Method and system for optimizing server power consumption, electronic device, and storage medium

Abstract

The present disclosure provides a method and system, electronic device and storage medium for optimizing server power consumption, wherein a data acquisition module collects operating status data of each server in a server cluster; a power consumption prediction module inputs the operating status data into a power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates power consumption prediction results for each server; a dynamic scheduling engine establishes a cluster energy consumption cost model based on the power consumption prediction results, and performs multi-objective optimization solutions in combination with task priority constraints to generate resource scheduling strategies for each server; a cluster controller performs resource scheduling on each server in the server cluster according to the resource scheduling strategies to adjust and optimize the power consumption of each server. Compared with related technologies, the present disclosure establishes a cluster energy consumption cost model through power consumption prediction results, performs multi-objective optimization solutions in combination with task priority constraints; achieves calculability of global energy consumption; and can achieve optimal global energy efficiency of the cluster.

Description

Technical Field

[0001] The present disclosure relates to the field of server technology, and in particular to a method and system for optimizing server power consumption, an electronic device, and a storage medium. Background Art

[0002] With the continued growth in demand for cloud computing, edge computing, and large-scale data processing, the deployment of multi-node server clusters is becoming increasingly widespread. Server power consumption management directly affects the energy efficiency and operating costs of the cluster. The operating status of a server cluster is highly dynamic and complex. In related technologies, rule-based management and control are widely used for optimizing server power consumption. However, this approach can only use historical data to passively optimize the power consumption of servers; it is unable to predict and make reasonable adjustments to computing scenarios that will occur in the future. In other technologies, time series are used for prediction, but this method has poor prediction accuracy, which can lead to delayed scheduling responses or unbalanced resource allocation, thus affecting the coordinated optimization of overall energy efficiency and service quality. Summary of the Invention

[0003] The present disclosure provides a method and system for optimizing server power consumption, an electronic device, and a storage medium, the main purpose of which is to solve the current problem of high power consumption of server clusters.

[0004] According to a first aspect of the present disclosure, a method for optimizing server power consumption is provided, comprising:

[0005] The data collection module collects the operating status data of each server in the server cluster;

[0006] The power consumption prediction module inputs the operating status data into the power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates the power consumption prediction results of each server;

[0007] The dynamic scheduling engine establishes a cluster energy cost model based on power consumption prediction results, and performs multi-objective optimization based on task priority constraints to generate resource scheduling strategies for each server.

[0008] The cluster controller schedules resources for each server in the server cluster according to the resource scheduling policy to adjust and optimize the power consumption of each server.

[0009] Optionally, the power consumption prediction model includes a load prediction sub-model and a power consumption prediction sub-model;

[0010] The power consumption prediction module inputs the operating status data into the power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates the power consumption prediction results of each server, including:

[0011] Input the operation status data into the load prediction sub-model to generate the load prediction results of each server;

[0012] The load prediction results are input into the power consumption prediction sub-model to generate the power consumption prediction results of each server.

[0013] Optionally, the dynamic scheduling engine builds a cluster energy cost model based on the power consumption prediction results, and performs multi-objective optimization based on task priority constraints to generate resource scheduling strategies for each server, including:

[0014] Get the task priority of the task queue in each server;

[0015] Based on the power consumption prediction results of each server, a cluster energy consumption cost model is constructed to calculate the predicted total power of the server cluster;

[0016] According to the principle of minimizing the predicted total power and combining the task priority constraints, a multi-objective optimization solution is performed on the task resource scheduling of each server. The multi-objective optimization solution is implemented through a genetic algorithm. The resource scheduling strategy includes at least one operation among task migration, node dormancy or frequency adjustment.

[0017] Optionally, the data collection module collects operating status data of each server in the server cluster, including:

[0018] The data collection module obtains the operating status data of each server, and the operating status data includes at least one of the node power consumption value, processor utilization, memory occupancy, network traffic, ambient temperature and task type label collected in real time;

[0019] Records are kept of the running status data of each server at different times.

[0020] Optionally, after the data collection module collects the operating status data of each server in the server cluster, the method further includes:

[0021] The data processing module performs at least one processing method of sliding window normalization processing, graph neural network embedding, and fault mode feature extraction on the operating status data to perform data preprocessing on the operating status data.

[0022] Optionally, the method further includes:

[0023] The training module constructs training data based on the operating status data collected by the data collection module; inputs the training data into the power consumption prediction model to be trained for model training, and obtains a trained power consumption prediction model.

[0024] Optionally, the training data is input into the power consumption prediction model to be trained for model training, including:

[0025] Inputting the training data into the load prediction sub-model of the power consumption prediction model to be trained for training, and generating a training load prediction result;

[0026] Inputting the training load prediction result into the power consumption prediction sub-model of the power consumption prediction model to be trained for training, and generating a training power consumption prediction result;

[0027] Perform dynamic time warping on the training load prediction results and the training power consumption prediction results to generate training prediction results;

[0028] Based on the training prediction results and the training scheduling results, incremental training data is generated, and the incremental training data is used to perform incremental learning training on the power consumption prediction model to be trained.

