A server cluster-oriented adaptive power management method and system

Abstract

The application relates to a server cluster-oriented adaptive power management method and system, and relates to the technical field of server cluster power management and energy efficiency optimization. Through acquisition of running state data of each node of the cluster and generation of a load prediction sequence, the running state and the prediction sequence are input into an upper-layer joint decision model to integrally decide power budget allocation and task scheduling, and each node power budget and task scheduling instructions are generated; based on the allocated power budget, a lower-layer prediction control model and an online-updated node power-performance response model are used for rolling optimization of a frequency adjustment sequence to generate a node frequency adjustment instruction; task scheduling and power consumption control are performed according to the frequency and the scheduling instruction, and actual power consumption and performance data are collected and fed back to the upper-layer model to update a decision strategy. Through double-layer cooperative closed-loop control, power-task integrated decision and node adaptive fine-grained frequency adjustment are realized, and the overall energy efficiency of the cluster is effectively improved under the condition of meeting differentiated service quality constraints.

Description

Technical Field

[0001] This invention relates to the field of server cluster power management and energy efficiency optimization technology, specifically to an adaptive power management method and system for server clusters. Background Technology

[0002] With the continued deployment of cloud computing, edge computing, and large-scale offline computing services, multi-node, heterogeneous server clusters have become the core infrastructure for various computing services. Especially under the "computing as a service" model, elastic server clusters generally exhibit operational characteristics such as drastic load fluctuations, heterogeneous task types (including both latency-sensitive and batch processing tasks), and significant differences in node hardware characteristics. Against this backdrop, data centers have introduced standardized management requirements for overall cluster power consumption constraints, hardware operational stability, and the quality of service for differentiated services. Within a limited power budget, and considering the real-time operating status of nodes and dynamic load changes, refined and adaptive power control and task allocation management are needed to achieve a synergistic balance between cluster energy efficiency and service performance.

[0003] However, current server clusters generally adopt a hierarchical and independent management and control model. Cluster-level power budget allocation, service task scheduling, and node-level hardware frequency adjustment are isolated from each other, and the various control links lack a unified linkage decision-making mechanism and timing coordination relationship. Existing management and control methods mostly rely on fixed control rules for static adjustment, making it difficult to combine short-term load forecast trends and changes in node hardware conditions for global overall planning. They cannot achieve fine-grained dynamic adaptation based on overall constraints. It is difficult to balance the overall energy efficiency benefits of the cluster with the operational guarantee requirements of different types of services under limited power resources, and it cannot effectively adapt to the differentiated characteristics of heterogeneous nodes and the dynamic fluctuations of load. At the same time, there is a lack of effective closed-loop feedback and policy linkage mechanisms between cluster-level decision-making and node-level control. As a result, global power allocation and task scheduling strategies cannot be dynamically optimized according to actual execution effects, making it difficult to continuously ensure the balance of overall energy efficiency and service quality of the cluster in scenarios with drastic load fluctuations.

[0004] Therefore, in order to meet the needs of refined and adaptive power management under large-scale and heterogeneous deployment of clusters, make up for the adaptation shortcomings of existing management and control modes, and overcome the technical bottlenecks of insufficient coordination in power allocation, task scheduling and node control, there is an urgent need in this field for an adaptive power management method and system for server clusters. Summary of the Invention

[0005] The main objective of this invention is to provide an adaptive power management method for server clusters, which aims to build a two-layer collaborative management and control system that combines global cluster decision-making with local node control. This system achieves integrated linkage between power budget allocation and task scheduling, and combines dynamic node model updates and rolling optimization and control to balance cluster performance and overall power consumption efficiency, thereby improving the adaptive adjustment capability and scenario adaptability of server cluster power management.

[0006] To overcome the shortcomings of existing technologies and achieve the above objectives, an adaptive power management method for server clusters is provided, the method comprising: S1: Obtain the running status data of each node in the server cluster, and generate a load prediction sequence for the future control time domain based on the running status data. The running status data includes the CPU utilization, memory utilization, current power consumption and task queue status of each node. S2: Input the operating status data and the load prediction sequence into the upper-level joint decision model, and make an integrated decision on power budget allocation and task scheduling through the upper-level joint decision model to generate the power budget and task scheduling instructions allocated to each node. S3: Based on the allocated power budget, each node generates frequency regulation commands for itself by using the locally deployed lower-level predictive control model and the online-updated node power-performance response model to perform rolling optimization of the frequency regulation sequence. S4: Based on the frequency adjustment command and the task scheduling command, perform task scheduling and power consumption control, collect the actual power consumption data and performance data after execution, and feed the actual power consumption data and performance data back to the upper-level joint decision model to update its decision strategy.

[0007] Preferably, in step S2, the integrated decision-making process for power budget allocation and task scheduling through the upper-level joint decision-making model includes: S21: The upper-level joint decision-making model models the cluster total power budget allocation and task scheduling as a unified Markov decision process; S22: Based on the Markov decision process, construct a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying service quality constraints; S23: Use the operating status data and the load prediction sequence as the state input of the Markov decision process, solve the Markov decision process, generate the power budget allocation value and task allocation result of each node, and map the task allocation result into the corresponding task scheduling instruction.

[0008] Preferably, in step S22, constructing a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying service quality constraints includes: S221: Obtain the type identifier of each task in the current task queue waiting to be scheduled, wherein the type identifier at least distinguishes between latency-sensitive tasks and batch processing tasks; S222: Based on the type distribution of the task queue, dynamically adjust the weight distribution between the performance optimization term and the power consumption optimization term in the multi-objective reward function; S223: The adjusted multi-objective reward function is used as the solution objective of the Markov decision process.

[0009] Preferably, in step S23, solving the Markov decision process includes: S231: Define the power budget allocation value and task allocation result of each node as the joint action space of the Markov decision process; S232: Solve the Markov decision process using a policy learning algorithm to obtain the optimal joint action policy; S233: Based on the optimal joint action strategy, output the power budget allocation value and task allocation result of each node simultaneously.

[0010] Preferably, in step S3, the rolling optimization of the frequency regulation sequence using the locally deployed lower-level predictive control model and the online-updated node power-performance response model includes: S31: Load the node power-performance response model updated in the previous control cycle as the current prediction model; S32: Based on the task type currently running on this node, construct a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term; S33: Using the current prediction model, predict the node power consumption and processing performance under different frequency settings in the future control time domain; S34: Based on the multi-objective rolling optimization function and the prediction results, solve the optimization problem that satisfies the power budget constraint allocated to this node, and generate the optimal frequency adjustment sequence; S35: Use the first frequency value in the optimal frequency adjustment sequence as the frequency adjustment command for this node; S36: Collect the actual operating data after the execution of this control cycle, update the node power-performance response model, and use it as the current prediction model for the next control cycle.

