction proposed in the present application can improve the energy efficiency ratio of the server by establishing a thermal-energy coupling model through multi-dimensional data acquisition and fusion analysis, realizing accurate perception by three-dimensional heat distribution calculation and LSTM prediction, dynamically adjusting the heat dissipation and power supply parameters by combining a multi-objective optimization algorithm, forming a closed-loop feedback mechanism, in addition to being able to realize dynamic collaborative optimization of heat dissipation control and energy consumption management, improve the server heat dissipation efficiency and prolong the service life of the hardware. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings incorporated in and forming a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the application.

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings can also provide further drawings based on these drawings for those of ordinary skill in the art without any creative effort.

[0043] Figure 1 The flowchart provided by an embodiment of the server BMC collaborative optimization method based on thermal perception and energy consumption prediction of the present application is shown in the figure.

[0044] Figure 2 The detailed flowchart of step S200 in the present application Figure 1

[0045] Figure 3 The detailed flowchart of step S230 in the present application Figure 2

[0046] Figure 4 The detailed flowchart of step S300 in the present application Figure 1

[0047] The flowchart provided by another embodiment of the server BMC collaborative optimization method based on thermal perception and energy consumption prediction of the present application is shown in the figure. Figure 5

[0048] The detailed flowchart of step S400 in the present application Figure 6 Figure 1

[0049] The flowchart provided by still another embodiment of the server BMC collaborative optimization method based on thermal perception and energy consumption prediction of the present application is shown in the figure. Figure 7

[0050] The structural diagram provided by an embodiment of the server BMC collaborative optimization system based on thermal perception and energy consumption prediction of the present application is shown in the figure. Figure 8

[0051] Explanation of the reference signs:

[0052] 10, memory; 20, processor.

[0053] The purpose of the present application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0054] ​​​​The technical solutions in the present application will be described clearly and completely in the present application combined with the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. The components of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] It should be understood that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0056] In the prior art, data center servers have long faced the challenge of balancing heat dissipation efficiency and energy consumption management. Traditional baseboard management controllers rely on fixed temperature thresholds to adjust fan speed, and use static power distribution strategies, resulting in coexistence of local overheating and waste of power resources. When the server runs high-load tasks, the CPU and GPU area temperature rises rapidly, but the fan cannot cool down in time due to response lag, while during low-load periods, it runs continuously at high speed, resulting in invalid energy consumption. When the room temperature fluctuates, the existing system lacks dynamic perception of heat flux density changes, making it difficult to achieve precise cooling control.

[0057] In order to solve the above problems, the research and development team observed the spatiotemporal correlation between uneven hardware temperature distribution and energy consumption fluctuations, and proposed a technical route combining thermodynamic analysis and energy consumption prediction. First, to address the problem of single thermal perception data, a multi-dimensional sensor data collection scheme was designed to capture heat transfer paths between hardware through heat flux density monitoring. Second, it was found that static cooling strategies cannot adapt to dynamic load changes, so a three-dimensional temperature prediction mechanism based on physical heat conduction model was constructed. At the same time, it was identified that energy consumption prediction needs to be combined with task queue characteristics, and a time series model was used to analyze load trends. Finally, it was established that dynamic balance of cooling control and power distribution can be achieved through a collaborative optimization algorithm.

[0058] Based on this, the embodiments of the present application provide a server BMC collaborative optimization method based on thermal perception and energy consumption prediction, referring to Figure 1 The server BMC collaborative optimization method based on thermal perception and energy consumption prediction comprises steps S100-S500, wherein:

[0059] In step S100, raw temperature data is acquired by a temperature sensor deployed on a server key hardware, raw heat flux density data is acquired by a heat flow sensor, real-time load data, historical load data and task queue information are acquired by a baseboard management controller (BMC), and environment parameters are acquired by a computer room environment monitoring system; and the raw temperature data and the raw heat flux density data are preprocessed to obtain processed temperature data and processed heat flux density data.

[0060] In step S200, based on the processed temperature data, the processed heat flux density data and the real-time load data, a thermal perception model is used for analysis and processing to acquire thermal bottleneck data and temperature prediction data.

[0061] In step S300, based on the real-time load data, the historical load data, the task queue information and the environment parameters, an energy consumption prediction model is used for prediction processing to acquire predicted energy consumption data.

[0062] In step S400, based on the thermal bottleneck data and the temperature prediction data and the predicted energy consumption data, a collaborative optimization control strategy is executed to acquire optimized fan rotating speed parameters and optimized power distribution parameters.

[0063] In step S500, the optimized fan rotating speed parameters are executed by the BMC to control the server fan rotating speed, and the optimized power distribution parameters are executed to adjust the power supply of each hardware component.