[0029] According to a second aspect of the present disclosure, a system for optimizing server power consumption is provided, comprising: a data acquisition module, a data processing module, a power consumption prediction module, a dynamic scheduling engine, and a cluster controller;

[0030] The data acquisition module collects the operating status data of each server in the server cluster and sends the operating status data to the data processing module;

[0031] The data processing module performs data preprocessing on the operating status data; the power consumption prediction module performs model inference or model training on the power consumption prediction model based on the preprocessed operating status data sent by the data processing module;

[0032] The dynamic scheduling engine generates resource scheduling strategies or training scheduling results for each server based on the power consumption prediction results or training prediction results generated by the power consumption prediction model;

[0033] The cluster controller schedules resources for each server according to the resource scheduling policy to optimize power consumption.

[0034] According to a third aspect of the present disclosure, there is provided an electronic device, including:

[0035] at least one processor; and

[0036] a memory communicatively connected to the at least one processor; wherein,

[0037] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method for optimizing server power consumption described in the first aspect.

[0038] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method for optimizing server power consumption as described in the first aspect.

[0039] According to a fifth aspect of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the method for optimizing server power consumption as described in the first aspect.

[0040] The present disclosure provides a method and system, electronic device, and storage medium for optimizing server power consumption, relating to the field of server technology. Compared to related technologies, the present disclosure utilizes operating status data input into a power consumption prediction model to predict power consumption changes within a preset timeframe and generate a power consumption prediction result. Compared to traditional threshold-based simple scheduling algorithms, this module achieves "early prediction" of power consumption changes, avoiding response lags and providing a forward-looking basis for dynamic scheduling. A cluster energy cost model is established based on the power consumption prediction results, and a multi-objective optimization solution is performed in conjunction with task priority constraints to generate a resource scheduling strategy. Global energy consumption is calculable by quantifying the relationship between the overall cluster power consumption and the static power consumption and load rate of each node. Combined with multi-objective optimization based on task priorities, a balance is achieved between minimizing power consumption and ensuring task performance, addressing the issues of "global load balancing and energy optimization" and "unable to balance energy efficiency and performance" in traditional methods. Resources are scheduled for each server according to the resource scheduling strategy to adjust and optimize power consumption. Multi-node collaboration is achieved through cluster-level unified scheduling, achieving optimal global cluster energy efficiency while balancing energy conservation and service availability.

[0041] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to better understand the present invention and do not constitute a limitation of the present invention.

[0043] Figure 1 A flow chart of a method for optimizing server power consumption provided by an embodiment of the present disclosure;

[0044] Figure 2 A flowchart of another method for optimizing server power consumption provided by an embodiment of the present disclosure;

[0045] Figure 3 A schematic diagram of the structure of a system for optimizing server power consumption provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0046] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0047] The following describes the server power consumption optimization method and system, electronic device, and storage medium according to embodiments of the present disclosure with reference to the accompanying drawings.

[0048] Figure 1 A flowchart of a method for optimizing server power consumption provided by an embodiment of the present disclosure is provided.

[0049] like Figure 1 As shown, the method comprises the following steps:

[0050] Step 101: The data collection module collects the operating status data of each server in the server cluster.

[0051] In an embodiment of the present disclosure, the data acquisition module continuously and in real time collects and records the operating status data of all servers in the cluster through sensor components and data monitoring tools deployed at each node of the server cluster. The collected operating status data covers key indicators of the server under different operating modes (such as load operation, sleep, low power consumption, etc.), including but not limited to hardware resource utilization data (such as CPU utilization, GPU utilization, memory occupancy, memory bandwidth, network traffic, network packet rate, etc.), energy consumption data (such as real-time power consumption of each node, total power consumption) and environmental parameters (such as the ambient temperature of the node, power supply status, etc.). During the collection process, the data is recorded at a preset time granularity (such as seconds) and forms a structured time series data set, which is stored in the system data unit to provide basic data support for subsequent power consumption prediction and dynamic scheduling.

[0052] Compared to traditional techniques that only collect single-dimensional data (such as monitoring only CPU utilization or power consumption), the data collected in this step covers multiple dimensions, including hardware load, energy consumption, and the environment. It also includes historical and real-time data from servers in different states, providing richer feature inputs for subsequent model learning and resolving the problem of one-sided predictions caused by the limited dimensions of traditional data collection. Through continuous real-time collection and time-series storage, the data is ensured to have time series characteristics, meeting the requirements of historical dependency learning for time-series prediction models (such as LSTM). Compared to traditional static data sampling (such as periodic interval sampling), this avoids the loss of time series features caused by data discontinuities, laying the foundation for accurate predictions. Unified data collection for all servers in the cluster breaks the limitations of traditional single-node independent collection, achieving global aggregation of cluster-level data. This provides data correlation support for subsequent multi-node collaborative scheduling and global energy consumption optimization, resolving the difficulties in collaborative optimization caused by isolated node data in traditional techniques.