[0011] Preferably, in step S32, constructing a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term includes: S321: Identify the type of task currently running on this node, wherein the task type includes at least latency-sensitive tasks and batch processing tasks; S322: Based on the task type of this node, adaptively adjust the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function; S323: The adjusted multi-objective rolling optimization function is used as the objective of the optimization problem.

[0012] Preferably, in step S36, collecting the actual operating data after the execution of this control cycle and updating the node power-performance response model includes: S361: Extract the real-time power consumption data, operating frequency data and resource utilization data collected in this control cycle to form a model update sample; S362: The parameters of the node power-performance response model are incrementally updated using the model update samples by employing a recursive gradient descent method. S363: Store the updated node power-performance response model to local memory.

[0013] Preferably, in step S4, feeding back the actual power consumption data and performance data to the upper-level joint decision-making model to update its decision-making strategy includes: S41: The lower-level predictive control model of each node integrates the actual power consumption data, performance data and the latest parameters and parameter change data of the node power-performance response model into global feedback data. S42: Input the global feedback data into the upper-level joint decision-making model to evaluate the actual effect of the current decision-making strategy; S43: Based on the evaluation results, update the decision strategy parameters of the upper-level joint decision-making model.

[0014] Preferably, before performing step S3, the method further includes: S301: The upper-level joint decision-making model generates task type prediction information for each node in the future control time domain based on the load prediction sequence; S302: The upper-level joint decision-making model will send the power budget, task allocation results and task type prediction information allocated to each node to the lower-level predictive control model of each node. S303: The lower-level predictive control model uses the power budget allocated to this node as the upper limit of the power constraint for rolling optimization. It determines the execution sequence of future tasks based on the received task allocation results and, in conjunction with the task type prediction information, pre-adjusts the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes an adaptive power management system for server clusters, the adaptive power management system for server clusters comprising: a memory, a processor, and an adaptive power management program for server clusters stored in the memory and executable on the processor, the adaptive power management program for server clusters being configured to implement the steps of the adaptive power management method for server clusters.

[0016] The beneficial effects of this invention are: 1. It constructs a two-layer closed-loop management and control architecture that deeply coordinates cluster global decision-making and node local control, and models power budget allocation and task scheduling as a unified decision-making process to achieve integrated linkage solution. It fundamentally solves the problems of cluster power control, task scheduling, node frequency adjustment layered fragmentation, serial decision-making leading to the inability to implement global optimization goals, and the mismatch between power resources and business computing power requirements, and realizes global overall optimization of cluster resources.

[0017] 2. Based on the node-local online dynamic update power-performance response model, combined with the finite time domain rolling optimization mechanism, fine-grained adaptive adjustment of node-level power consumption is achieved. This solves the problem that fixed control rules cannot adapt to the drift of node hardware characteristics and dynamic load fluctuations. Under the premise of strictly adhering to global power budget constraints, it effectively takes into account the differentiated service quality assurance requirements of heterogeneous services.

[0018] 3. A two-way closed-loop feedback mechanism from node execution data to global decision optimization has been established, enabling continuous iteration and adaptation of control strategies. This solves the problems of lacking complete closed-loop feedback, being unable to adapt to complex scenarios such as configuration changes, hardware aging, and operating condition fluctuations during long-term cluster operation, and improves the long-term operational stability, scenario compatibility, and adaptability of the power management system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the steps of the adaptive power management method for server clusters according to an embodiment of the present invention; Figure 2 This is a structural diagram of the adaptive power management system for server clusters as described in an embodiment of the present invention.

[0020] Reference numerals 10 and 20 are used to indicate memory and processor, respectively. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described content is only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] The features and effects of the present invention will be further described in detail below with reference to the embodiments.

[0023] like Figure 1 As shown, this invention provides an adaptive power management method for server clusters, comprising: S1: Obtain the running status data of each node in the server cluster, and generate a load prediction sequence for the future control time domain based on the running status data. The running status data includes the CPU utilization, memory utilization, current power consumption and task queue status of each node. S2: Input the operating status data and the load prediction sequence into the upper-level joint decision model, and make an integrated decision on power budget allocation and task scheduling through the upper-level joint decision model to generate the power budget and task scheduling instructions allocated to each node. S3: Based on the allocated power budget, each node generates frequency regulation commands for itself by using the locally deployed lower-level predictive control model and the online-updated node power-performance response model to perform rolling optimization of the frequency regulation sequence. S4: Based on the frequency adjustment command and the task scheduling command, perform task scheduling and power consumption control, collect the actual power consumption data and performance data after execution, and feed the actual power consumption data and performance data back to the upper-level joint decision model to update its decision strategy.

[0024] In this embodiment, the upper-layer joint decision-making model refers to the cluster-level global decision-making model deployed on the cluster management node. It is used to achieve global allocation of power budget and global planning of task scheduling at the cluster level. It has a configurable fixed decision cycle, which can be adapted and adjusted according to the macro-level changes in cluster services. In typical general scenarios, the decision cycle can be configured to 5 minutes to adapt to changes in the cluster's macro-level load. For ultra-low latency service scenarios, it can be configured to 1 minute, and for ultra-large-scale clusters with more than 1000 nodes, it can be configured to 10 minutes. The lower-layer predictive control model refers to the node-level real-time control model deployed locally on each computing node. It is used to achieve frequency adjustment and power consumption control at the single-node level. It has a configurable fixed control cycle, and its control cycle is a fixed positive integer multiple of the decision cycle of the upper-layer joint decision-making model, adapting to the collaborative needs of changes in the cluster's macro-level load and fine-grained real-time control of nodes. In typical scenarios, the upper-layer 5-minute decision cycle corresponds to the lower-layer 10-second control cycle (50x matching), and in ultra-low latency scenarios, the upper-layer 1-minute decision cycle corresponds to the lower-layer 2-second control cycle (30x matching). That is, let the upper-layer decision cycle be T. u The lower-level control cycle is T. l Satisfying T u =k×T l k is a positive integer; the timing synchronization mechanism is as follows: each upper-level decision cycle T uDivided into k consecutive lower-level control cycles T l The upper-level joint decision-making model in the kth T l In the first half of the cycle, global decisions are made, generating power budgets, task scheduling instructions, and task type prediction information for each node, and in the k-th T... l The data is distributed to all nodes during the latter half of the cycle; the lower-level predictive control model of each node is deployed during the k-th T cycle. l Before the end, complete the parameter pre-adjustment, in the first T of the next upper-level decision cycle. l At the beginning, new global decision instructions are executed to ensure that the control timing of the upper and lower layers is fully aligned and there is no control gap.