[0064] In this embodiment, the temperature sensor refers to a thermocouple or a digital temperature sensor installed on the surface of a CPU, a GPU, a memory module and a power module, which can be implemented by a DS18B20 sensor to capture the change of the surface temperature of the hardware. The heat flow sensor refers to a heat flux detection device embedded in the heat dissipation path of the hardware, which can be implemented by a thin film heat flow sensor to measure the heat transfer rate per unit area. The preprocessing refers to filtering and denoising and removing outliers of the raw data, which can be implemented by a sliding average filtering algorithm to eliminate the interference of the sensor signal. The thermal perception model refers to a three-dimensional thermal field calculation model based on finite element analysis, which can be constructed by COMSOL Multiphysics software to simulate the heat conduction process inside the server. The energy consumption prediction model refers to a neural network model with time series processing capability, which can be implemented by an LSTM network structure to learn the mapping relationship between the load characteristics and the energy consumption. The collaborative optimization control strategy refers to a multi-objective optimization algorithm balancing the heat dissipation demand and the energy consumption target, which can be implemented by an NSGA-II genetic algorithm to generate a Pareto optimal solution set.

[0065] In this embodiment, the temperature sensor array collects the surface temperature of the CPU, GPU and other core components in real time, and the heat flow sensor monitors the heat transfer efficiency between the heat sink and the case. The collected raw data is processed by Kalman filtering to eliminate signal fluctuations caused by environmental electromagnetic interference. The preprocessed temperature data and heat flux density data are input into the thermal perception model. The model discretizes the internal space of the server into cubic units through three-dimensional grid division, and calculates the heat conduction equation of each unit combined with the thermal conductivity coefficient of the hardware material. When the temperature gradient of a certain area exceeds the safety threshold and the heat flux density is continuously abnormal, it is marked as a thermal bottleneck area. The energy consumption prediction model synchronously analyzes the task queue characteristics within the next five minutes, and predicts the energy consumption trend of each hardware component combined with the historical load pattern. The optimization algorithm dynamically calculates the fan speed combination and power supply parameters according to the distribution of the thermal bottleneck position and the predicted energy consumption curve, so that the high heat generation area obtains targeted cooling, and the power supply voltage of the low load hardware is reduced.

[0066] In this embodiment, a three-dimensional thermal field model is constructed by fusing heat flux density data, which can accurately identify hidden thermal conduction bottlenecks. Unlike static threshold control methods, predictive temperature analysis is used to adjust the cooling strategy in advance, avoiding temperature oscillation caused by response lag. Compared with independently running cooling and energy management systems, the collaborative optimization algorithm realizes dynamic matching of fan power consumption and hardware power supply, for example, when it is predicted that the GPU will enter a high load state, the fan speed in the corresponding area is increased in advance and the power supply phase is adjusted, preventing overheating and reducing energy waste. In this way, the present application realizes accurate monitoring and prediction of the internal thermal distribution of the server, effectively eliminating the risk of local overheating; the dynamic power distribution strategy adjusts the power supply parameters in real time according to the load changes, reducing unnecessary energy consumption; the collaborative action of the thermal perception model and the energy consumption prediction model enables the cooling system to respond to temperature change trends in advance, reducing energy loss caused by frequent fan start-stop; the multi-objective optimization algorithm balances cooling efficiency and energy consumption while ensuring safe operation of the hardware, overall improving the economy and reliability of server operation.

[0067] In a possible implementation, with reference to Figure 2 , step S200 includes steps S210-S230, in which:

[0068] Step S210, performing three-dimensional grid division processing on the processed temperature data, processed heat flux density data and real-time load data, dividing the internal space of the server into a plurality of grid units and mapping to corresponding hardware components;

[0069] Step S220, based on the principle of physical heat conduction, performing heat conduction calculation processing on each grid unit to generate three-dimensional thermal distribution data;

[0070] At step S230, temperature change trend prediction processing is performed based on the three-dimensional heat distribution data to obtain temperature prediction data; and heat bottleneck identification processing is performed based on the three-dimensional heat distribution data and the processed heat flux density data to obtain heat bottleneck data.

[0071] In this embodiment, the three-dimensional grid division processing refers to discretizing the internal space of the server into a plurality of regularly arranged three-dimensional units, which can be specifically implemented by using a finite element grid generation algorithm. Each grid unit is mapped to the position coordinates of hardware components such as CPUs, GPUs, and memory slots, so that temperature data can be located to a specific hardware region. The physical heat conduction principle calculation processing refers to establishing a heat conduction differential equation based on Fourier's law, and specifically can use a finite difference method to iteratively calculate each grid unit, simulate the dynamic process of heat conduction in the internal space of the server by solving the temperature gradient and heat flow relationship between adjacent grid units. The temperature change trend prediction processing refers to extrapolating calculation of the three-dimensional heat distribution data by a time series analysis model, which can specifically use an autoregressive integrated moving average model to predict the temperature fluctuation trajectory of each grid unit in the future time step in combination with the power consumption change rate parameter in the real-time load data. The heat bottleneck identification processing refers to coupling analysis of heat flux density and temperature gradient, which can specifically use a thermal resistance network model to calculate the heat flow impedance value of each grid unit. When the heat flux density of a certain region is continuously higher than the critical threshold and the temperature gradient shows nonlinear growth, it is determined that the region is a heat bottleneck.