[0053] In step 102 , the power consumption prediction module inputs the operation status data into the power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates a power consumption prediction result of each server.

[0054] In an embodiment of the present disclosure, the power consumption prediction module receives the operating status data output by the data acquisition module, pre-processes it, and then inputs it into a preset power consumption prediction model. This model is based on a time series data modeling method. By learning the correlation between historical operating status data and power consumption changes, it predicts the power consumption change trend of each server within a preset time period (such as the next few minutes to tens of minutes), outputs power consumption change data containing time series characteristics, and forms a power consumption prediction result for each server. Among them, the pre-processing of the operating status data includes time series alignment of the data, feature normalization, and noise filtering to ensure that the data input to the model meets the requirements of time series modeling; the power consumption prediction model continuously optimizes its ability to capture long-term dependencies in historical data through continuous iterative learning, thereby improving the accuracy of the prediction results.

[0055] Compared to the prediction method based on the ARIMA model (which relies only on trend and periodic analysis of historical data), the power consumption prediction model adopted by this disclosure can effectively mine the long-term temporal dependencies in the operating status data, solving the problem that traditional models cannot capture sudden changes in dynamic load scenarios, making the prediction results more consistent with the power consumption fluctuation patterns of the actual server operation. By outputting power consumption change data within a preset time period, it provides advance time for subsequent dynamic scheduling. Compared with simple judgments based on thresholds (which can only passively respond to the current load status), it can actively predict future power consumption trends and avoid the problem of untimely energy consumption optimization due to scheduling lags.

[0056] In step 103 , the dynamic scheduling engine establishes a cluster energy consumption cost model based on the power consumption prediction result, and performs multi-objective optimization solution in combination with task priority constraints to generate a resource scheduling strategy for each server.

[0057] In the disclosed embodiment, the dynamic scheduling engine receives the power consumption prediction results for each server from the power consumption prediction module. Combining the basic power consumption parameters and load characteristics of each server in the cluster, it constructs a cluster-level energy cost model to quantify the relationship between the overall cluster energy consumption and the operating status of each node. Simultaneously, the engine introduces task priority parameters as constraints and solves objectives such as "minimizing cluster energy consumption" and "ensuring high-priority task performance" through a multi-objective optimization algorithm. The engine then outputs a resource scheduling policy that includes operations such as task allocation adjustment and node operating status control. The policy must clearly define the specific control methods and execution sequence for each server.

[0058] Compared to static rule-based scheduling methods (which only formulate policies for a single node or local area), the disclosed embodiments establish a cluster energy cost model to quantitatively assess the overall energy consumption of the cluster, resolving the problem of "local optimization while global energy consumption remains high" caused by the lack of multi-node coordination in traditional methods. By combining task priority constraints for multi-objective optimization, compared to strategies that simply pursue power reduction (which can easily lead to delays for high-priority tasks), the disclosed embodiments ensure the execution efficiency of critical tasks during energy consumption optimization, resolving the contradiction in traditional technologies of "difficulty balancing energy efficiency and performance."

[0059] Step 104 : The cluster controller performs resource scheduling on each server in the server cluster according to the resource scheduling policy to adjust and optimize the power consumption of each server.

[0060] In an embodiment of the present disclosure, a cluster controller receives resource scheduling policies output by a dynamic scheduling engine, breaks down the policies into specific execution instructions for each server, and distributes them to the corresponding server nodes via the cluster management interface to implement resource scheduling for the server cluster. Scheduling operations include task migration (reallocating computing tasks between nodes), node operating state adjustments (such as switching to sleep mode or adjusting hardware operating parameters), and resource configuration changes (such as adjusting the operating frequency of computing units) to respond in real time to power consumption prediction results and changes in task requirements. During the scheduling process, the controller continuously collects execution feedback data from each node to verify the scheduling effect and provide a basis for subsequent policy iterations.

[0061] Compared with the control method based on single-node static rules (which can only optimize the power consumption of local nodes and easily lead to an imbalance in the overall load of the cluster), the cluster controller uniformly executes the scheduling strategy to achieve multi-node coordinated adjustment, solving the contradiction between global load balancing and energy consumption optimization in traditional methods and improving the overall energy efficiency of the cluster.

[0062] The present disclosure provides a method for optimizing server power consumption. Compared with related technologies, the present disclosure inputs operating status data into a power consumption prediction model to predict power consumption change data within a preset time period and generate power consumption prediction results. Compared with the traditional "simple scheduling algorithm based on thresholds", this module achieves "early prediction" of power consumption changes, avoids response lag problems, and provides a forward-looking basis for dynamic scheduling. Based on the power consumption prediction results, a cluster energy consumption cost model is established, and a multi-objective optimization solution is performed in combination with task priority constraints to generate a resource scheduling strategy. By quantifying the relationship between the overall power consumption of the cluster and the static power consumption and load rate of each node, the global energy consumption can be calculated. Combined with the multi-objective optimization of task priorities, a balance is found between "minimizing power consumption" and "guaranteeing task performance", solving the problems of "prominent contradictions between global load balancing and energy consumption optimization" and "unable to balance energy efficiency and performance" in traditional methods. According to the resource scheduling strategy, resources are scheduled for each server to adjust and optimize power consumption. Multi-node collaboration is achieved through cluster-level unified scheduling. The optimal global energy efficiency of the cluster can be achieved while taking into account energy saving and service availability.