[0025] The switching rules between upper and lower layers are as follows: after the upper layer issues new power budget and task allocation instructions, the lower layer predictive control model does not interrupt the currently executing control cycle. After the current control cycle is completed, new global decision parameters are loaded when the next control cycle starts, and new rolling optimization is performed. This ensures the integrity of the control logic within each lower layer control cycle and avoids performance jitter caused by frequent switching.

[0026] The fault tolerance rule for lost feedback data is as follows: If the global feedback data of a certain node is lost during the upload cycle, the upper-layer joint decision model will use the feedback data of the previous complete upload cycle of that node when updating the policy; if no feedback data is received from that node for three consecutive decision cycles, the node will be marked as an abnormal node, and new latency-sensitive tasks will be stopped from being assigned to it. The power budget will be reduced to 60% of the node's rated power until communication is restored and the feedback data is uploaded normally.

[0027] The node power-performance response model refers to a lightweight prediction model deployed locally on the computing node. It fits the nonlinear mapping relationship between node hardware power consumption, processing performance, and real-time operating parameters. It can be dynamically updated online according to the node's operating status. It is constructed using a two-layer fully connected neural network. The inputs are node CPU utilization, operating frequency, and memory utilization; the outputs are the node's predicted power consumption and predicted processing performance. The load prediction sequence refers to node load change trend data in the future control time domain, generated based on historical and real-time node operating status data. Its time domain length covers at least one complete upper-level decision cycle and all lower-level control cycles contained within that cycle. It is generated using an LSTM time series prediction model. The input is the node's historical operating data for the past 24 hours; the output time domain length is completely consistent with the upper-level decision cycle.

[0028] This embodiment employs a two-layer control architecture that combines cluster-level global decision-making with node-level local control. The output of the upper-layer joint decision-making model serves as the input constraint for the lower-layer predictive control model, while the actual execution results of the lower-layer predictive control model serve as the optimization basis for the upper-layer joint decision-making model, forming a complete closed-loop control link. Operating status data is collected in real-time through the out-of-band management channels of each computing node and the standard performance monitoring interface of the operating system. The collection frequency matches the control cycle of the lower-layer predictive control model. The out-of-band management channels can read the node's overall power consumption and hardware temperature data via the standard IPMI protocol, adapting to the native BMC interface of mainstream server manufacturers. Data within the operating system can be accessed through standard system calls in the Linux system's / proc directory. For data reading, Windows systems can use the built-in performance counter API, eliminating the need for additional low-level data acquisition programs. For common abnormal scenarios such as node offline or network communication interruptions, offline nodes execute a default conservative frequency control strategy. Once communication is restored, the latest global decision information is synchronized to prevent uncontrolled node power consumption. For example, the conservative frequency control strategy locks the node's CPU frequency to 60% of its rated frequency by default, preventing uncontrolled power consumption while ensuring the availability of basic services. For abnormal scenarios where the CPU temperature exceeds a safe threshold, a forced frequency derating operation is triggered. For example, when the CPU temperature exceeds 85°C, the frequency is reduced by 5% every control cycle until the temperature drops below 75°C, prioritizing hardware safety.

[0029] This embodiment provides two feasible two-tier architecture implementation schemes. The first is a highly reliable master-slave architecture: the upper-layer joint decision-making model is deployed on a distributed decision-making cluster consisting of three master and slave management nodes, using the Raft consensus protocol to synchronize decision data. When a single management node fails, the remaining nodes automatically take over, avoiding uncontrolled cluster power management caused by a single point of failure. The lower-layer predictive control model is deployed on the out-of-band management controller of the node, without occupying business computing resources. The second is a large-scale partitioned architecture: the upper-layer joint decision-making model adopts a distributed decision-making architecture based on availability zones. Partition decisions are first completed in each sub-region, and then the master management node completes global convergence, avoiding overload of single-node computing pressure. The lower-layer predictive control model supports differentiated control cycle configuration for single nodes. High-load nodes use a 2-second control cycle to ensure response speed, while low-load nodes use a 30-second control cycle to reduce computing overhead. The core processes and closed-loop logic of the two schemes are completely consistent, fully covering the implementation needs of clusters of different sizes.

[0030] This embodiment constructs a two-layer closed-loop control architecture that coordinates upper-layer global decision-making with lower-layer local control. It integrates cluster power budget allocation and task scheduling into a unified decision-making process. Combined with node-level fine-grained frequency rolling optimization control, it achieves deep collaboration between global optimization and local adaptive adjustment of cluster power management. This solves the core technical problems in existing technologies, such as the disconnect between cluster power allocation and task scheduling, mismatch in upper and lower layer control timing, and the inability to balance cluster energy efficiency and service quality under dynamic load scenarios. Through load prediction and closed-loop feedback optimization mechanisms, it can adaptively adapt to dynamic changes in cluster load, improving overall cluster energy efficiency while ensuring cluster service quality. It also provides implementation solutions adaptable to multiple scenarios, possessing strong scenario adaptability and engineering implementation value.

[0031] In one feasible implementation, step S2, which involves making an integrated decision on power budget allocation and task scheduling through the upper-level joint decision model, includes: S21: The upper-level joint decision-making model models the cluster total power budget allocation and task scheduling as a unified Markov decision process; S22: Based on the Markov decision process, construct a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying service quality constraints; S23: Use the operating status data and the load prediction sequence as the state input of the Markov decision process, solve the Markov decision process, generate the power budget allocation value and task allocation result of each node, and map the task allocation result into the corresponding task scheduling instruction.

[0032] In step S22, constructing a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying quality of service constraints includes: S221: Obtain the type identifier of each task in the current task queue waiting to be scheduled, wherein the type identifier at least distinguishes between latency-sensitive tasks and batch processing tasks; S222: Based on the type distribution of the task queue, dynamically adjust the weight distribution between the performance optimization term and the power consumption optimization term in the multi-objective reward function; S223: The adjusted multi-objective reward function is used as the solution objective of the Markov decision process.

