Cloud data center energy consumption optimization method based on model health

Abstract

The application discloses a cloud data center energy consumption optimization method based on model health degree, relates to the technical field of cloud computing and data center energy-saving control, and discloses the following technical scheme. A time slice division and operation data collection mechanism is set to collect operation data. A power consumption and temperature prediction model online update mechanism is set to perform online parameter update on a previously established server power consumption prediction model and server temperature prediction model. A unified model health degree used for representing the health state of the power consumption prediction model is constructed. A multi-dimensional linkage control parameter determination mechanism is set to determine control parameters. Joint energy consumption optimization solving is performed to obtain an operation scheme. A health degree low prediction and processing mechanism is set to predict the operation state of a physical server. When the model health degree is high, an aggressive energy-saving strategy is adopted, and joint energy consumption optimization solving and a long-term closed-loop operation mechanism are combined, so that a comprehensive balance between energy-saving benefits, model mismatch risks, virtual machine migration disturbance and optimization calculation overheads is achieved in the long-term operation process of the data center.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and data center energy-saving control technology, specifically a cloud data center energy consumption optimization method based on model health. Background Technology

[0002] With the rapid development of cloud computing, big data, and artificial intelligence, data center scale continues to expand, and the number and power density of servers are constantly increasing, leading to a rapid increase in overall data center energy consumption and a significant increase in operating costs and carbon emission pressure. To reduce energy consumption, the industry has proposed various energy-saving control technologies, mainly including virtual machine placement and migration optimization, server frequency and power management, cooling system optimization and supply air temperature control, and joint optimization frameworks. On the other hand, in data center operation and maintenance, building energy conservation, and machine learning operation and maintenance, existing work has proposed using predictive models or health indicators to evaluate the status of systems and models. For example, cooling effect indicators are calculated based on temperature and airflow distribution to assess the health of the cooling system and whether there are localized overcooling or hot spots; model health is constructed based on indicators such as prediction error and data distribution drift to trigger model retraining or anomaly alarms; and in some control schemes, model uncertainty or confidence is introduced into the control algorithm to weigh energy-saving benefits against the risk of exceeding constraints. However, existing technologies still have certain shortcomings. Most joint energy optimization methods assume that the power consumption prediction model and temperature prediction model are reliable throughout their entire life cycle, but lack continuous evaluation of the model's health status. When the model is mismatched or the parameters are aging, aggressive energy-saving strategies are still implemented, which can easily lead to energy consumption prediction errors, rack temperature exceeding limits, or service quality degradation. Existing studies on health or confidence levels often use health for a single purpose, such as for alarms, simple correction of a certain control parameter, or triggering model retraining. They do not coordinate and control a unified model health index with virtual machine placement strategies, temperature safety margins, virtual machine migration scale, and energy consumption optimization trigger cycles. In the long-term operation scenario of cloud data centers, a balance needs to be struck between energy-saving benefits, model mismatch risks, virtual machine migration disturbances, and optimization computational overhead. Existing solutions do not adequately consider these factors and lack a closed-loop mechanism for long-term operation, making it difficult to maintain stable and reliable energy-saving effects under conditions of continuous changes in business and environment. Summary of the Invention

[0003] The purpose of this invention is to provide a cloud data center energy consumption optimization method based on model health, so as to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a cloud data center energy consumption optimization method based on model health, the method comprising the following steps: S1. The time slice division and runtime data collection mechanism is set up to divide the runtime of the cloud data center into continuous optimized time slices and collect runtime data of physical servers and virtual machines running on physical servers. S2. Set up an online update mechanism for power consumption and temperature prediction models to update the parameters of the pre-established server power consumption prediction model and server temperature prediction model online based on the resource utilization rate, historical power measurement value and historical temperature measurement value in the historical data buffer. S3. Construct a first health index to characterize the health status of the power consumption prediction model and a second health index to characterize the health status of the temperature prediction model, and then weight and fuse the first health index and the second health index to obtain a unified model health. S4. A multi-dimensional linkage control parameter determination mechanism is set up to compare the unified model health with the first threshold and the second threshold, and to determine the control parameters through a pre-set piecewise continuous mapping function about the unified model health. S5. The joint energy consumption optimization solution obtains the placement scheme of virtual machines to target physical servers, the processor operating frequency of each physical server and the air supply temperature setpoint of the cooling system, and optimizes the scheme during long-term operation. A low health prediction and handling mechanism is set up to predict the operating status of physical servers during energy consumption optimization. If it is predicted that the temperature of a physical server exceeds the upper limit of the temperature safety margin and the health of the unified model is continuously lower than the first threshold, a conservative control strategy is triggered.