[0072] In this embodiment, the internal space of the server is first divided into a plurality of cubic units by three-dimensional grid division, and each unit is associated with specific hardware such as CPU heat sinks and power modules. Then, the heat exchange between each unit and its adjacent units is calculated based on the heat conduction equation to generate a three-dimensional heat field model containing temperature gradient and heat flow direction. Based on this model, the correlation between historical temperature data and real-time load is analyzed by a time series prediction algorithm to deduce the temperature change curve of each hardware region in the next five minutes. At the same time, the density data collected by the heat flow sensor is superimposed with the three-dimensional heat distribution to identify the region where the heat flow is blocked and the temperature continuously rises, such as the local high-temperature area formed by the airflow blockage between the memory slots.

[0073] In this embodiment, the thermal state of each hardware region is accurately quantified by three-dimensional thermal field modeling, such as identifying the abnormal thermal flow impedance of the GPU backplane caused by poor contact of the cooling fins, so as to specifically enhance the air volume of this region. In addition, the scheme adjusts the cooling strategy in advance through temperature trend prediction before the hardware load mutation, such as increasing the speed curve of the corresponding cooling fan in advance when it is predicted that the temperature of the SSD controller will exceed the threshold value after 30 seconds. Through the above technical solutions, the present application realizes the fine analysis of the internal thermal field of the server, and solves the problem of mismatching of cooling resources caused by uneven heat distribution. The heat accumulation process of the hardware gap is accurately restored through three-dimensional heat conduction calculation, such as detecting the heat retention area formed between the mainboard power supply module and the side wall of the case due to narrow space. Combined with the time sequence prediction model, the temperature change direction can be predicted at the initial stage of hardware load fluctuation, such as generating cooling control instructions in advance to avoid temperature overshoot when the CPU utilization increases suddenly due to virtual machine migration. The coupling analysis of heat flux density and temperature gradient effectively distinguishes between temporary hot spots and structural thermal bottlenecks, such as identifying the persistent heat resistance increase caused by aging of the thermal paste, rather than the temporary temperature rise caused by instantaneous high load.

[0074] In a feasible implementation, with reference to Figure 3 , step S230 includes steps S231-S233, wherein:

[0075] Step S231, based on the three-dimensional thermal distribution data, calculating the thermal load data of each hardware component;

[0076] Step S231, based on the thermal load data, identifying the grid unit whose temperature exceeds the hardware safety threshold, and obtaining the hot spot position data;

[0077] Step S231, based on the processed heat flux density data, if the hot spot position data has not been relieved for a plurality of time steps and the heat flux density exceeds the preset threshold, it is determined as thermal bottleneck data.

[0078] In this embodiment, the thermal load data refers to the difference between the heat generated by the hardware component per unit time and its cooling capacity, which can be calculated by using the heat conduction equation combined with the temperature rise rate of the three-dimensional grid unit, and is used to quantify the difference in cooling demand of different hardware regions. The hot spot position data refers to the local region coordinates in the three-dimensional thermal distribution whose temperature exceeds the hardware design safety range, which can be generated by real-time comparison of the grid unit temperature and the preset safety threshold, and is used to accurately locate the overheating area corresponding to the actual hardware. The plurality of time steps refers to a predetermined continuous judgment period, for example, it can be 5 sampling intervals, which is used to exclude the interference of instantaneous temperature fluctuation on the judgment of thermal bottleneck. The preset threshold refers to the critical value of heat flux density dynamically adjusted according to the hardware cooling capacity, which can be set by experimental data fitting or historical operation data analysis, and is used to judge the actual load state of the cooling system.

[0079] In this embodiment, the internal space of the server is first discretized into a plurality of interrelated physical units through three-dimensional grid division, and the thermal load data of each unit is independently calculated, thereby accurately reflecting the dynamic balance state of heat generation and dissipation of different hardware components in the spatial dimension. When the temperature of a certain grid unit continuously exceeds the hardware safety threshold, the corresponding physical location is marked as a hot spot area. Further combined with the heat flux density data, if the hot spot area does not show a temperature drop trend in a plurality of continuous time periods, and the heat flux density of the corresponding position is continuously higher than the system carrying capacity, it is determined as a persistent thermal bottleneck. This spatio-temporal joint criterion mechanism effectively distinguishes between accidental temperature rise and substantial heat dissipation abnormalities through double verification conditions, avoiding fan misoperation caused by transient load fluctuations.

[0080] In this embodiment, by dynamically correlating the three-dimensional heat distribution and the heat flux density, a heat dissipation state evaluation model is established at the hardware level granularity, and a time persistence judgment condition is introduced, which overcomes the response lag problem caused by sensor sampling delay in traditional methods, and through the double filtering of heat flux density and time dimension, the number of invalid speed regulation is significantly reduced, to accurately identify the persistent thermal bottleneck area inside the server, and avoid hardware performance degradation or damage caused by local overheating. At the same time, by excluding the interference of transient temperature fluctuations on heat dissipation control, energy waste caused by frequent start-stop of the fan is reduced, and efficient directional allocation of heat dissipation resources is realized. In addition, the dynamic threshold determination mechanism based on heat flux density data can adapt to the heat dissipation characteristics of different hardware configurations, and improve the robustness of the system under different load scenarios.