[0063] In order to clearly illustrate the embodiment of the present disclosure, this embodiment provides a flowchart of another method for optimizing server power consumption.

[0064] like Figure 2 As shown, the method comprises the following steps:

[0065] Step 201: The data collection module collects the operating status data of each server in the server cluster.

[0066] As a specific implementation method, the data acquisition module collects the operating status data of each server in the server cluster, including: the data acquisition module obtains the operating status data of each server, the operating status data includes at least one of the node power consumption value, processor utilization, memory occupancy, network traffic, ambient temperature and task type label collected in real time; records and retains the operating status data of each server at different times.

[0067] Specifically, in step 201, the data collection module collects operational status data from each server at a preset time granularity (e.g., per second) using sensor components (such as power consumption meters and temperature sensors) deployed at each node in the server cluster and out-of-band data collection tools. Node power consumption reflects the server's real-time energy consumption level, while processor utilization (including CPU / GPU), memory usage (including memory bandwidth), and network traffic (including network packet rate and HCA utilization) quantify hardware resource load. Ambient temperature characterizes the physical environment in which the node operates, and task type tags distinguish between different computing tasks requiring different resources. During the collection process, the module transmits data via a dedicated IPMI channel to ensure stable and real-time data acquisition. Furthermore, a structured data format is used to record and retain operational status data for each server at different times (e.g., during sleep, high load, and idle states), forming a time-series historical database.

[0068] By collecting multiple types of data, including node power consumption, hardware load, environmental parameters, and task labels, the core characteristics of server operation are captured. Compared to traditional methods that only collect a single load metric, this provides a multi-feature fusion input foundation for subsequent power consumption prediction models (such as the dual LSTM model), addressing the problem of biased model predictions caused by insufficient data dimensionality. Recording and retaining operational status data at different times creates a complete historical time series dataset, enabling the LSTM model to learn the dynamic changes in server power consumption over time (such as the long-term dependence between load fluctuations and power consumption changes). Compared to static data sampling, this improves the model's adaptability to dynamic load scenarios. Data transmission is achieved through out-of-band collection tools and a dedicated channel (IPMI), reducing the use of internal server computing resources. The structured data format facilitates cleaning and feature extraction in subsequent data processing modules (such as sliding window normalization and graph neural network embedding), providing reliable data support for high-precision training of prediction models.

[0069] Step 202: Input the operation status data into the load prediction sub-model to generate the load prediction result of each server.

[0070] Step 203: Input the load prediction result into the power consumption prediction sub-model to generate the power consumption prediction result of each server.

[0071] Specifically, in steps 202 and 203, the load prediction sub-model (LSTM1) and the power consumption prediction sub-model (LSTM2) adopt a unified network architecture design, both consisting of a 3-layer LSTM network, an attention mechanism (Attention) layer, and a fully connected layer.

[0072] The load prediction sub-model receives preprocessed operating status data (including load-related features such as CPU / GPU utilization, memory usage, network traffic, and task type labels), and extracts short-term fluctuations and long-term dependency features from the time series layer by layer through a three-layer LSTM network. Each LSTM unit controls the flow of information through forget gates, input gates, and output gates to enhance the capture of dynamic load trends. Subsequently, the attention mechanism layer assigns differentiated weights to the features of each time step to highlight the impact of key load mutation moments. Finally, a fully connected layer maps the features into load prediction results within a preset duration (such as 5 minutes) (such as processor utilization timing curves, memory bandwidth change trends, etc.).

[0073] The power consumption prediction sub-model (LSTM2) uses the load prediction results output by the load prediction sub-model as its core input. It also integrates the historical power consumption data of each server (such as the power consumption baseline under different loads). Through the same three-layer LSTM network, it learns the temporal correlation between load and power consumption. The attention mechanism layer focuses on the power consumption-sensitive features during high-load periods. The fully connected layer ultimately outputs the power consumption prediction results for the corresponding time period (such as the sequence of node power consumption values ​​changing over time).