[0005] Furthermore, in step S1: the time slice division and runtime data acquisition mechanism is used to divide the runtime of the cloud data center into continuous optimized time slices. Within each optimized time slice, resource utilization information of each physical server and the virtual machines running on it is collected, such as CPU utilization, memory utilization, power measurement values, and processor temperature or memory temperature measurement values. The collected data is then stored in the historical data buffer.

[0006] Furthermore, in step S2: the online update mechanism for the power consumption and temperature prediction model updates the pre-established server power consumption prediction model and server temperature prediction model online based on the resource utilization rate, historical power measurement value, and historical temperature measurement value in the historical data buffer; the online update uses a recursive least squares algorithm with a forgetting factor or a Kalman filter algorithm to incrementally correct the model parameters, enabling the model to track changes in the data center environment, business load structure, and equipment performance over time; using the updated power consumption prediction model and temperature prediction model, the power consumption prediction value and temperature prediction value of each physical server in the current optimization time slice are calculated respectively, in order to track the slow changes in the data center environment, load structure, and equipment performance over time, and improve the adaptability of the prediction model in the long-term operation process.

[0007] Furthermore, in step S3: within a preset sliding time window ending with the current optimization time slice, power prediction error sequences and temperature prediction error sequences are constructed respectively. Power prediction error statistics and temperature error statistics are calculated based on the absolute or squared values ​​of the errors. Using a monotonically decreasing normalization function, the power prediction error statistics are mapped to a first health index, and the temperature prediction error statistics are mapped to a second health index. Weighting coefficients are set to weight and fuse the first and second health indices to obtain a unified model health index H, where the value of H ranges from 0 to 1.

[0008] Furthermore, the length of the sliding time window is adaptively adjusted according to the fluctuation cycle of the cloud data center's business load to achieve a balance between the drastic fluctuations in health caused by an excessively short window length and the lag in health caused by an excessively long window length. The first health index is calculated based on the absolute or squared value of the power prediction error within the sliding time window to obtain an error statistic, which is mapped to the 0-1 interval through a monotonically decreasing normalization function. This index is used to characterize the health status of the power prediction model within the sliding time window. The second health index is calculated based on the absolute or squared value of the temperature prediction error within the sliding time window to obtain an error statistic, which is also mapped to the 0-1 interval through a monotonically decreasing normalization function. This index is used to characterize the health status of the temperature prediction model within the sliding time window. When the error is small, the health status corresponding to the error tends to 1; when the error exceeds a preset upper limit, the health status corresponding to the error tends to 0. The unified model health status is used to jointly characterize the statistical reliability of the power prediction model and the temperature prediction model, as well as the risk level of temperature constraints, power constraints, or service level agreement constraints being violated due to model mismatch.

[0009] Furthermore, in step S4: the multi-dimensional linkage control parameter determination mechanism is used to set the first threshold H. low Second threshold H high, where 0 < H low <H high <1, and compare the unified model health H with the first threshold and the second threshold, and determine the control parameters through a pre-set piecewise continuous mapping function about the unified model health. The control parameters include the type of virtual machine placement optimization algorithm, the temperature safety margin of the server operation, the upper limit of the number of virtual machine migrations allowed to be performed in each optimization time slice, and the trigger cycle length between two adjacent energy consumption optimizations.

[0010] Furthermore, the piecewise continuous mapping function is configured as follows: If H≥H high If the unified health of the prediction model is determined to be in the high health range, indicating high reliability, then an aggressive energy-saving strategy is adopted. Heuristic virtual machine placement algorithms, such as first-fit and best-fit algorithms, are used to reduce algorithm complexity and shorten solution time. The temperature safety margin is set to a relatively small value, allowing server or rack operating temperatures close to the safety limit, in order to increase the cooling system's air supply temperature and reduce cooling energy consumption. The upper limit for the number of virtual machines to be migrated is set to a relatively large value to facilitate more thorough virtual machine consolidation. Finally, the energy consumption optimization trigger cycle is set to a long cycle, such as triggering once every multiple time slices, to reduce optimization computational overhead. If H≤H low If the unified health of the prediction model is determined to be in a low health range, indicating low reliability, a conservative control strategy is adopted. A metaheuristic virtual machine placement algorithm, such as genetic algorithm, particle swarm optimization, or simulated annealing, is used to find conservative solutions that meet safety constraints in a larger solution space. The temperature safety margin is set to a relatively large value, and the upper limit of the allowable target temperature is lowered to increase temperature redundancy. The upper limit of the number of virtual machines to be migrated is set to a relatively small value to reduce business disruption. The energy consumption optimization trigger cycle is set to a short cycle to adjust the control strategy more frequently based on changes in model error. If H low <H<H high The unified health of the prediction model is determined to be in the intermediate transition range. A hybrid strategy of heuristic and metaheuristic algorithms is adopted. The iteration number or weight ratio of the two algorithms is adjusted according to the value of H, so that the heuristic algorithm accounts for a larger proportion when the model health is higher, and the metaheuristic algorithm accounts for a larger proportion when the health is lower. Furthermore, by using linear interpolation, the temperature safety margin, the upper limit of the number of virtual machines to migrate, and the trigger cycle length are set to values ​​that change continuously and monotonically with H, so as to avoid drastic changes in control parameters when the health fluctuates slightly.