[0081] In a feasible implementation manner, the energy consumption prediction model comprises a pre-trained long short-term memory network model; with reference to Figure 4 , step S300 comprises steps S310-S320, wherein:

[0082] Step S310, the real-time load data, historical load data, task queue information and environmental parameters are subjected to time sequence alignment processing and normalization processing to obtain processed time sequence feature data;

[0083] Step S320, the processed time sequence feature data is input into the pre-trained long short-term memory network model for time series analysis processing to output the predicted energy consumption sequence of the future time period.

[0084] In this embodiment, the time sequence alignment processing refers to synchronizing data from different sources according to a unified timestamp, which can be implemented by using a sliding window matching or an interpolation method to eliminate the time offset caused by the difference in sampling frequency of the sensor. The normalization processing refers to converting data of different dimensions to a unified numerical range, which can be implemented by using a minimum-maximum normalization or Z-score standardization method to eliminate the interference of data distribution difference on model training. The pre-trained long short-term memory network model refers to an LSTM neural network pre-trained by historical load and energy consumption data, which can be implemented by using a recurrent neural network structure containing a forget gate, an input gate and an output gate to capture the nonlinear time sequence correlation between load fluctuation and energy consumption.

[0085] In this embodiment, real-time load data, historical load data, task queue information and environmental parameters are first input to the data processing module. Through time sequence alignment processing, the task submission time in the task queue is matched with the load data collected by the sensor through timestamp matching, for example, linear interpolation method is used to fill the data of missing time points. Subsequently, the normalization processing maps the data of different dimensions such as CPU utilization, memory occupancy and environmental temperature to the [0, 1] interval to form standardized time sequence feature data. The processed feature data is input to the pre-trained LSTM model, which stores historical state through memory unit and dynamically adjusts information flow through gating mechanism to learn the complex correlation pattern between server load change and energy consumption. For example, when intensive computing tasks appear in the task queue, the LSTM model can predict the rising trend of server energy consumption in the next 5 minutes based on the periodicity of historical load and the current environmental temperature, and generate the corresponding predicted energy consumption sequence.

[0086] It can be understood that the existing scheme uses linear regression or fixed threshold for energy consumption prediction, which cannot handle the burstiness and periodicity of server load. For example, the traditional method only estimates energy consumption according to the current CPU utilization, ignoring the influence of the tasks to be executed in the task queue, resulting in that the prediction result lags behind the actual load change. While the present scheme jointly analyzes multi-source time sequence data through the LSTM model, which can identify the periodicity in load fluctuation and the influence of burst tasks, for example, predicting the energy consumption peak in advance before the GPU acceleration task is submitted, thereby improving the prediction accuracy, solving the problem of waste of power resources caused by static energy consumption prediction, and realizing the dynamic prediction of future energy consumption trend of the server. Through time sequence alignment and normalization processing, the interference of time offset and dimension difference of multi-source data on the prediction model is eliminated; through the LSTM model, the time sequence correlation between load, task queue and environmental parameters is modeled, which improves the accuracy of the prediction result. The predicted energy consumption sequence generated thereby can provide a basis for dynamic power distribution, for example, reducing power supply redundancy during low load period or adjusting power supply configuration in advance before energy consumption peak, thereby optimizing the utilization rate of power resources.

[0087] In an implementable embodiment, referring to Figure 5 , the method further comprises steps S330-S350, wherein:

[0088] In step S330, the confidence data of the predicted energy consumption sequence is calculated based on the historical prediction error data recorded in the training phase of the pre-trained long short-term memory network model.

[0089] In step S340, if the confidence data is lower than a preset value, a thermal influence quantization value is calculated according to the thermal bottleneck data.

[0090] In step S350, the thermal influence quantization value is input as a dynamic correction factor into the energy consumption prediction model to regenerate a corrected predicted energy consumption sequence.

[0091] In this embodiment, the historical prediction error data refers to the deviation statistical result of the predicted value and the actual energy consumption value of the model in the training phase, which can be specifically realized by using a sliding window method to calculate the standard deviation of the prediction error of the past N time steps, for evaluating the fluctuation range of the current prediction result. The confidence data refers to the reliability index of the prediction sequence, which can be specifically realized by calculating the confidence interval width through the error distribution probability density function, for judging whether to trigger the correction mechanism. The thermal influence quantization value refers to the numerical representation of the influence of the thermal bottleneck on the energy consumption, which can be specifically realized by using the product of the heat flux density and the temperature gradient for integral operation, for quantifying the additional influence of local overheating on the overall energy consumption of the system. The dynamic correction factor refers to the compensation coefficient for adjusting the prediction result, which can be specifically realized by linearly weighting and fusing the thermal influence quantization value and the historical error data, for converting the thermodynamic parameters into a correction signal recognizable by the model.