[0074] Compared to shallower structures, a three-layer LSTM network can more deeply exploit the multi-level features in time series data, addressing the problem of traditional single LSTM layers failing to capture long-term dependencies in complex load scenarios. The attention mechanism layer enhances the influence of key time-step features through weight distribution, improving prediction accuracy for critical scenarios such as sudden load changes and periods of high power consumption. The two sub-models utilize a unified architecture to ensure consistency in feature processing logic between load and power prediction. Dynamic Time Warping (DTW) sequence alignment reduces prediction bias caused by time series misalignment. Incremental learning and hierarchical dataset partitioning enable the model to adapt to the hardware characteristics and load patterns of different server clusters, addressing the poor generalization of fixed models in heterogeneous environments. The hierarchical modeling of load and power predictions allows power prediction results to be directly linked to the root causes of load variations, providing more interpretable input for subsequent dynamic scheduling engines (such as multi-objective optimization based on genetic algorithms). Compared to traditional black-box models, this significantly reduces the deviation between scheduling strategies and actual power optimization requirements.

[0075] Step 204: Obtain the task priority of the task queue in each server.

[0076] Specifically, in step 204, by interacting with the server cluster's task management system, priority information for all pending tasks in each server's task queue is obtained in real time. Task priorities are divided based on preset rules (such as task service level agreements (SLAs), user permissions, and task types). For example, high-priority tasks may include real-time data processing and critical business requests, while low-priority tasks may include offline data analysis and log backup. This priority information is bound to the corresponding task in the form of structured data (such as tags or numerical rankings) and stored in the scheduling engine's task information database, serving as a constraint input for the subsequent multi-objective optimization solution.

[0077] By clarifying task priorities, the dynamic scheduling engine prioritizes the performance of high-priority tasks (e.g., reducing migration delays and avoiding resource preemption) when generating resource scheduling policies. This addresses the "energy efficiency and performance imbalance" and ensures that power consumption is optimized without impacting critical business operations. Task priorities are incorporated as constraints into the cluster energy cost model, ensuring that the weights assigned to optimization objectives (e.g., minimizing power consumption and meeting SLAs for high-priority tasks) are more aligned with actual business needs. Compared to traditional indiscriminate scheduling, this avoids policy failures caused by confusion about task importance and enhances the practicality of scheduling results. Real-time access to task queue priority information enables timely response to dynamic changes in the task queue (e.g., the addition of new high-priority tasks), enabling dynamic adjustment of scheduling policies. This addresses the inability of static resource allocation to adapt to fluctuations in task priority and improves the system's adaptability to complex business scenarios.

[0078] Step 205 : constructing a cluster energy consumption cost model based on the power consumption prediction results of each server to calculate the predicted total power of the server cluster.

[0079] Specifically, in step 205, the dynamic scheduling engine constructs a cluster energy cost model based on the power consumption prediction results of each server and the hardware parameters of each node in the cluster (such as the static power consumption baseline and the maximum power consumption threshold). This model uses a quantitative formula to calculate the predicted total power of the server cluster. The formula is as follows: =

[0080] in, is the total power predicted for the cluster; N is the total number of server nodes; is the current power consumption of the i-th server (the power consumption at the current moment collected by the data acquisition module); is the maximum power consumption of the i-th server (the upper limit of power consumption when the node is running at full load); is the load rate of the i-th server calculated based on the power consumption prediction result (calculated comprehensively by the predicted CPU utilization, memory occupancy, etc.).

[0081] During the model building process, the dynamic scheduling engine maps the power consumption prediction results of each server (such as power consumption change data within a preset time period) to the load rate of the corresponding node , substitute the above formula to calculate the predicted total power of the cluster as a whole, while retaining the power consumption composition details of each node, providing fine-grained energy consumption data support for subsequent multi-objective optimization.

[0082] Through the formulated cluster energy consumption cost model, the decentralized power consumption prediction results of each server are integrated into the total power calculation of the cluster level, solving the problem that the traditional method can only evaluate the power consumption of a single node and cannot quantify the overall energy consumption of the cluster, and providing a calculable quantitative basis for global energy efficiency optimization. Accurately reflect the correlation between load and power consumption: the load rate is introduced into the model The square term of the dynamic power consumption in CMOS circuits fits the nonlinear relationship between dynamic power consumption and load (frequency, voltage) (the document mentions that dynamic power consumption is proportional to the square of the voltage and the frequency). Compared to traditional linear models, this more accurately captures the actual impact of load changes on power consumption, improving the accuracy of total power prediction. It also supports efficient multi-objective optimization solutions: The model retains node-level power consumption details (such as static power consumption and dynamic power consumption contribution). This enables the dynamic scheduling engine to formulate targeted strategies (such as prioritizing sleep strategies for nodes with high static power consumption) during subsequent optimization processes (e.g., incorporating task priority constraints). This significantly improves the targetedness and efficiency of optimization solutions compared to black-box total power evaluation.

[0083] In step 206, a multi-objective optimization solution is performed on the task resource scheduling of each server according to the principle of minimizing the predicted total power and combining the task priority constraint. The multi-objective optimization solution is implemented by a genetic algorithm. The resource scheduling strategy includes at least one operation of task migration, node dormancy or frequency adjustment.