[0011] Furthermore, the virtual machine placement optimization algorithm includes heuristic algorithms, metaheuristic algorithms, and hybrid algorithms combining heuristic and metaheuristic algorithms; the temperature safety margin for server operation is the maximum allowable control temperature for the server or rack; the upper limit of the number of virtual machine migrations allowed to be performed in each optimization time slice is used to limit the business disturbances caused by virtual machine migrations; the trigger cycle length between two adjacent energy consumption optimizations is used to control the call frequency and optimization computational overhead of energy consumption optimization solutions.

[0012] Furthermore, in step S5: Under the constraints of the virtual machine placement optimization algorithm type, temperature safety margin, upper limit of virtual machine migration quantity, and trigger cycle determined in step S4, the total energy consumption of the cloud data center in the current optimization time slice is used as the objective function, and the resource capacity of each physical server, the temperature safety margin, and the service level agreement constraints are used as constraints. The placement scheme of each virtual machine to the target physical server, the processor operating frequency of each physical server, and the air supply temperature setpoint of the cooling system are solved. The operation of the cloud data center in the current optimization time slice is controlled according to the solved placement scheme, processor operating frequency, and air supply temperature setpoint. After the current optimization time slice ends, the next optimization time slice is entered and the process returns to step S1. Thus, a comprehensive balance is achieved between energy-saving benefits, model mismatch risk, virtual machine migration disturbance, and optimization computational overhead during long-term operation.