[0092] In this embodiment, in the energy consumption prediction process, first, the confidence index of the current prediction sequence is calculated through the statistical characteristics of the historical prediction error. When the confidence is lower than a preset threshold, it indicates that the time sequence characteristics of the current load and environmental parameters are insufficient to support accurate prediction, at which time the thermodynamic parameters need to be introduced for compensation. By extracting the heat flux density and temperature gradient data of the thermal bottleneck region, the heat energy accumulation rate of the region is calculated and compared with the historical average thermal influence level to generate a normalized thermal influence quantization value. The quantization value is converted into a dynamic correction factor through a preset mapping relationship, which acts on the output layer neurons of the LSTM model in a superimposed manner, so as to offset and compensate the original prediction sequence. For example, when the heat flux density of the thermal bottleneck region is continuously higher than the threshold value, the correction factor will increase the predicted energy consumption value in proportion, reflecting the additional power consumption caused by the increase of heat dissipation demand.

[0093] In this embodiment, by establishing a confidence trigger mechanism and a thermodynamic parameter compensation channel, the thermal environment parameters are actively introduced for correction when the prediction reliability is insufficient, solving the prediction deviation problem caused by abnormal heat distribution. Therefore, when the server appears local overheating or heat dissipation efficiency decreases, the confidence decrease state of the prediction model can be effectively identified, and the prediction result is corrected in real time through the thermodynamic parameters, avoiding the energy consumption estimation distortion caused by thermal environment mutation. For example, in the scene where the GPU cluster produces thermal bottleneck due to poor heat dissipation, the corrected prediction sequence can accurately reflect the influence of fan power consumption increase on total energy consumption, thereby providing reliable input for power distribution strategy.

[0094] In a feasible implementation manner, the reference Figure 6 , step S400 includes steps S410-S430, wherein:

[0095] Step S410, generating fan speed adjustment suggestion data based on the thermal bottleneck data and temperature prediction data;

[0096] Step S420, generating power distribution suggestion data based on the predicted energy consumption data;

[0097] Step S430, performing parameter optimization processing on the fan speed adjustment suggestion data and the power distribution suggestion data through a multi-objective optimization algorithm to obtain optimized fan speed parameters and optimized power distribution parameters.

[0098] In this embodiment, the thermal bottleneck data refers to the heat dissipation bottleneck region information identified through three-dimensional heat distribution data and heat flux density data analysis, which can be specifically realized by temperature change trend and heat flux density threshold judgment of grid elements, and is used to locate the hardware region that needs to be adjusted preferentially. The temperature prediction data refers to the future temperature change trend calculated based on the heat conduction model, which can be specifically realized by using a time series prediction algorithm, and is used to predict the heat dissipation demand in advance. The predicted energy consumption data refers to the future energy consumption sequence predicted by the load and environmental parameters, which can be specifically generated by using a long short-term memory network model, and is used to dynamically match the power distribution strategy. The multi-objective optimization algorithm refers to an optimization model considering the heat dissipation efficiency and energy consumption index at the same time, which can be specifically realized by using a genetic algorithm or a particle swarm algorithm, and is used to balance the conflict relationship between fan speed and power distribution.

[0099] In this embodiment, first, the hardware region that needs to be adjusted in priority is determined according to the thermal bottleneck data, and the fan speed adjustment suggestion is generated in combination with the temperature prediction data, for example, the speed of the corresponding fan is increased in advance when it is predicted that the temperature of a certain region will exceed the safety threshold. At the same time, based on the predicted energy consumption data, the future load trend of each hardware component is analyzed, and the suggestion of dynamically adjusting the power supply voltage and current is generated, for example, the power supply power of a certain component is reduced when it is predicted that the load of the component will decrease. Subsequently, the fan speed adjustment suggestion data and the power supply distribution suggestion data are input into a multi-objective optimization algorithm, and the optimal parameter combination that can both inhibit the temperature rise of the thermal bottleneck region and reduce the overall energy consumption is obtained through iterative calculation. Thus, the dynamic coordination of heat dissipation control and energy consumption management is realized, and the waste of resources or the risk of hardware overheating caused by a single adjustment strategy is avoided.

[0100] In this embodiment, by fusing the thermal prediction and energy consumption prediction data and using a multi-objective optimization algorithm for joint parameter optimization, the temperature change trend can be responded in advance, the load demand can be dynamically matched, and the mutual restriction problem between heat dissipation and energy consumption can be solved, so that the heat dissipation and energy consumption control can be coordinated in real time during the operation of the server, the response speed of the fan speed regulation can be improved, and the local overheating or energy waste caused by lag regulation can be avoided. At the same time, the power supply distribution parameters are dynamically adjusted to match the future load trend, and the invalid power consumption is reduced. Finally, the balance optimization of hardware stability and overall energy efficiency is realized.