[0084] Specifically, in step 206, the dynamic scheduling engine minimizes the predicted total power calculated by the cluster energy cost model as its core optimization objective. Task priority constraints (such as the upper limit on response delays for high-priority tasks and the resource allocation threshold for low-priority tasks) are incorporated into the optimization framework, and a genetic algorithm is used to solve the multi-objective optimization problem. During the genetic algorithm implementation, the task allocation plan and operating status parameters (such as whether the server is dormant and the operating frequency level) of each server are used as genetic codes to construct an initial population. Using the "predicted total power reduction" and "task priority satisfaction" (such as the SLA achievement rate for high-priority tasks) as fitness functions, the population is subjected to selection, crossover, and mutation operations. Constraints (such as the upper limit on the number of task migrations per time) are also set to filter out solutions that do not meet actual operational requirements. After multiple generations of iteration, the algorithm outputs the optimal solution, which is a resource scheduling strategy that includes specific operations: task migration: allocating some tasks of high-load nodes (low-priority tasks are migrated first) to low-load nodes to balance the node load; node sleep: for nodes whose predicted idle time exceeds the threshold, switch to deep sleep mode to reduce static power consumption; frequency adjustment: through dynamic voltage frequency scaling (DVFS) technology, reduce the processor operating frequency for nodes with medium and low load rates to reduce dynamic power consumption.

[0085] Through multi-objective optimization, the genetic algorithm minimizes predicted total power while ensuring the execution of high-priority tasks. This addresses the conflict between global load balancing and energy optimization, and the inability to balance energy efficiency and performance. Compared to single-objective optimization, this significantly improves the strategy's practical applicability. The introduction of constraints (such as a limit on the number of migrations) avoids execution risks caused by over-optimization (such as network congestion caused by excessive task migrations). The combined operation of task migration, node hibernation, and frequency adjustment covers the main power optimization scenarios for multi-node server clusters, achieving more significant optimization results than traditional single-adjustment methods (the example in the document shows a power consumption reduction of up to 40%).

[0086] The iterative solution process of the genetic algorithm can quickly respond to changes in power consumption prediction results. Combined with the dynamic update of task priorities, the scheduling strategy can adapt to fluctuations in cluster load (such as the sudden influx of high-priority tasks) in real time, solving the problem that traditional static rule scheduling cannot adapt to dynamic workloads.

[0087] In step 207 , the cluster controller performs resource scheduling on each server in the server cluster according to the resource scheduling policy to adjust and optimize the power consumption of each server.

[0088] Specifically, in step 207, the cluster controller receives the resource scheduling policy output by the dynamic scheduling engine, breaks down the policy into specific execution instructions for each server node, and sends them to the corresponding node through the cluster management interface (such as an IPMI dedicated channel). Scheduling operations include: Task migration execution: According to the task migration plan specified by the policy, tasks that meet the migration conditions (such as low-priority tasks) in the task queue of the high-load node are transferred to the low-load node through the cluster's internal communication protocol. During the migration process, the task status is monitored in real time to ensure data consistency; Node state control: For nodes marked as "idle" by the policy, deep sleep mode is triggered (non-essential hardware components are turned off to reduce static power consumption), and the fast wake-up mechanism is retained (preset wake-up signals are used to shorten recovery time); Hardware parameter adjustment: For nodes that require frequency adjustment, the DVFS interface is called to dynamically reduce the CPU / GPU operating voltage and clock frequency. The adjustment range is determined based on the load rate prediction results in the policy.

[0089] During the scheduling process, the cluster controller continuously collects execution feedback data from each node (such as task migration completion rate, node power consumption change, and task response latency) and feeds this data into the dynamic scheduling engine and power consumption prediction module as input for iterative policy optimization and incremental model learning. The cluster controller uniformly executes cross-node scheduling policies, breaking the limitations of traditional single-node independent control and addressing the issue of "not considering multi-node collaborative optimization," significantly reducing overall cluster power consumption. Task migration prioritizes neighboring low-load nodes, retains a fast wake-up mechanism for dormant nodes, and adjusts frequency to match load demand, all of which optimize power consumption while limiting the impact on service performance.

[0090] To train the power prediction model, the disclosed embodiment uses a data acquisition module to acquire the server's operating status data at different times. This operating status data is processed and divided into training data. Before using the training data to train the power prediction model, the training data may be preprocessed using, but not limited to, the following methods, including at least one of sliding window normalization, graph neural network embedding, and fault mode feature extraction.

[0091] After preprocessing the training data, the power prediction model can be trained in the following ways, but not limited to: inputting the training data into the load prediction sub-model of the power consumption prediction model to be trained for training to generate a training load prediction result; inputting the training load prediction result into the power consumption prediction sub-model of the power consumption prediction model to be trained for training to generate a training power consumption prediction result; dynamically time-regulating the training load prediction result and the training power consumption prediction result to generate a training prediction result; generating incremental training data based on the training prediction result and the training scheduling result, and using the incremental training data to perform incremental learning training on the power consumption prediction model to be trained.