[0013] Furthermore, a low health prediction and handling mechanism is set up to predict the operating status of a physical server in the next one or M optimization time slices. If the physical server's temperature exceeds the upper limit of the temperature safety margin and the unified model's health H remains below the first threshold H, the mechanism will be implemented. low In such cases, a conservative control strategy is triggered, which includes: reducing the processor operating frequency of the physical server; migrating some virtual machines running on the physical server to other physical servers; increasing the global temperature safety margin of the data center; thereby reducing the risk of temperature exceeding limits in the event of model mismatch, and maintaining the operational safety and business stability of the cloud data center before model recovery or retraining.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a unified model health metric by performing sliding statistics and normalizing the power consumption prediction error and temperature prediction error. This unifies the statistical reliability of the power consumption prediction model and the temperature prediction model within a certain time window into a single index, allowing the model health status to be quantified. Model health not only reflects the magnitude of the error but also characterizes the risk level of temperature exceedances, energy consumption deviations, and service quality breaches, providing a unified risk measurement basis for subsequent strategy adjustments. By introducing a unified model health metric, the reliability and risk level of the prediction model are explicitly characterized. Furthermore, by linking model health with key control variables in multiple dimensions, the robustness of joint optimization is improved. This invention does not merely use model health for alarms or to trigger model retraining. Instead, it utilizes a unified model health score to simultaneously drive four key control variables: virtual machine placement algorithm type, temperature safety margin, virtual machine migration limit, and energy consumption optimization trigger cycle. Multi-dimensional linkage control is achieved through segmented continuous mapping. When model health is high, the system adopts a more aggressive energy-saving strategy while ensuring safety constraints. When model health decreases, the system automatically switches to a more conservative control strategy, effectively avoiding excessively aggressive energy consumption optimization during model mismatch stages. This significantly improves the robustness of joint optimization in long-term operating scenarios. Furthermore, segmented continuous mapping and smooth interpolation are used to avoid drastic strategy switching. This invention divides the model health interval into a high health interval, a low health interval, and an intermediate transition interval by setting a first threshold and a second threshold. In the intermediate transition interval, a linear interpolation method is used to make the temperature safety margin, the upper limit of the number of virtual machines to migrate, and the energy consumption optimization trigger cycle change continuously and monotonically with H. This segmented continuous mapping method avoids the control parameters from jumping between different health intervals, reduces the impact of frequent strategy switching on system stability and business performance, and is conducive to the smooth transition of data centers in actual engineering operations. This invention balances energy-saving effects, business disruption, and computational overhead. In joint energy consumption optimization, it explicitly introduces two control variables: an upper limit on the number of virtual machine migrations and an energy consumption optimization trigger cycle, dynamically adjusted by model health. On one hand, by limiting the scale of virtual machine migrations in each time slice, it avoids significant disruption to business operations caused by frequent, large-scale migrations. On the other hand, it extends the energy consumption optimization trigger cycle when model health is high to reduce the number of optimization calls and computational overhead, and shortens the trigger cycle when model health is low to improve the system's responsiveness to model mismatch and environmental changes, thus achieving a balance between energy-saving benefits, business disruption, and computational overhead. Through an online parameter update mechanism, this invention tracks changes in the data center environment, load structure, and equipment performance over time, maintaining... This invention maintains the effectiveness of power consumption prediction and temperature prediction models. Through model health assessment within a sliding time window, it can promptly identify trends in model performance degradation and adjust joint optimization strategies and key control parameters under the drive of unified model health, enabling the energy consumption optimization process to have adaptive capabilities for long-term operation. Even during periods of persistently low model health, triggering conservative control strategies can still ensure the operational safety and business stability of the cloud data center, providing a buffer time window for subsequent model recovery or retraining. In summary, based on existing virtual machine scheduling, frequency control, and cooling control technologies, this invention introduces a unified model health and multi-dimensional linkage control mechanism, improving the reliability and long-term effectiveness of cloud data center energy consumption optimization from a holistic system perspective, and has promising engineering application prospects. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process flow of the cloud data center energy consumption optimization method based on model health of the present invention; Figure 2 This is a schematic diagram of the model health calculation process for the cloud data center energy consumption optimization method based on model health of the present invention; Figure 3 This is a schematic diagram illustrating the linkage relationship between model health-driven control parameters in the cloud data center energy consumption optimization method based on model health of the present invention. Figure 4 This is a schematic diagram of a typical cloud data center deployment structure for the cloud data center energy consumption optimization method based on model health of the present invention. Figure 5 This is a schematic diagram of the conservative control strategy of the cloud data center energy consumption optimization method based on model health of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figures 1-5 As shown, this invention provides a technical solution: a cloud data center energy consumption optimization method based on model health, such as... Figure 1 As shown, the method includes the following steps: S1. The time slice division and runtime data collection mechanism is set up to divide the runtime of the cloud data center into continuous optimized time slices and collect runtime data of physical servers and virtual machines running on physical servers. S2. Set up an online update mechanism for power consumption and temperature prediction models to update the parameters of the pre-established server power consumption prediction model and server temperature prediction model online based on the resource utilization rate, historical power measurement value and historical temperature measurement value in the historical data buffer. S3. Construct a first health index to characterize the health status of the power consumption prediction model and a second health index to characterize the health status of the temperature prediction model, and then weight and fuse the first health index and the second health index to obtain a unified model health. S4. A multi-dimensional linkage control parameter determination mechanism is set up to compare the unified model health with the first threshold and the second threshold, and to determine the control parameters through a pre-set piecewise continuous mapping function about the unified model health. S5. The joint energy consumption optimization solution obtains the placement scheme of virtual machines to target physical servers, the processor operating frequency of each physical server and the air supply temperature setpoint of the cooling system, and optimizes the scheme during long-term operation. A low health prediction and handling mechanism is set up to predict the operating status of physical servers during energy consumption optimization. If it is predicted that the temperature of a physical server exceeds the upper limit of the temperature safety margin and the health of the unified model is continuously lower than the first threshold, a conservative control strategy is triggered.

[0018] In step S1: The time slice division and runtime data acquisition mechanism is used to divide the runtime of the cloud data center into continuous optimized time slices. Within each optimized time slice, resource utilization information, power measurement values, and processor temperature or memory temperature measurement values ​​of each physical server and the virtual machines running on it are collected, and the collected data is stored in the historical data buffer. For example, the runtime of the cloud data center is divided into continuous optimized time slices, each time slice being 10 minutes long. For each optimized time slice, the monitoring system periodically collects: resource indicators such as CPU utilization and memory utilization of each physical server and the virtual machines on it; server power measurement values, including overall power consumption or power consumption of key components; server processor temperature, memory temperature, and some ambient temperatures.

[0019] In step S2: as Figure 2 As shown, the online update mechanism for the power consumption and temperature prediction models updates the pre-established server power consumption and temperature prediction models online based on resource utilization, historical power measurements, and historical temperature measurements in the historical data buffer. The online update uses a recursive least squares algorithm with a forgetting factor or a Kalman filter algorithm to incrementally correct the model parameters, enabling the model to track changes in the data center environment, business load structure, and equipment performance over time. Using the updated power consumption and temperature prediction models, the predicted power consumption and temperature values ​​for each physical server within the current optimization time slice are calculated to track the slow changes in the data center environment, load structure, and equipment performance over time, improving the adaptability of the prediction models during long-term operation.