[0101] In a feasible implementation manner, the step S410 comprises determining a hardware region to be adjusted in priority based on the thermal bottleneck data; and generating a fan speed adjustment suggestion in advance in combination with the temperature prediction data to obtain fan speed adjustment suggestion data.

[0102] In this embodiment, the thermal bottleneck data refers to the continuously high heat load region identified through three-dimensional heat distribution and heat flux density analysis, which can be specifically realized by using grid heat conduction calculation combined with time series monitoring, and is used to locate the position of the hardware component that needs to be cooled in priority. The temperature prediction data refers to the future temperature change trend derived based on a physical heat conduction model, which can be specifically realized by using a time series prediction algorithm combined with a thermodynamic equation modeling, and is used to predict the heat load evolution direction of a specific region.

[0103] In this embodiment, when it is detected that a certain processor region presents heat flux density exceeding the standard and temperature gradient anomaly for three consecutive monitoring periods, the region is marked as thermal bottleneck data. In combination with the future five-minute temperature rise curve of the region predicted based on the LSTM model, a scheme of increasing the speed of the corresponding cooling fan in stages is generated in advance. For example, when it is predicted that the temperature will break through the safety threshold after two minutes, the fan speed increasing program is started immediately, instead of waiting for the actual temperature to exceed the standard before responding.

[0104] In this embodiment, after identifying the graphics card region as the current thermal bottleneck, the corresponding cooling fan speed is individually increased to a preset safety value, solving the problems of resource allocation homogenization and response lag in traditional cooling control. The targeted intensive cooling of high heat flux density areas in the server avoids over-driving the fans in non-critical areas, while the predictive speed regulation eliminates the hardware protective frequency reduction phenomenon caused by sudden temperature increases, allowing the server to reduce overall cooling energy consumption while maintaining safe operating temperatures.

[0105] In one possible implementation, step S420 includes analyzing the future load trend of each hardware component based on the predicted energy consumption data; generating a suggestion for dynamically adjusting the supply voltage and current to obtain power distribution suggestion data.

[0106] In this embodiment, the predicted energy consumption data refers to the power consumption sequence data in the future time period output by the energy consumption prediction model, which can be generated by using a long short-term memory network to perform time series analysis on real-time load, historical load, and environmental parameters, and is used to reflect the power demand changes of hardware components in future operation. The future load trend refers to the calculation task intensity and power consumption fluctuation law of hardware components in the subsequent operation period analyzed from the predicted energy consumption data, which can be obtained by using a sliding window algorithm to fit the trend of the predicted energy consumption sequence, and is used to identify the power demand differences of different hardware components. Dynamic adjustment of the supply voltage and current refers to real-time adjustment of the supply parameters of hardware components according to the load trend, which can be realized by a voltage adjustment circuit and a current control chip of a programmable power supply module, so that the supply parameters can match the actual operating state of the hardware components.

[0107] In this embodiment, the scheme first analyzes the load intensity change curve of each hardware component in the future time window through the predicted energy consumption data output by the energy consumption prediction model. For example, when it is predicted that the GPU will enter a high-load rendering task in the next 5 minutes, the corresponding supply demand curve of the GPU will show a steep upward trend. Based on this trend analysis, the system generates a supply voltage increase suggestion for the GPU, and generates a supply current reduction suggestion for the storage device in a low-load state. Further, by comparing the predicted load trend with the current power distribution strategy, hardware components with power redundancy or insufficient supply are identified, and a set of differentiated voltage and current adjustment parameters are generated. These adjustment parameters are transmitted to the programmable power supply module through the power management interface of the baseboard management controller, realizing dynamic allocation of power resources according to the actual needs of hardware components.

[0108] In this embodiment, by predicting future load trends, the power supply parameters are adjusted in advance, for example, the supply voltage of the CPU is raised before it enters high-load calculation, to avoid performance degradation caused by power supply delay. At the same time, for hardware components with periodic load fluctuations, intermittent voltage reduction can be implemented according to the predicted curve to reduce power consumption during inefficient operation. In this way, the application can accurately adjust the power supply parameters according to the future operating state of the hardware components, eliminating the power redundancy caused by static allocation. When the server faces a sudden load, the power supply capacity of the key hardware can be raised in advance to avoid system instability caused by insufficient voltage. For hardware components with periodic load changes, the power supply current during non-peak periods can be dynamically reduced to achieve the dual goals of energy consumption optimization and hardware protection.

[0109] In a possible implementation, the method further comprises the following steps after step S500: Figure 7 In step S500, the method further comprises steps S610-S620, wherein:

[0110] In step S610, the temperature data and the energy consumption data after execution are monitored.

[0111] In step S620, the temperature data and the energy consumption data are fed back to the thermal perception model and the energy consumption prediction model to update the model parameters of the thermal perception model and the energy consumption prediction model.