[0033] In step S23, solving the Markov decision process includes: S231: Define the power budget allocation value and task allocation result of each node as the joint action space of the Markov decision process; S232: Solve the Markov decision process using a policy learning algorithm to obtain the optimal joint action policy; S233: Based on the optimal joint action strategy, output the power budget allocation value and task allocation result of each node simultaneously.

[0034] In this embodiment, integrated decision-making refers to the simultaneous modeling, solving, and output of two strongly coupled decision-making processes: the allocation of the cluster's total power budget among nodes and the scheduling of cluster tasks among nodes, rather than a serial, step-by-step execution approach. This achieves deep linkage and global optimization between the two decision-making processes. The joint Markov decision process specifically refers to the Markov decision model constructed to adapt to the coupled scenario of cluster power allocation and task scheduling in this invention. Its state space and action space simultaneously cover both power allocation and task scheduling dimensions, achieving unified modeling of the two decision-making processes. The state space includes the upper limit of the cluster's global total power budget, the real-time operating status data of each node, the future load trend corresponding to the load prediction sequence, and the full information of the current queue of tasks to be scheduled, ensuring that the modeling covers all influencing factors of the two decision-making processes. When the task allocation result is mapped to the task scheduling instruction, the global allocation result is decomposed into the task execution queue of a single node, the task start sequence, and the task resource reservation quota information. Standardized scheduling instructions are generated according to the preset priority of the tasks and issued to the corresponding computing nodes, adapting to the instruction formats of mainstream cluster scheduling frameworks such as K8s and YARN.

[0035] In this embodiment, the task type identifier can be directly obtained through the metadata tag submitted with the task. For example, tasks tagged with "latency_sensitive" are classified as latency-sensitive tasks, and tasks tagged with "batch_processing" are classified as batch processing tasks. Tasks without metadata tags can be automatically classified based on their runtime characteristics: tasks with a single-task runtime < 1 minute and an end-to-end latency threshold < 100ms are classified as latency-sensitive tasks, and tasks with a single-task runtime > 10 minutes and no strict latency requirements are classified as batch processing tasks. This automatic classification requires no manual annotation. For first-time runs without metadata tags, a temporary classification rule is used by default: tasks with a single-task preset runtime < 1 minute and an end-to-end latency threshold < 100ms are classified as latency-sensitive tasks, and tasks with a single-task preset runtime > 10 minutes and no strict latency requirements are classified as batch processing tasks. The classification results are updated based on the actual runtime characteristics after the task is completed. Latency-sensitive tasks include online business tasks with strict end-to-end latency requirements, while batch processing tasks include offline business tasks with lenient completion time requirements. The fixed expression for the multi-objective reward function is R = ω1. perf ×R perf +ω1 pow ×R pow , where ω1 perf ω1 is the weighting coefficient of the performance optimization term. powThese are the weighting coefficients for the power consumption optimization term; both are non-negative real numbers and satisfy the constraint ω1. perf +ω1 pow =1; To avoid performance drop or power consumption runaway caused by extreme weights, the value range of both is limited to [0.2, 0.8] to ensure a basic balance between performance and energy efficiency; Performance optimization item R perf Used to quantify the compliance of cluster service quality. The specific calculation method is as follows: In the current decision period, the ratio of the number of latency-sensitive tasks that have been completed in the cluster and whose end-to-end processing latency is lower than the preset SLA latency threshold of the task to the total number of latency-sensitive tasks completed in the same period. The value range is [0,1]. The higher the value, the better the compliance of cluster service quality. End-to-end processing latency refers to the total time from task submission to completion, which can be calculated using the native logs of the cluster scheduling system. The SLA latency threshold is the maximum allowable latency marked by metadata tags when the task is submitted. Unmarked tasks use the general default threshold for their respective task types (100ms for latency-sensitive tasks and 30min for batch processing tasks). Power consumption optimization term R pow Used to quantify the optimization of cluster energy efficiency; the specific calculation method is as follows: The value range is [-0.2, 1]; Among them, P act The average total power consumption of the cluster during the current decision-making cycle is calculated by aggregating the power consumption data collected by the BMC of each node; P total P represents the upper limit of the total power budget allocated by the data center for this cluster, a fixed configuration value for cluster management; when the actual total power consumption of the cluster exceeds the total power budget, P... act / P total >1, R pow If the value is negative, a power consumption overrun penalty is triggered, with a lower limit of -0.2 to avoid excessive penalty for a single decision causing strategy oscillation.

[0036] The directly implementable quantitative rule for dynamic weight adjustment is: when latency-sensitive tasks account for ≥70%, set ω1. perf =0.8、ω1 pow =0.2, prioritizing business performance; when batch processing tasks account for ≥80%, set ω1. perf =0.3、ω1 pow =0.7, prioritizing the improvement of cluster energy efficiency; in intermediate scenarios, weights are calculated by linear interpolation based on task proportion, which can adaptively adapt to different business scenarios without manual debugging.

[0037] In this embodiment, the joint action space refers to the integrated action set formed by binding the power budget allocation value of each node with the task allocation result. The power budget allocation value and the task allocation result are mutually constrained linkage variables, rather than two independent decision dimensions, to ensure the global adaptability of the decision result. To ensure the effectiveness of integrated decision-making, this embodiment sets a power-task strong coupling constraint for the joint action space. The core constraint formula is: the total computing power requirement for task allocation on a single node ≤ the maximum computing power that the node can provide under the power budget, so as to avoid invalid decisions where task allocation exceeds the node's power carrying capacity.

[0038] The specific quantification rules for the coupling constraints are as follows: Calculation of total computing power requirement for single-node task allocation: For all tasks allocated to this node, the computing power requirement is calculated separately for each task type and then summed. For latency-sensitive tasks, the single-task computing power requirement = peak computing power requirement × 1.2 (SLA redundancy coefficient); for batch processing tasks, the single-task computing power requirement = average computing power requirement. The task computing power requirement is obtained based on the average / peak CPU utilization of historical task data. For tasks without historical data, the requirement is determined based on the computing power specifications indicated at the time of task submission. Calculation of maximum available computing power under the power budget: Based on the node power-performance response model, the maximum stable computing power output value corresponding to the lowest operating frequency that satisfies the SLA latency threshold of the currently allocated task under the power budget constraint allocated to this node is calculated as the upper limit of the coupling constraint's computing power. Constraint effectiveness rules: During integrated decision-making, only when the node satisfies the above coupling constraints is the corresponding joint action considered a valid action and included in the strategy solution scope.