[0020] In step S3: as Figure 3 As shown, within a preset sliding time window ending at the current optimization time slice, power prediction error sequences and temperature prediction error sequences are constructed respectively. Power prediction error statistics and temperature error statistics are calculated based on the absolute or squared value of the error. Using a monotonically decreasing normalization function, the power prediction error statistics are mapped to a first health index, and the temperature prediction error statistics are mapped to a second health index. Weighting coefficients are set to weight and fuse the first and second health indices to obtain a unified model health index H, where the value of H ranges from 0 to 1.

[0021] The length of the sliding time window is adaptively adjusted according to the fluctuation cycle of the cloud data center's business load to achieve a balance between the drastic fluctuations in health caused by an excessively short window length and the lag in health caused by an excessively long window length. The first health index is calculated based on the absolute or squared value of the power prediction error within the sliding time window, and mapped to the 0-1 interval through a monotonically decreasing normalization function. This index is used to characterize the health status of the power prediction model within the sliding time window. The second health index is calculated based on the absolute or squared value of the temperature prediction error within the sliding time window, and mapped to the 0-1 interval through a monotonically decreasing normalization function. This index is used to characterize the health status of the temperature prediction model within the sliding time window. Specifically, when the error is small, the corresponding health index tends to be 1, and when the error exceeds a preset upper limit, the corresponding health index tends to be 0. The unified model health index ranges from 0 to 1, and is used to jointly characterize the statistical reliability of the power prediction model and the temperature prediction model, as well as the risk level of temperature constraints, power constraints, or service level agreement constraints being violated due to model mismatch.

[0022] In step S4: as Figure 3 As shown, the multi-dimensional linkage control parameter determination mechanism is used to set the first threshold H. low Second threshold H high The unified model health H is compared with the first threshold and the second threshold, and the control parameters are determined by a pre-set piecewise continuous mapping function for the unified model health. The control parameters include the type of virtual machine placement optimization algorithm, the temperature safety margin of the server operation, the upper limit of the number of virtual machine migrations allowed to be performed in each optimization time slice, and the trigger cycle length between two adjacent energy consumption optimizations.

[0023] The piecewise continuous mapping function is configured as follows: If H≥H high If the unified health of the prediction model is determined to be in the high health range, indicating high reliability, then an aggressive energy-saving strategy is adopted. Heuristic virtual machine placement algorithms, such as first-fit and best-fit algorithms, are used to reduce algorithm complexity and shorten solution time. The temperature safety margin is set to a relatively small value, allowing server or rack operating temperatures close to the safety limit, in order to increase the cooling system's air supply temperature and reduce cooling energy consumption. The upper limit for the number of virtual machines to be migrated is set to a relatively large value to facilitate more thorough virtual machine consolidation. Finally, the energy consumption optimization trigger cycle is set to a long cycle, such as triggering once every multiple time slices, to reduce optimization computational overhead. If H≤H lowIf the unified health of the prediction model is determined to be in a low health range, indicating low reliability, a conservative control strategy is adopted. A metaheuristic virtual machine placement algorithm, such as genetic algorithm, particle swarm optimization, or simulated annealing, is used to find conservative solutions that meet safety constraints in a larger solution space. The temperature safety margin is set to a relatively large value, and the upper limit of the allowable target temperature is lowered to increase temperature redundancy. The upper limit of the number of virtual machines to be migrated is set to a relatively small value to reduce business disruption. The energy consumption optimization trigger cycle is set to a short cycle to adjust the control strategy more frequently based on changes in model error. If H low <H<H high The unified health of the prediction model is determined to be in the intermediate transition range. A hybrid strategy of heuristic and metaheuristic algorithms is adopted. The iteration number or weight ratio of the two algorithms is adjusted according to the value of H, so that the heuristic algorithm accounts for a larger proportion when the model health is higher, and the metaheuristic algorithm accounts for a larger proportion when the health is lower. Furthermore, by using linear interpolation, the temperature safety margin, the upper limit of the number of virtual machines to migrate, and the trigger cycle length are set to values ​​that change continuously and monotonically with H, so as to avoid drastic changes in control parameters when the health fluctuates slightly. Through the above multi-dimensional linkage strategy, the unified model health H not only determines the aggressiveness of the energy-saving strategy, but also coordinates the scale of virtual machine migration and optimizes computing overhead, achieving a risk-reward balance during long-term operation.

[0024] The types of virtual machine placement optimization algorithms include heuristic algorithms, metaheuristic algorithms, and hybrid algorithms combining heuristic and metaheuristic algorithms; the temperature safety margin for server operation is the maximum allowable control temperature of the server or rack; the upper limit of the number of virtual machine migrations allowed per optimization time slice is used to limit the business disturbances caused by virtual machine migrations; the trigger cycle length between two adjacent energy consumption optimizations is used to control the call frequency and optimization computational overhead of energy consumption optimization solutions.