[0092] Specifically, to effectively train the power prediction model, this embodiment uses a data acquisition module to acquire server operating status data at different times (including time-series data such as CPU utilization, memory usage, network traffic, node power consumption, and ambient temperature). After processing, this data is divided into training data for model training. The specific process is as follows:

[0093] Before using training data to train the power prediction model, the data must be preprocessed using at least one of sliding window normalization, graph neural network embedding, and fault mode feature extraction. Sliding window normalization: This involves segmenting time series data into continuous subsequences based on a preset time window (e.g., 1-minute units). The feature values ​​of each subsequence (e.g., CPU utilization, power consumption) are normalized (mapped to the range 0-1) to eliminate magnitude deviations caused by hardware differences between servers and ensure input feature consistency. Graph neural network embedding: This involves treating a multi-node server cluster as a topological graph (nodes represent servers, and edges represent inter-node communication relationships). A graph neural network is used to extract spatial correlation features of nodes (e.g., load coupling relationships between adjacent nodes), compensating for the fact that single-node time series data often ignores cluster coordination. Fault mode feature extraction: Feature engineering is used to extract fault mode labels (e.g., "burst traffic" or "memory leak risk") for abnormal conditions (e.g., transient high loads, precursors to hardware failures) seen in historical data, enhancing the model's adaptability to non-steady-state scenarios.

[0094] The preprocessed training data is used to train the power prediction model (including the load prediction sub-model LSTM1 and the power consumption prediction sub-model LSTM2). The specific steps are as follows: Load prediction sub-model training: Load features (such as CPU / GPU utilization, memory bandwidth, and network traffic) from the training data are fed into the trained LSTM1. A three-layer LSTM network learns the temporal dependencies of the loads. The attention layer enhances the features of key time steps, and the fully connected layer outputs the training load prediction results (e.g., a series of load rates for the next five minutes). Power consumption prediction sub-model training: The training load prediction results output by LSTM1 are fed into the trained LSTM2. The historical power consumption features from the training data (e.g., power consumption baselines under different loads) are also integrated. The mapping between load and power consumption is learned using the same network structure, and the training power consumption prediction results are output (e.g., a series of power consumption for the corresponding time period). Dynamic Time Warping (DTW) alignment: DTW is performed on the training load prediction results and the training power consumption prediction results to correct for any timing misalignments between the two series (e.g., the delay between load changes and power consumption responses). This generates spatiotemporally aligned training prediction results, improving the model's accuracy in learning the load-power consumption relationship. Incremental learning and training: The training predictions generated during model training and actual scheduling results (such as actual power consumption data after task migration and frequency adjustment) are added as new samples to the training queue, triggering the incremental learning mechanism. Model parameters (such as the hidden layer weights and attention layer weights of LSTM) are iteratively optimized to continuously adapt the model to the dynamic operating characteristics of the server cluster. During training, historical load and power consumption data from a preset time period is used as the sample set, divided into training, validation, and test sets in a 7:2:1 ratio. The mean squared error (MSE) is used as the loss function, and early stopping is used to prevent model overfitting.

[0095] It should be noted that the embodiments of the present disclosure may include multiple steps. For the convenience of description, these steps are numbered, but these numbers do not limit the execution time slots or execution order between the steps; these steps can be implemented in any order, and the embodiments of the present disclosure do not limit this.

[0096] Corresponding to the above-mentioned method for optimizing server power consumption, the present disclosure also provides a system for optimizing server power consumption. Since the system embodiment of the present disclosure corresponds to the above-mentioned method embodiment, details not disclosed in the system embodiment can be referred to the above-mentioned method embodiment and will not be further described in this disclosure.

[0097] Figure 3 A schematic diagram of a server power consumption optimization system provided by an embodiment of the present disclosure is shown in FIG. Figure 3As shown, it includes: a data acquisition module 31, a data processing module 32, a power consumption prediction module 33, a dynamic scheduling engine 34, and a cluster controller 35;

[0098] The data collection module 31 collects the operating status data of each server in the server cluster and sends the operating status data to the data processing module 32;

[0099] The data processing module 32 performs data preprocessing on the operating status data; the power consumption prediction module 33 performs model inference or model training on the power consumption prediction model according to the preprocessed operating status data sent by the data processing module 32;

[0100] The dynamic scheduling engine 34 generates a resource scheduling strategy or a training scheduling result for each server based on the power consumption prediction result or the training prediction result generated by the power consumption prediction model;

[0101] The cluster controller 35 performs resource scheduling on each server according to the resource scheduling policy to optimize power consumption.

[0102] It should be noted that the above explanation of the method embodiment is also applicable to the system of this embodiment, and the principles are the same, which is no longer limited in this embodiment.

[0103] For descriptions of features in the embodiments corresponding to the system for optimizing server power consumption, reference may be made to the relevant descriptions of the embodiments corresponding to the method for optimizing server power consumption, which will not be detailed here.

[0104] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above-mentioned server power consumption optimization method embodiments.

[0105] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the steps of any of the above-mentioned server power consumption optimization method embodiments when running.

[0106] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0107] An embodiment of the present application further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of any of the above-mentioned server power consumption optimization method embodiments are implemented.