[0039] Policy learning algorithms refer to deep reinforcement learning algorithms that adapt to continuous state space and continuous action space. They have the characteristics of stable training and high sample utilization. They can adapt to the global decision-making needs of large-scale cluster scenarios. PPO (proximal policy optimization) algorithm is preferred. The training environment is a simulation environment built with historical load data of the cluster. The convergence condition is that the reward value fluctuation is ≤5% for 100 consecutive decision cycles.

[0040] This embodiment provides two exemplary implementation schemes for integrated decision-making. The first is a continuous space scheme for large-scale clusters: it adopts the aforementioned PPO algorithm and continuous state and action space design, which is suitable for medium-to-large-scale clusters with more than 100 nodes and can achieve refined global optimal decision-making. The second is a lightweight discrete space scheme for small-to-medium-scale clusters: it simplifies the state space and action space of the joint Markov decision-making process. The state space only retains the core information of the cluster's total power budget, the CPU utilization of each node, and the queue of tasks to be scheduled. The joint action space adopts a discrete design with fixed power budget levels and task allocation node numbers, transforming continuous space optimization into discrete space optimization. The solution algorithm adopts the Deep Q-Network (DQN) algorithm adapted to the discrete action space, reducing the training and inference computation by more than 60%. It can run directly on ordinary configuration management nodes without the need for dedicated GPU computing power. The decision cycle is fixed at 3 minutes, adapting to the load change speed of small-to-medium-scale clusters. The core logic of the two schemes is completely consistent, fully covering the deployment needs of clusters of different sizes.

[0041] This embodiment models the two tightly coupled decision-making processes of cluster power budget allocation and task scheduling as a unified joint Markov decision process. It constructs a multi-objective reward function that dynamically adjusts to adapt to task types, defines a joint action space that binds power allocation and task scheduling, and obtains the globally optimal decision result through a deep reinforcement learning strategy to solve the problem simultaneously. This completely solves the technical problem in the prior art where power allocation and task scheduling are disconnected due to serial step-by-step decision-making, and the global optimization objective cannot be achieved. By dynamically adjusting the weights of the multi-objective reward function according to task types, it can adaptively adapt to different business load scenarios of the cluster, achieving a dynamic balance between performance and energy efficiency. It also provides two adaptation schemes: fine-grained for large-scale clusters and lightweight for small and medium-sized clusters, and has strong robustness and scenario adaptability.

[0042] When the total power budget of the cluster is temporarily reduced, the upper-layer joint decision-making model triggers an emergency decision-making mode, adjusting the power budget and task scheduling according to the following priorities: First priority: Ensure that the SLA of all latency-sensitive tasks is met, allocate sufficient power budget to nodes running latency-sensitive tasks, and ensure that their minimum operating frequency is not lower than the frequency lower limit required by the task SLA; Second priority: Reduce the power budget of batch processing task nodes proportionally, and simultaneously suspend the scheduling of low-priority batch processing tasks until the total power budget allocation value of the cluster does not exceed the new total power budget upper limit; In emergency decision-making mode, the decision cycle is temporarily shortened to 1 / 2 of the original cycle to speed up the adjustment response speed. After the total power budget returns to stability, switch back to the normal decision-making mode.

[0043] In step S3, the rolling optimization of the frequency regulation sequence using the locally deployed lower-level predictive control model and the online-updated node power-performance response model includes: S31: Load the node power-performance response model updated in the previous control cycle as the current prediction model; S32: Based on the task type currently running on this node, construct a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term; S33: Using the current prediction model, predict the node power consumption and processing performance under different frequency settings in the future control time domain; S34: Based on the multi-objective rolling optimization function and the prediction results, solve the optimization problem that satisfies the power budget constraint allocated to this node, and generate the optimal frequency adjustment sequence; S35: Use the first frequency value in the optimal frequency adjustment sequence as the frequency adjustment command for this node; S36: Collect the actual operating data after the execution of this control cycle, update the node power-performance response model, and use it as the current prediction model for the next control cycle.

[0044] In step S32, constructing a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term includes: S321: Identify the type of task currently running on this node, wherein the task type includes at least latency-sensitive tasks and batch processing tasks; S322: Based on the task type of this node, adaptively adjust the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function; S323: The adjusted multi-objective rolling optimization function is used as the objective of the optimization problem.

[0045] In step S36, collecting the actual operating data after the execution of this control cycle and updating the node power-performance response model includes: S361: Extract the real-time power consumption data, operating frequency data and resource utilization data collected in this control cycle to form a model update sample; S362: The parameters of the node power-performance response model are incrementally updated using the model update samples by employing a recursive gradient descent method. S363: Store the updated node power-performance response model to local memory.

[0046] In this embodiment, rolling optimization refers to a finite-time-domain rolling optimization mechanism based on the model predictive control framework. Within each control cycle, the optimal frequency adjustment sequence for the future preset time domain is solved, and only the first control instruction in the sequence is executed. The next control cycle is re-solved based on the latest operating data, achieving real-time tracking and adaptive adjustment of node load fluctuations. The default prediction time domain length is set to 5 lower-level control cycles, ensuring both predictive foresight and avoiding excessive computational overhead and decreased prediction accuracy due to an excessively long prediction time domain. The initial state of the node power-performance response model is pre-trained using node hardware factory characteristic test data. The pre-training samples must cover CPU utilization from 0% to 100% and the entire frequency range, with at least 1000 effective samples. The pre-training convergence condition is a prediction power consumption error ≤ 5%. Frequency adjustment instructions are directly issued to the node CPU's standard power management module for execution, achieving real-time hardware-level frequency adjustment. The timing matching rules for frequency adjustment and task scheduling are as follows: For latency-sensitive tasks, the "frequency adjustment first, then startup" timing rule is executed: frequency adjustment is completed first according to the target frequency corresponding to the task type, and the task is started only after the CPU frequency stabilizes within ±2% of the target frequency, thus avoiding excessive latency caused by insufficient frequency during task startup. For batch processing tasks, the "frequency adjustment and task startup are executed synchronously" timing rule is executed, completing frequency adjustment at the same time as task startup, prioritizing the tracking accuracy of the power budget.