[0025] In step S5: Under the constraints of the virtual machine placement optimization algorithm type, temperature safety margin, upper limit of virtual machine migration number and triggering cycle determined in step S4, the total energy consumption of the cloud data center in the current optimization time slice is used as the objective function, and the resource capacity, temperature safety margin and service level agreement constraints of each physical server are used as the constraints to solve for the placement scheme of each virtual machine to the target physical server, the processor operating frequency of each physical server and the air supply temperature setpoint of the cooling system. The objective function is the overall energy consumption of the data center within the current optimization time slice, including server computing energy consumption and cooling system energy consumption. Constraints include CPU, memory, storage, and network resource capacity constraints for each physical server; server or rack temperature must not exceed the upper limit determined by the temperature safety margin; virtual machine performance must not be lower than that specified in the service level agreement; and the number of virtual machines migrated must not exceed the upper limit for migration per time slice. For virtual machine placement, the control module generates candidate placement schemes based on the current virtual machine load and server resource status, and solves them using selected optimization algorithms: a heuristic algorithm is used to quickly obtain an approximate solution in the high health range, while a meta-heuristic algorithm is used for a deeper search in the low health range. For processor frequency, the control module determines the operating frequency of each server based on the target load and service level agreement constraints, reducing the frequency to lower power consumption while meeting performance requirements. For the cooling system, under temperature safety margin constraints, the control module sets differentiated air supply temperatures based on the predicted rack temperature and airflow distribution, ensuring sufficient cooling in high heat load areas while appropriately increasing the air supply temperature in low heat load areas to reduce cooling energy consumption. The cloud data center is controlled to operate within the current optimization time slice according to the placement scheme, processor operating frequency and air supply temperature setpoint obtained from the solution. After the current optimization time slice ends, the next optimization time slice is entered and the process returns to step S1. This achieves a comprehensive balance between energy-saving benefits, model mismatch risk, virtual machine migration disturbance and optimization computing overhead during long-term operation.

[0026] like Figure 4 As shown, a cloud data center includes multiple racks, each containing several physical servers and network switching equipment; each physical server runs multiple virtual machines through a virtualization platform to support different business applications; the data center uses room air conditioning or row-level air conditioning for cooling, and the cooling system can adjust the supply air temperature and air volume; In terms of software architecture, an energy consumption optimization and control module is deployed on top of the virtualization management platform. This module interacts with the monitoring system, virtual machine scheduling module, and cooling system control module. Obtain resource utilization, power consumption, and temperature measurements for each physical server from the monitoring system; The virtual machine migration operation is performed through the virtual machine scheduling module; Send processor frequency setting instructions to the server management system; Send the air temperature setpoint to the cooling system control module; Based on the aforementioned monitoring data and control interface, the energy consumption optimization control module implements the model health assessment and multi-dimensional linkage control strategy of this invention.

[0027] like Figure 5As shown, a low health prediction and handling mechanism is set up to predict the operating status of a physical server in the next one or M optimization time slices. If the physical server's temperature exceeds the upper limit of the temperature safety margin and the unified model's health H remains below the first threshold H, the mechanism will be implemented. low In such cases, a conservative control strategy is triggered, which includes: reducing the processor operating frequency of the physical server; migrating some virtual machines running on the physical server to other physical servers; increasing the global temperature safety margin of the data center; reducing the risk of temperature exceeding limits in the event of model mismatch; and maintaining the operational safety and business stability of the cloud data center before model recovery or retraining.

[0028] Example 1: The method of the present invention is applied to a cloud data center platform, which includes several racks and a large number of physical servers, and uses a virtualization platform to support a variety of online business applications.

[0029] The optimization time slice length is set to 10 minutes, and the sliding time window length is the most recent 12 optimization time slices, corresponding to approximately 2 hours of historical data, which are used to calculate power consumption prediction error and temperature prediction error.

[0030] The weighting coefficients for the unified model health level H are set as follows: power consumption health level weight w1 = 0.4, temperature health level weight w2 = 0.6, to reflect the priority of temperature safety; the first threshold H low Set to 0.6, the second threshold H high We set H to 0.9. When H is less than 0.6, the prediction model is considered to be in a low health state. When H is greater than 0.9, the prediction model is considered to be in a high health state. The interval between the two is the transition interval.

[0031] The following boundary parameters are set for temperature safety margin, maximum number of virtual machines to be migrated, and energy consumption optimization trigger cycle: If H ≥ 0.9, the prediction model is considered to be in a high health state. Therefore, the temperature safety margin is set to a smaller value, such as 3℃. The maximum number of virtual machines that can be migrated in each optimization time slice is set to a larger value, such as 200. The energy consumption optimization trigger cycle is set to a longer cycle, such as triggering joint optimization once every 6 time slices.