[0108] An embodiment of the present application also provides another computer program product, including a non-volatile computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps in any of the above-mentioned server power consumption optimization method embodiments.

[0109] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0110] The above is a detailed introduction to a method and system for optimizing server power consumption, an electronic device, and a storage medium provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for optimizing server power consumption, characterized in that: include: The data collection module collects the operating status data of each server in the server cluster, wherein the operating status data includes at least one of the node power consumption value, processor utilization, memory occupancy, network traffic, ambient temperature and task type label collected in real time; The power consumption prediction module inputs the operating status data into the power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates the power consumption prediction results of each server. The power consumption prediction model includes a load prediction sub-model and a power consumption prediction sub-model, including: inputting the operating status data into the load prediction sub-model to generate the load prediction results of each server; inputting the load prediction results into the power consumption prediction sub-model to generate the power consumption prediction results of each server; The dynamic scheduling engine establishes a cluster energy consumption cost model based on the power consumption prediction results, and performs multi-objective optimization solution in combination with task priority constraints to generate a resource scheduling strategy for each server, including: obtaining the task priority of the task queue in each server; Based on the power consumption prediction results of the servers, constructing the cluster energy consumption cost model to calculate the predicted total power of the server cluster; According to the principle of minimizing the predicted total power and combining the task priority constraints, a multi-objective optimization solution is performed on the task resource scheduling of each server, wherein the multi-objective optimization solution is implemented by a genetic algorithm, and the resource scheduling strategy includes at least one operation of task migration, node dormancy, or frequency adjustment; The cluster controller performs resource scheduling on each server in the server cluster according to the resource scheduling policy, so as to adjust and optimize the power consumption of each server.

2. The method for optimizing server power consumption according to claim 1, wherein: The data collection module collects the operating status data of each server in the server cluster, including: The data acquisition module obtains the operating status data of each server; The operation status data of each server at different times are recorded and retained.

3. The method for optimizing server power consumption according to claim 1, wherein: After the data collection module collects the operating status data of each server in the server cluster, the method further includes: The data processing module performs at least one processing method of sliding window normalization processing, graph neural network embedding, and fault mode feature extraction on the operating status data to perform data preprocessing on the operating status data.

4. The method for optimizing server power consumption according to claim 1, wherein: The method further comprises: The training module constructs training data based on the operating status data collected by the data collection module; inputs the training data into the power consumption prediction model to be trained to perform model training, and obtains a trained power consumption prediction model.

5. The method for optimizing server power consumption according to claim 4, wherein: Inputting the training data into the power consumption prediction model to be trained for model training includes: Inputting the training data into the load prediction sub-model of the power consumption prediction model to be trained for training, and generating a training load prediction result; Inputting the training load prediction result into the power consumption prediction sub-model of the power consumption prediction model to be trained for training to generate a training power consumption prediction result; Performing dynamic time warping on the training load prediction result and the training power consumption prediction result to generate a training prediction result; Based on the training prediction result and the training scheduling result, incremental training data is generated, and the incremental training data is used to perform incremental learning training on the power consumption prediction model to be trained.

6. A system for optimizing server power consumption, characterized in that: include: Data acquisition module, data processing module, power consumption prediction module, dynamic scheduling engine, cluster controller; The data acquisition module collects the operating status data of each server in the server cluster and sends the operating status data to the data processing module, wherein the operating status data includes at least one of the node power consumption value, processor utilization, memory occupancy, network traffic, ambient temperature and task type label collected in real time; The data processing module performs data preprocessing on the operating status data; the power consumption prediction module performs model inference or model training on the power consumption prediction model based on the preprocessed operating status data sent by the data processing module, inputs the operating status data into the power consumption prediction model, predicts the power consumption change data of each server within a preset time period, and generates the power consumption prediction results of each server. The power consumption prediction model includes a load prediction sub-model and a power consumption prediction sub-model, including: inputting the operating status data into the load prediction sub-model to generate the load prediction results of each server; inputting the load prediction results into the power consumption prediction sub-model to generate the power consumption prediction results of each server; The dynamic scheduling engine generates a resource scheduling strategy or a training scheduling result for each server based on the power consumption prediction result or the training prediction result generated by the power consumption prediction model, including: obtaining a task priority of a task queue in each server; Based on the power consumption prediction results of the servers, constructing the cluster energy consumption cost model to calculate the predicted total power of the server cluster; According to the principle of minimizing the predicted total power and combining the task priority constraints, a multi-objective optimization solution is performed on the task resource scheduling of each server, wherein the multi-objective optimization solution is implemented by a genetic algorithm, and the resource scheduling strategy includes at least one operation of task migration, node dormancy, or frequency adjustment; The cluster controller performs resource scheduling on each server according to the resource scheduling policy to optimize power consumption.

7. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method for optimizing server power consumption according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to enable the computer to execute the method for optimizing server power consumption according to any one of claims 1 to 5.

Images