[0047] In this embodiment, the task type identification rule is consistent with the task type classification rule of the upper-level joint decision-making model, ensuring the consistency of the upper and lower-level control logic. The fixed expression of the multi-objective rolling optimization function is J=ω2. pow ×J pow +ω2 perf ×J perf +ω2 smooth ×J smooth The constraint is: the predicted node power consumption at all times in the prediction time domain is less than or equal to the node power budget allocated by the upper layer, where ω2 pow ω2 is the weighting coefficient of the node-level power tracking error penalty term. perf To maintain the weighting coefficient of the penalty term for node-level performance, ω2 smooth These are the weighting coefficients of the node-level frequency fluctuation suppression term. All three are non-negative real numbers and satisfy the constraint condition ω2. pow +ω2 perf +ω2 smooth =1; Power tracking error penalty term J pow This is used to quantify the tracking accuracy of node power consumption against the upper-layer allocated power budget. Specifically, it is calculated as the mean of the sum of squares of the differences between the predicted node power consumption and the upper-layer allocated node power budget for each control step within the prediction time domain. The formula is: Where n is the total number of control steps in the prediction time domain (default n=5, consistent with the length of the prediction time domain), P pred,k The predicted power consumption of the node at the k-th control step (calculated using the node power-performance response model), P budget This represents the upper limit of the power budget allocated to this node by the upper layer; the smaller the value, the higher the power tracking accuracy. Performance retention penalty J perf This is used to quantify the performance guarantee of node task processing. Specifically, it is calculated as the sum of the deviation penalties between the predicted task processing latency and the SLA latency threshold for each control step within the prediction time domain. The formula is: Among them, T pred,k T is the predicted processing delay of the task running within the k-th control step (calculated using the node power-performance response model). threshold This is the SLA delay threshold for this task, which is completely consistent with the threshold rules of the upper-level decision-making process. When the predicted delay is lower than the threshold, the penalty value is 0. When it exceeds the threshold, the penalty is amplified by the square of the deviation. The smaller the value, the better the task performance guarantee effect. Frequency fluctuation suppression term J smooth This is used to quantify the fluctuation range of node CPU frequency, avoiding hardware wear and performance jitter caused by frequent frequency adjustments. Specifically, it is calculated as the sum of the absolute values ​​of the differences between two adjacent control step frequency setpoints within the prediction time domain. The formula is: Among them, f k f is the CPU frequency setting value for the k-th control step. k-1 This is the CPU frequency setting value for the previous control step. When k=1, f0 is the final CPU frequency value of the node in the previous control cycle before the start of the current control cycle, serving as the initial reference value for the prediction time domain. A smaller value indicates less frequency fluctuation and better hardware stability. The implementable rule for adaptive weight adjustment is: when the node is running a latency-sensitive task, set ω2. perf =0.7、ω2 pow =0.2、ω2 smooth =0.1, prioritizing task performance; when the node runs batch processing tasks, set ω2. perf =0.2、ω2 pow =0.7、ω2 smooth =0.1, prioritizing power tracking accuracy and energy efficiency.

[0048] In this embodiment, the recursive gradient descent method refers to a parameter optimization method suitable for online incremental updates of lightweight models. It updates model parameters using only real-time samples collected in the current period, eliminating the need to store historical samples, thus consuming minimal computational resources. It is well-suited for real-time operation on local nodes and does not interfere with normal business execution. Specifically, incremental updates are triggered only when the model prediction error exceeds a preset threshold (default 5%). The update step size is adaptively adjusted based on the magnitude of the prediction error; the larger the error, the larger the step size, ensuring model update accuracy while reducing unnecessary computational overhead. The incremental update process is executed in the out-of-band management controller on the local node, without consuming computational resources for node business operations and without interfering with the execution of normal business tasks. Specifically, a recursive least squares algorithm with a forgetting factor can be used to complete the incremental update of model parameters. The forgetting factor λ is fixed at 0.95 to ensure that the weight of new samples is higher than that of historical samples, adapting to the slow drift of hardware characteristics. The adaptive step size rule is: learning rate η = η0 × (1 + |e| / e0), where η0 is the base learning rate (fixed at 1 × 10⁻⁶). -4 e represents the relative prediction error of the model in the current cycle, and e0 is the error trigger threshold (fixed at 5%), enabling adaptive adjustment where the larger the error, the larger the update step size. Simultaneously, update boundary constraints are set: the adjustment range of model parameters in a single incremental update must not exceed ±10% of the original parameters to avoid sudden changes in model parameters caused by single sample anomalies; when the prediction error is below 2% for three consecutive control cycles, incremental model updates are paused to reduce computational overhead.

[0049] This embodiment provides two implementation schemes for rolling optimization. The first is a general-purpose x86 computing node scheme: it adopts the aforementioned single-CPU frequency optimization logic and a 2-layer fully connected neural network model, which is suitable for most general-purpose server nodes. The second is a CPU / GPU heterogeneous computing node scheme: it expands the rolling optimization dimension, simultaneously optimizing the operating frequency of the CPU and GPU. The multi-objective rolling optimization function adds a GPU power consumption tracking term and a GPU performance maintenance term. The node power-performance response model simultaneously adds GPU operating parameters as input, which is suitable for the power management requirements of heterogeneous computing scenarios such as AI training and inference. The core processes and rolling optimization logic of the two schemes are completely consistent, fully covering the deployment requirements of different types of computing nodes.

[0050] This embodiment addresses the technical problem in existing technologies where node-level power control uses fixed models and parameters, making it unable to adapt to the time-varying drift and fine-grained load fluctuations of node hardware characteristics. By deploying a lower-level predictive control model based on model predictive control locally on the node, and constructing a multi-objective rolling optimization function adapted to the node's task type, it utilizes an online incrementally updated node power-performance response model to predict power consumption and performance. Frequency adjustment commands are generated through a finite-time domain rolling optimization mechanism. Through a single-step execution and rolling solution control method, real-time tracking and rapid response to node load fluctuations are achieved. While strictly adhering to the power budget allocated by the upper layer, it adaptively ensures task execution performance while avoiding hardware losses caused by frequent frequency fluctuations, improving the accuracy and adaptability of node-level power control, and covering the deployment needs of both general-purpose and heterogeneous computing nodes.

[0051] In a preferred embodiment, before performing step S3, the method further includes: S301: The upper-level joint decision-making model generates task type prediction information for each node in the future control time domain based on the load prediction sequence; S302: The upper-level joint decision-making model will send the power budget, task allocation results and task type prediction information allocated to each node to the lower-level predictive control model of each node. S303: The lower-level predictive control model uses the power budget allocated to this node as the upper limit of the power constraint for rolling optimization. It determines the execution sequence of future tasks based on the received task allocation results and, in conjunction with the task type prediction information, pre-adjusts the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function.