[0032] If H≤0.6, the prediction model is judged to be in a low health state, and the temperature safety margin is taken as a larger value, such as 8℃; the upper limit of the number of virtual machines allowed to be migrated in each optimization time slice is a smaller value, such as 40; the energy consumption optimization trigger cycle is a shorter cycle, such as triggering joint optimization once every 2 time slices; If 0.6 < H < 0.9, the prediction model is judged to be in the intermediate transition range. The temperature safety margin is linearly interpolated by H between 3℃ and 8℃; the upper limit of the number of virtual machines to be migrated is linearly interpolated by H between 40 and 200; the energy consumption optimization triggering cycle is linearly interpolated by H between 2 time slices and 6 time slices; the virtual machine placement algorithm adopts a hybrid strategy of heuristic algorithm and metaheuristic algorithm, in which the iteration ratio of heuristic search and metaheuristic search changes continuously with H.

[0033] When the unified model health score H = 0.75 calculated by the sliding time window within a certain time period, H is between H low With H high Between these values, the temperature safety margin is between 3℃ and 8℃; the maximum number of virtual machines that can be migrated per optimization time slice is between 40 and 200; the energy consumption optimization trigger cycle is between 2 and 6 time slices; the virtual machine placement algorithm adopts a hybrid strategy of heuristic search and metaheuristic search.

[0034] As can be seen, when the model health changes from high to low, the temperature safety margin gradually increases, the virtual machine migration scale gradually decreases, and the energy consumption optimization trigger cycle gradually shortens. The above control variables change continuously with H, avoiding policy abrupt changes caused by slight fluctuations in health.

[0035] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cloud data center energy consumption optimization method based on model health, characterized by: The method includes the following steps: S1. The time slice division and runtime data collection mechanism is set up to divide the runtime of the cloud data center into continuous optimized time slices and collect runtime data of physical servers and virtual machines running on physical servers. S2. Set up an online update mechanism for power consumption and temperature prediction models to update the parameters of the pre-established server power consumption prediction model and server temperature prediction model online based on the resource utilization rate, historical power measurement value and historical temperature measurement value in the historical data buffer. S3. Construct a first health index to characterize the health status of the power consumption prediction model and a second health index to characterize the health status of the temperature prediction model, and then weight and fuse the first health index and the second health index to obtain a unified model health. S4. A multi-dimensional linkage control parameter determination mechanism is set up to compare the unified model health with the first threshold and the second threshold, and to determine the control parameters through a pre-set piecewise continuous mapping function about the unified model health. S5. The joint energy consumption optimization solution obtains the placement scheme of virtual machines to target physical servers, the processor operating frequency of each physical server and the air supply temperature setpoint of the cooling system, and optimizes the scheme during long-term operation. A low health prediction and handling mechanism is set up to predict the operating status of physical servers during energy consumption optimization. If it is predicted that the temperature of a physical server exceeds the upper limit of the temperature safety margin and the health of the unified model is continuously lower than the first threshold, a conservative control strategy is triggered.

2. The cloud data center energy consumption optimization method based on model health as described in claim 1, characterized in that: In step S1: The time slice division and runtime data acquisition mechanism is used to divide the runtime of the cloud data center into continuous optimized time slices. In each optimized time slice, the resource utilization information and power measurement values ​​of each physical server and the virtual machines running on it, as well as the measurement values ​​of processor temperature or memory temperature, are collected, and the collected data is stored in the historical data buffer.

3. The cloud data center energy consumption optimization method based on model health as described in claim 1, characterized in that: In step S2: the online update mechanism for the power consumption and temperature prediction model updates the pre-established server power consumption prediction model and server temperature prediction model online based on the resource utilization rate, historical power measurement value, and historical temperature measurement value in the historical data buffer; the online update uses a recursive least squares algorithm with a forgetting factor or a Kalman filter algorithm to incrementally correct the model parameters; using the updated power consumption prediction model and temperature prediction model, the power consumption prediction value and temperature prediction value of each physical server in the current optimization time slice are calculated respectively.

4. The cloud data center energy consumption optimization method based on model health as described in claim 1, characterized in that: In step S3: Within a preset sliding time window ending with the current optimization time slice, power prediction error sequences and temperature prediction error sequences are constructed respectively. Power prediction error statistics and temperature error statistics are calculated based on the absolute or squared values ​​of the errors. Using a monotonically decreasing normalization function, the power prediction error statistics are mapped to a first health index, and the temperature prediction error statistics are mapped to a second health index. Weighting coefficients are set to weight and fuse the first and second health indices to obtain a unified model health index H. The value range of the unified model health index is 0 to 1.