[0112] In this embodiment, the temperature data and the energy consumption data after execution are monitored by continuously collecting thermodynamic parameters and power consumption parameters in the actual operating state through sensors deployed on the server hardware. Specifically, a combination of a distributed temperature sensor array and a smart meter can be used to obtain real working condition data after the optimization strategy is executed. The model parameter update refers to dynamically adjusting the weight matrix of the machine learning model according to the difference between the actual running data and the predicted data. Specifically, the gradient descent algorithm can be used to correct the internal parameter mapping relationship of the model through the back propagation mechanism.

[0113] In this embodiment, after the fan speed and power supply distribution adjustment is completed, the temperature sensor and the energy consumption metering device continuously collect physical layer feedback signals to form a monitoring data set containing time stamps. The data set is input to the loss function calculation module of the thermal perception model to update the material thermal resistance coefficient in the heat conduction calculation by comparing the difference between the predicted temperature distribution and the actual temperature distribution. At the same time, the energy consumption data is input to the error correction module of the energy consumption prediction model to adjust the forgetting gate weight parameter of the long short-term memory network by calculating the mean square error between the predicted energy consumption sequence and the actual energy consumption curve. This double feedback mechanism enables the model to adapt to the heat dissipation efficiency decay caused by server hardware aging and the heat exchange coefficient fluctuation caused by the change of the computer room environment temperature and humidity.

[0114] In some embodiments, the temperature data feedback channel can configure a sliding time window mechanism, for example, using actual temperature data in the past 24 hours as a training sample set, and updating the model parameters through a batch gradient descent algorithm. Energy consumption data feedback can be combined with a real-time load feature extraction module, for example, when a sudden computing task is detected, an online learning mechanism of the prediction model is automatically triggered.

[0115] In this embodiment, by establishing a linkage mechanism of real-time data acquisition and model parameter updating, the thermal perception model can automatically correct the change of thermal resistance caused by dust accumulation on the heat sink, and the energy consumption prediction model can adapt to the change of power consumption characteristics of new computing chips, overcoming the problem of prediction deviation accumulation of static models in long-term operation, and realizing the continuous self-optimization ability of the server cooling and energy consumption control system. The thermal perception model corrects the heat conduction calculation parameters through actual temperature feedback, improving the accuracy of heat distribution prediction; the energy consumption prediction model calibrates the load feature recognition ability through actual energy consumption data, enhancing the prediction reliability under different working conditions. The collaborative updating mechanism of the two effectively suppresses the exponential growth trend of model prediction error, ensuring the applicability of the optimization strategy in the whole life cycle of the server.

[0116] In the embodiments of the present application, the server BMC collaborative optimization method based on thermal perception and energy consumption prediction establishes a thermal-energy coupling model through multi-dimensional data acquisition and fusion analysis, realizes accurate perception through three-dimensional heat distribution calculation and LSTM prediction, dynamically adjusts the cooling and power supply parameters through a multi-objective optimization algorithm, forms a closed-loop feedback mechanism, and can improve the energy efficiency ratio of the server. In addition, it can also realize dynamic collaborative optimization of cooling control and energy consumption management, improve the server cooling efficiency, and prolong the service life of the hardware.

[0117] It should be noted that the above examples are only for understanding the present application and do not constitute a limitation on the server BMC collaborative optimization method based on thermal perception and energy consumption prediction of the present application. Further simple transformations based on this technical concept are within the scope of protection of the present application.

[0118] The present application also provides a server BMC collaborative optimization system based on thermal perception and energy consumption prediction, which is described with reference to Figure 8 The server BMC collaborative optimization system based on thermal perception and energy consumption prediction comprises a memory 10, a processor 20, and a server BMC collaborative optimization program based on thermal perception and energy consumption prediction stored on the memory 10 and executable on the processor 20. The server BMC collaborative optimization program based on thermal perception and energy consumption prediction is configured to implement the steps of the server BMC collaborative optimization method based on thermal perception and energy consumption prediction.

[0119] The server BMC cooperative optimization system based on heat perception and energy consumption prediction provided in the application adopts the server BMC cooperative optimization method based on heat perception and energy consumption prediction in the above embodiment, and can improve the energy efficiency ratio of the server. Compared with the prior art, the beneficial effects of the server BMC cooperative optimization system based on heat perception and energy consumption prediction provided in the application are the same as those of the server BMC cooperative optimization method based on heat perception and energy consumption prediction provided in the above embodiment, and other technical features in the server BMC cooperative optimization system based on heat perception and energy consumption prediction are the same as those disclosed in the above embodiment method, and will not be repeated here.