[0052] In this embodiment, the task type prediction information is generated based on the future task arrival time sequence corresponding to the load prediction sequence and the historical distribution data of task types. The prediction time domain length is completely matched with the rolling optimization time domain length of the lower-level predictive control model. The upper-level joint decision model synchronously distributes the power budget, task allocation results, and task type prediction information to the corresponding computing nodes through the cluster internal management network. The distribution timing is within the last lower-level control cycle of each upper-level decision cycle, ensuring that the lower-level predictive control model has obtained complete global decision information before executing the rolling optimization of the next upper-level decision cycle. A distribution fault tolerance rule is set. If the distribution delay exceeds one lower-level control cycle, the node uses the power budget and weight rules of the previous cycle until the new information is synchronized, avoiding control logic confusion. After receiving the allocated power budget, the lower-level predictive control model sets it as a hard upper limit for power constraints throughout the entire rolling optimization time domain. The predicted power consumption of each node within the entire optimization time domain must not exceed this upper limit. Based on the task allocation results, it determines the start time, resource requirements, and execution duration of each task within the future control time domain, and constructs performance constraints accordingly. The weight pre-adjustment based on task type prediction information is consistent with the weight adjustment rules corresponding to the node's local real-time task type, ensuring the continuity of control logic. The weight conflict resolution rules are as follows: the lower-level predictive control model uses the weights corresponding to the task type prediction information pre-issued by the upper layer as the base weights; when the locally identified task type matches the predicted type, the base weights are used; when the locally identified task type differs from the predicted type, the weights corresponding to the local real-time task type are used, and the weight switch is completed immediately. The task type deviation information is uploaded to the upper-level joint decision model along with the current global feedback data to optimize the subsequent task type prediction accuracy. During weight switching, a single control cycle linear transition is used to achieve a smooth switch, avoiding control logic oscillations caused by sudden weight changes.

[0053] This embodiment addresses the technical problems of information asymmetry between upper and lower layers, the inability of lower-level control to adapt to future load changes, and lag in response to sudden load changes by proactively distributing power budget, task allocation results, and task type prediction information to the lower-level predictive control model after the upper-level joint decision-making model completes global decision-making. This allows the lower-level model to set power constraint limits, determine task execution sequences, and pre-adjust optimization function weights in advance. By proactively distributing global decision information, deep collaboration between upper-level global decision-making and lower-level local control is achieved. This enables node local control to adapt to future load changes in advance, rather than responding passively, improving the response speed to sudden load changes. It effectively avoids service performance degradation and service level agreement (SLA) breaches caused by sudden load changes, further improving the utilization rate of cluster power budget and the stability of system operation.

[0054] In one feasible implementation, step S4, feeding back the actual power consumption data and performance data to the upper-level joint decision-making model to update its decision-making strategy, includes: S41: The lower-level predictive control model of each node integrates the actual power consumption data, performance data and the latest parameters and parameter change data of the node power-performance response model into global feedback data. S42: Input the global feedback data into the upper-level joint decision-making model to evaluate the actual effect of the current decision-making strategy; S43: Based on the evaluation results, update the decision strategy parameters of the upper-level joint decision-making model.

[0055] In this embodiment, the integration cycle of global feedback data is completely consistent with the decision cycle of the upper-level joint decision-making model. After each upper-level decision-making cycle ends, the lower-level predictive control model of each node integrates the running data of all lower-level control cycles within this cycle into a standardized global feedback data packet, which is then uploaded to the upper-level joint decision-making model of the cluster management node through the cluster internal management network. The evaluation dimensions of the actual effect of the decision strategy include the overall energy efficiency of the cluster, the compliance status of the cluster's global service quality, the execution deviation rate of the power budget of each node, and the compliance status of the task execution performance of each node. The update of the decision strategy parameters adopts an incremental update method. Based on the global feedback data of all nodes, the deviation between the actual execution effect of the current strategy and the expected optimization target is calculated. The incremental update of the strategy network parameters is completed through the gradient descent algorithm. A full update is triggered only when the deviation between the strategy execution effect and the expected target exceeds a preset threshold (default 10%), so as to avoid the fluctuation of the cluster operation status caused by large-scale strategy updates. The high-frequency update of the node local model and the low-frequency update of the upper-level decision strategy form a two-layer closed-loop feedback mechanism. The two work together to continuously optimize the control effect of the entire system. The parameter change refers to the average of the relative changes in the parameters of the node power-performance response model after the current update and the model parameters at the end of the previous upper-level decision cycle. It is extracted by calculating the absolute value of the relative change rate of the weight parameters at each layer of the model, and taking the average of all parameters as the final parameter change, used to characterize the degree of drift in node hardware characteristics. The upper-level policy linkage rule is as follows: when the parameter change of a node exceeds 10% for two consecutive decision cycles, it is determined that the node's hardware characteristics have drifted significantly. In subsequent integrated decision-making, the upper-level joint decision model lowers the upper limit of the node's power budget utilization rate (e.g., from 100% to 85%), and simultaneously reduces the proportion of latency-sensitive tasks allocated to the node, until the parameter change of the node is less than 5% for two consecutive cycles, at which point normal decision configuration is restored.

[0056] This embodiment constructs a two-layer closed-loop feedback mechanism from the node level to the cluster level, integrating the actual operating data and model parameter change data of the node into global feedback data, which is then uploaded to the upper-layer joint decision-making model to achieve iterative optimization of the global decision-making strategy. This solves the technical problems in existing cluster power management schemes, such as the lack of complete closed-loop feedback, the inability of decision-making strategies to adapt to changes in load characteristics during long-term cluster operation, and hardware aging drift. Through the coordinated cooperation of node-level fine-grained high-frequency updates and cluster-level macro-level low-frequency updates, it ensures both rapid response to fine-grained load fluctuations and continuous optimization of the global decision-making strategy, enabling the entire power management system to maintain optimal operating status for a long time and continuously ensure a balance between energy efficiency and performance throughout the entire life cycle of the cluster.

[0057] like Figure 2 As shown, the present invention also provides an adaptive power management system for server clusters, including: a memory 10, a processor 20, and an adaptive power management program for server clusters stored on the memory 10 and executable on the processor 20, wherein the adaptive power management program for server clusters is configured to implement the steps of the adaptive power management method for server clusters.