5. The cloud data center energy consumption optimization method based on model health according to claim 4, characterized in that: The length of the sliding time window is adaptively adjusted according to the fluctuation cycle of the cloud data center's business load, so as to achieve a balance between the sharp fluctuations in health caused by an excessively short window length and the lag in health caused by an excessively long window length; the first health index is calculated based on the absolute value or square value of the power consumption prediction error within the sliding time window to obtain an error statistic, and is mapped to the 0 to 1 interval through a monotonically decreasing normalization function. The second health index is calculated based on the absolute or squared value of the temperature prediction error within the sliding time window to obtain an error statistic, and is mapped to the 0 to 1 interval through a monotonically decreasing normalization function; wherein, when the error is small, the health index corresponding to the error tends to 1, and when the error exceeds a preset upper limit, the health index corresponding to the error tends to 0.

6. The cloud data center energy consumption optimization method based on model health as described in claim 1, characterized in that: In step S4: the multi-dimensional linkage control parameter determination mechanism is used to set the first threshold H. low Second threshold H high The unified model health H is compared with the first threshold and the second threshold, and the control parameters are determined by a pre-set piecewise continuous mapping function for the unified model health. The control parameters include the type of virtual machine placement optimization algorithm, the temperature safety margin of the server operation, the upper limit of the number of virtual machine migrations allowed to be performed in each optimization time slice, and the trigger cycle length between two adjacent energy consumption optimizations.

7. The cloud data center energy consumption optimization method based on model health as described in claim 6, characterized in that: The piecewise continuous mapping function is configured as follows: If H≥H high If the unified health of the prediction model is determined to be in the high health range, and the reliability of the prediction model is determined to be high, then an aggressive energy-saving strategy is adopted. A heuristic virtual machine placement algorithm is adopted, setting the temperature safety margin to a relatively small value, setting the upper limit of the number of virtual machines to be migrated to a relatively large value, and setting the energy consumption optimization trigger cycle to a long cycle. If H≤H low If the unified health of the prediction model is determined to be in the low health range, and the reliability of the prediction model is determined to be low, a conservative control strategy is adopted. A metaheuristic virtual machine placement algorithm is adopted, with the temperature safety margin set to a relatively large value, the upper limit of the number of virtual machines to be migrated set to a relatively small value, and the energy consumption optimization trigger cycle set to a short cycle. If H low <H<H high The unified health of the prediction model is determined to be in the intermediate transition range. A hybrid algorithm of heuristic and metaheuristic algorithms is adopted, and the temperature safety margin, the upper limit of the number of virtual machines to migrate, and the trigger cycle length are set to values ​​that change continuously and monotonically with H through linear interpolation.

8. The cloud data center energy consumption optimization method based on model health as described in claim 6, characterized in that: The virtual machine placement optimization algorithm includes heuristic algorithms, metaheuristic algorithms, and hybrid algorithms of heuristic and metaheuristic algorithms; the temperature safety margin of the server operation is the maximum allowable control temperature of the server or rack; the upper limit of the number of virtual machine migrations allowed to be performed in each optimization time slice is used to limit the business disturbance caused by virtual machine migration; the trigger cycle length between two adjacent energy consumption optimizations is used to control the call frequency and optimization computational overhead of energy consumption optimization solutions.

9. The cloud data center energy consumption optimization method based on model health as described in claim 6, characterized in that: In step S5: Under the constraints of the virtual machine placement optimization algorithm type, temperature safety margin, upper limit of virtual machine migration quantity, and trigger cycle determined in step S4, the total energy consumption of the cloud data center in the current optimization time slice is used as the objective function, and the resource capacity of each physical server, the temperature safety margin, and the service level agreement constraints are used as constraints to solve for the placement scheme of each virtual machine to the target physical server, the processor operating frequency of each physical server, and the air supply temperature setpoint of the cooling system; the operation of the cloud data center in the current optimization time slice is controlled according to the solved placement scheme, processor operating frequency, and air supply temperature setpoint. After the current optimization time slice ends, the next optimization time slice is entered and the process returns to step S1.

10. The cloud data center energy consumption optimization method based on model health as described in claim 6, characterized in that: A low health prediction and handling mechanism is set up to predict the operating status of a physical server in the next one or M optimization time slices. If the physical server's temperature exceeds the upper limit of the temperature safety margin and the unified model's health H remains below the first threshold H, the mechanism will be implemented. low In such cases, a conservative control strategy is triggered; the conservative control strategy is to reduce the processor operating frequency of the physical server, migrate some virtual machines running on the physical server to other physical servers, and improve the global temperature safety margin of the data center.

Images