[0120] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0121] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any equivalent structural transformation made by using the content of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A server BMC collaborative optimization method based on thermal sensing and energy consumption prediction, characterized in that, The method includes: Raw temperature data is obtained by temperature sensors deployed on key server hardware, raw heat flux density data is obtained by heat flux sensors, real-time load data, historical load data and task queue information are obtained by the Baseboard Management Controller (BMC), and environmental parameters are obtained by the data center environment monitoring system. The raw temperature data and raw heat flux density data are preprocessed to obtain processed temperature data and processed heat flux density data. Based on the processed temperature data, processed heat flux density data, and real-time load data, the thermal sensing model is used for analysis and processing to obtain thermal bottleneck data and temperature prediction data. Based on the real-time load data, historical load data, task queue information, and environmental parameters, an energy consumption prediction model is used to perform prediction processing to obtain predicted energy consumption data. Based on the thermal bottleneck data, temperature prediction data, and predicted energy consumption data, a collaborative optimization control strategy is executed to obtain optimized fan speed parameters and optimized power distribution parameters. The BMC executes the optimized fan speed parameters to control the server fan speed, and executes the optimized power distribution parameters to adjust the power supply to each hardware component; The energy consumption prediction model includes a pre-trained long short-term memory network model; the step of using the energy consumption prediction model to perform prediction processing based on the real-time load data, historical load data, task queue information, and environmental parameters to obtain predicted energy consumption data includes: The real-time load data, historical load data, task queue information, and environmental parameters are subjected to time-series alignment and normalization processing to obtain processed time-series feature data. The processed time-series feature data is input into a pre-trained long short-term memory network model for time series analysis to output a predicted energy consumption sequence for future time periods. The method further includes: Based on the historical prediction error data recorded during the training phase of the pre-trained long short-term memory network model, the confidence data of the predicted energy consumption sequence is calculated. If the confidence level data is lower than the preset value, then the thermal impact quantification value is calculated based on the thermal bottleneck data; The quantified value of thermal impact is used as a dynamic correction factor and input into the energy consumption prediction model to regenerate the corrected predicted energy consumption sequence. The step of executing a collaborative optimization control strategy based on the thermal bottleneck data, temperature prediction data, and predicted energy consumption data to obtain optimized fan speed parameters and optimized power distribution parameters includes: Based on the aforementioned thermal bottleneck data and temperature prediction data, fan speed adjustment suggestion data is generated; Based on the predicted energy consumption data, power allocation recommendation data is generated; The fan speed regulation suggestion data and power allocation suggestion data are optimized using a multi-objective optimization algorithm to obtain optimized fan speed parameters and optimized power allocation parameters.

2. The server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in claim 1, characterized in that, The step of analyzing and processing the processed temperature data, processed heat flux density data, and real-time load data using a thermal sensing model to obtain thermal bottleneck data and temperature prediction data includes: The processed temperature data, processed heat flux density data, and real-time load data are subjected to three-dimensional mesh generation, dividing the internal space of the server into multiple mesh units and mapping them to the corresponding hardware components. Based on the physical principle of heat conduction, heat conduction calculations are performed on each grid cell to generate three-dimensional heat distribution data; Based on the three-dimensional thermal distribution data, temperature change trend prediction processing is performed to obtain temperature prediction data; based on the three-dimensional thermal distribution data and the processed heat flux density data, thermal bottleneck identification processing is performed to obtain thermal bottleneck data.

3. The server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in claim 2, characterized in that, The step of performing thermal bottleneck identification processing based on the three-dimensional heat distribution data and the processed heat flux density data to obtain thermal bottleneck data includes: Based on the three-dimensional thermal distribution data, calculate the thermal load data of each hardware component; Based on the heat load data, grid cells whose temperature exceeds the hardware safety threshold are identified, and hotspot location data is obtained. Based on the heat flux density data processed in step [1], if the hot spot location data does not improve for multiple consecutive time steps and the heat flux density exceeds a preset threshold, it is determined to be heat bottleneck data.

4. The server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in claim 1, characterized in that, The step of generating fan speed control suggestion data based on the thermal bottleneck data and temperature prediction data includes: Based on the thermal bottleneck data, determine the hardware area to be adjusted first; Based on the temperature prediction data, fan speed adjustment suggestions are generated in advance to obtain fan speed adjustment suggestion data.

5. The server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in claim 1, characterized in that, The step of generating power allocation recommendation data based on the predicted energy consumption data includes: Based on the predicted energy consumption data, analyze the future load trends of each hardware component; Generate suggestions for dynamically adjusting supply voltage and current to obtain power distribution recommendation data.

6. The server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in claim 1, characterized in that, After the steps of executing the optimized fan speed parameters through the BMC to control the server fan speed and executing the optimized power distribution parameters to adjust the power supply to each hardware component, the method further includes: Monitor temperature and energy consumption data after execution; The temperature data and energy consumption data are fed back to the thermal sensing model and the energy consumption prediction model to update the model parameters of the thermal sensing model and the energy consumption prediction model.

7. A server BMC collaborative optimization system based on thermal sensing and energy consumption prediction, characterized in that, The server BMC collaborative optimization system based on thermal sensing and energy consumption prediction includes: a memory, a processor, and a server BMC collaborative optimization program based on thermal sensing and energy consumption prediction stored in the memory and executable on the processor. The server BMC collaborative optimization program based on thermal sensing and energy consumption prediction is configured to implement the steps of the server BMC collaborative optimization method based on thermal sensing and energy consumption prediction as described in any one of claims 1 to 6.

Images