[0058] The adaptive power management system for server clusters provided by this invention adopts the adaptive power management method for server clusters in the above embodiments. It can integrate cluster power budget allocation and task scheduling through a two-layer collaborative closed-loop architecture of upper-layer joint decision-making and lower-layer predictive control. Combined with a node-level online updated power-performance response model, it achieves rolling optimization of frequency adjustment sequences, realizing deep collaboration between global cluster energy efficiency optimization and node-specific adaptive control. This effectively improves the overall energy efficiency of the cluster while strictly meeting the differentiated service quality constraints of latency-sensitive and batch processing services. Compared with the prior art, the beneficial effects of the adaptive power management system for server clusters provided by this invention are the same as those of the adaptive power management method for server clusters provided in the above embodiments. Furthermore, other technical features of the adaptive power management system for server clusters are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0059] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An adaptive power management method for server clusters, characterized in that, The method includes: S1: Obtain the running status data of each node in the server cluster, and generate a load prediction sequence for the future control time domain based on the running status data. The running status data includes the CPU utilization, memory utilization, current power consumption and task queue status of each node. S2: Input the operating status data and the load prediction sequence into the upper-level joint decision model, and make an integrated decision on power budget allocation and task scheduling through the upper-level joint decision model to generate the power budget and task scheduling instructions allocated to each node. S3: Based on the allocated power budget, each node generates frequency regulation commands for itself by using the locally deployed lower-level predictive control model and the online-updated node power-performance response model to perform rolling optimization of the frequency regulation sequence. S4: Based on the frequency adjustment command and the task scheduling command, perform task scheduling and power consumption control, collect the actual power consumption data and performance data after execution, and feed the actual power consumption data and performance data back to the upper-level joint decision model to update its decision strategy.

2. The adaptive power management method for server clusters according to claim 1, characterized in that, In step S2, the integrated decision-making process for power budget allocation and task scheduling through the upper-level joint decision-making model includes: S21: The upper-level joint decision-making model models the cluster total power budget allocation and task scheduling as a unified Markov decision process; S22: Based on the Markov decision process, construct a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying service quality constraints; S23: Use the operating status data and the load prediction sequence as the state input of the Markov decision process, solve the Markov decision process, generate the power budget allocation value and task allocation result of each node, and map the task allocation result into the corresponding task scheduling instruction.

3. The adaptive power management method for server clusters according to claim 2, characterized in that, In step S22, constructing a multi-objective reward function that simultaneously includes the objectives of maximizing cluster energy efficiency and satisfying quality of service constraints includes: S221: Obtain the type identifier of each task in the current task queue waiting to be scheduled, wherein the type identifier at least distinguishes between latency-sensitive tasks and batch processing tasks; S222: Based on the type distribution of the task queue, dynamically adjust the weight distribution between the performance optimization term and the power consumption optimization term in the multi-objective reward function; S223: The adjusted multi-objective reward function is used as the solution objective of the Markov decision process.

4. The adaptive power management method for server clusters according to claim 2, characterized in that, In step S23, solving the Markov decision process includes: S231: Define the power budget allocation value and task allocation result of each node as the joint action space of the Markov decision process; S232: Solve the Markov decision process using a policy learning algorithm to obtain the optimal joint action policy; S233: Based on the optimal joint action strategy, output the power budget allocation value and task allocation result of each node simultaneously.

5. The adaptive power management method for server clusters according to claim 1, characterized in that, In step S3, the rolling optimization of the frequency regulation sequence using the locally deployed lower-level predictive control model and the online-updated node power-performance response model includes: S31: Load the node power-performance response model updated in the previous control cycle as the current prediction model; S32: Based on the task type currently running on this node, construct a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term; S33: Using the current prediction model, predict the node power consumption and processing performance under different frequency settings in the future control time domain; S34: Based on the multi-objective rolling optimization function and the prediction results, solve the optimization problem that satisfies the power budget constraint allocated to this node, and generate the optimal frequency adjustment sequence; S35: Use the first frequency value in the optimal frequency adjustment sequence as the frequency adjustment command for this node; S36: Collect the actual operating data after the execution of this control cycle, update the node power-performance response model, and use it as the current prediction model for the next control cycle.

6. The adaptive power management method for server clusters according to claim 5, characterized in that, In step S32, constructing a multi-objective rolling optimization function that includes a power tracking error term, a performance preservation term, and a frequency fluctuation suppression term includes: S321: Identify the type of task currently running on this node, wherein the task type includes at least latency-sensitive tasks and batch processing tasks; S322: Based on the task type of this node, adaptively adjust the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function; S323: The adjusted multi-objective rolling optimization function is used as the objective of the optimization problem.

7. The adaptive power management method for server clusters according to claim 5, characterized in that, In step S36, collecting the actual operating data after the execution of this control cycle and updating the node power-performance response model includes: S361: Extract the real-time power consumption data, operating frequency data and resource utilization data collected in this control cycle to form a model update sample; S362: The parameters of the node power-performance response model are incrementally updated using the model update samples by employing a recursive gradient descent method. S363: Store the updated node power-performance response model to local memory.

8. The adaptive power management method for server clusters according to claim 1, characterized in that, In step S4, feeding back the actual power consumption data and performance data to the upper-level joint decision-making model to update its decision-making strategy includes: S41: The lower-level predictive control model of each node integrates the actual power consumption data, performance data and the latest parameters and parameter change data of the node power-performance response model into global feedback data. S42: Input the global feedback data into the upper-level joint decision-making model to evaluate the actual effect of the current decision-making strategy; S43: Based on the evaluation results, update the decision strategy parameters of the upper-level joint decision-making model.

9. The adaptive power management method for server clusters according to claim 1, characterized in that, Before performing step S3, the method further includes: S301: The upper-level joint decision-making model generates task type prediction information for each node in the future control time domain based on the load prediction sequence; S302: The upper-level joint decision-making model will send the power budget, task allocation results and task type prediction information allocated to each node to the lower-level predictive control model of each node. S303: The lower-level predictive control model uses the power budget allocated to this node as the upper limit of the power constraint for rolling optimization. It determines the execution sequence of future tasks based on the received task allocation results and, in conjunction with the task type prediction information, pre-adjusts the weight distribution between the power tracking error penalty term and the performance preservation penalty term in the multi-objective rolling optimization function.

10. An adaptive power management system for server clusters, characterized in that, The adaptive power management system for server clusters includes: a memory, a processor, and an adaptive power management program for server clusters stored in the memory and executable on the processor, the adaptive power management program for server clusters being configured to implement the steps of the adaptive power management method for server clusters as described in any one of claims 1 to 9.

Images