A method and system for intelligent allocation of computing server resources and energy efficiency optimization

Abstract

This invention relates to the field of cloud computing resource scheduling, and more particularly to a method and system for intelligent allocation and energy efficiency optimization of computing server resources. The method includes: collecting and normalizing the status data of computing cluster nodes; inputting a spatiotemporal graph attention network model to extract topological connectivity and temporal load features; using the Gumbel Copula function to calculate the joint probability distribution to generate burst load prediction results; constructing a virtual penalty queue based on the prediction results and latency constraint parameters and obtaining the current queue length; defining a Lyapunov function to establish an optimization model aimed at minimizing drift and energy consumption penalties and generating a resource scheduling strategy; and accordingly executing virtual machine migration and adjusting processor operating frequency and voltage. This invention effectively avoids the risk of cascading downtime caused by extreme concurrent loads, and reduces the overall physical operating energy consumption of the cluster while ensuring service level agreements and task latency.

Description

Technical Field

[0001] This invention relates to the field of cloud computing resource scheduling, and in particular to a method and system for intelligent allocation and energy efficiency optimization of computing server resources. Background Technology

[0002] With the widespread deployment of cloud computing and microservice architecture, the computing server clusters in large data centers usually need to carry massive and highly dynamic concurrent computing tasks. In order to effectively control the overall operating costs while ensuring the service level agreements of different business flows, the underlying system must collect and analyze the hardware and software status characteristics of each computing node in real time, and then use the global resource scheduler to dynamically map and intelligently allocate the task queues and physical energy consumption budgets in the computing network.

[0003] However, most existing conventional resource scheduling schemes are based on the assumption of stable distribution and passive threshold alarms and delayed expansion. When faced with sudden extreme high-pressure concurrent traffic, it is difficult to predict the tail extreme correlation of multiple underlying nodes congestion at the same time. It is very easy to cause cascading diffusion of node resource exhaustion and partial cluster downtime due to untimely response. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method and system for intelligent allocation of computing server resources and energy efficiency optimization, aiming to improve the response lag problem caused by the lack of tail extreme value correlation prediction in existing scheduling schemes.

[0005] In a first aspect, the present invention provides the following technical solution: a method for intelligent allocation of computing server resources and energy efficiency optimization, comprising: S1. Collect status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and normalize the collected status data. S2. Input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between the computing nodes; S3. Based on the time-series load characteristics, use the Gumbel Copula function to calculate the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window, and generate a burst load prediction result. S4. Based on the burst load prediction results and the delay constraint parameters in the parameters of the task to be processed, construct a virtual penalty queue representing the delay deviation state, and obtain the current queue length of the virtual penalty queue. S5. Define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current moment, and generate a resource scheduling strategy by solving the optimization model. S6. Perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the amount of tasks allocated to each computing node.

[0006] Preferably, in step S1, the step of normalizing the collected state data includes: The hardware timing data and the network bandwidth are timestamped using a time window interpolation algorithm to compensate for the phase deviation of asynchronous acquisition. Based on the anomaly detection model, hardware noise and network transmission spikes in the state data are identified and removed to obtain a cleaned data sequence. The cleaned data sequence is scaled and projected spatially using an extremum mapping function to generate a dimensionless feature matrix that meets the network input standard, which serves as the normalized state data.

[0007] Preferably, in step S2, the step of extracting the topological connections and temporal load characteristics between the computing nodes includes: Using the computing nodes as vertices and network bandwidth and physical links as edges, a heterogeneous topology graph representing the architecture of the computing server cluster is constructed. The heterogeneous topology graph is subjected to neighborhood aggregation using graph convolution kernels to extract the topological connectivity features that characterize resource contention and communication congestion radius; By using a gated recurrent unit network to process node state sequences and by adaptively learning the decay weights of historical states, the time-series load features that capture the load evolution trend are extracted.

[0008] Preferably, in step S3, the step of calculating the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window includes: By using kernel density estimation to fit the temporal load characteristics of each computing node, an edge cumulative distribution function characterizing the resource utilization fluctuation of a single node is constructed. The maximum likelihood estimation method is used in conjunction with samples whose historical concurrent load exceeds the preset threshold to solve for the correlation parameters of the Gumbel Copula function, which characterizes the correlation strength of the tail extrema of multiple nodes. Substituting the edge cumulative distribution function into the parameter-calibrated Gumbel Copula function, the joint probability distribution of multiple computing nodes simultaneously exceeding the preset threshold is solved through numerical integration.

[0009] Preferably, in step S3, the step of generating the burst load prediction result includes: The joint probability distribution is compared and verified with a preset probability threshold to filter out a subset of nodes with concurrency overload risk. By combining graph clustering algorithms to perform connectivity analysis on the subset of overload risk nodes, the cascading diffusion path of the predicted resource depletion is depicted. The overload risk node subset, the cascade diffusion path, and the corresponding occurrence probability are encapsulated into a risk feature vector, which is then output as the burst load prediction result.

[0010] Preferably, in step S4, the step of constructing the virtual penalty queue representing the delay deviation state includes: The delay constraint parameters in the parameters of the task to be processed are analyzed to extract the preset service time limit and maximum default tolerance rate of different priority business flows; For each priority service flow, an independent state machine is instantiated, and a discrete-time queuing update equation driven by the task arrival rate and the actual processing rate is established. The virtual penalty queue, which represents the degree of service level agreement violation, is generated by accumulating the task timeout penalty term within the scheduling time slot using the queuing update equation.

[0011] Preferably, in step S5, the step of establishing an optimization model aimed at minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time includes: The Lyapunov function characterizing the queuing congestion state of the computing server cluster is defined as half the sum of the squares of the current queue lengths of the virtual penalty queue. Calculate the conditional expectation change of the Lyapunov function between adjacent scheduling time slots, obtain the upper bound of the drift term, and perform a product operation between the energy efficiency weight parameter and the energy consumption penalty at the current time. By linearly combining the drift upper bound term with the energy consumption penalty at the current moment after the product operation, the optimization model that transforms long-term stochastic optimization into single-step deterministic extremum solution is constructed.

[0012] Preferably, in step S5, the step of generating a resource scheduling strategy by solving the optimization model includes: By introducing Lagrange multipliers, the resource capacity constraint in the optimization model is relaxed to the objective function, and the original problem is decomposed into multiple subproblems that can be computed in parallel. An alternating direction multiplier algorithm is used to perform iterative communication between the central controller and each computing node, synchronously updating resource allocation variables until the two-sided gradient converges. Based on the converged variable matrix, the task routing and forwarding matrix and the node power budget allocation vector are extracted and integrated to form the resource scheduling strategy that indicates the virtual machine location and hardware power consumption limit.

[0013] Preferably, in step S6, the step of adjusting the operating frequency and voltage of the processors in each of the computing nodes includes: The node power budget allocation vector in the resource scheduling strategy is parsed and mapped to the underlying physical state instruction set of the processor in each computing node; The physical state instruction set is sent to the dynamic adjustment module through the baseboard management controller, and the power supply voltage and operating frequency of the processor are smoothly switched using a phase-locked loop circuit; A temperature monitoring closed loop based on a digital thermal sensor is established. When the core junction temperature reaches a preset core temperature threshold, thermal throttling protection is triggered to execute the task.

[0014] Secondly, the present invention provides the following technical solution: a computing server resource intelligent allocation and energy efficiency optimization system, comprising: The data acquisition and normalization module is used to acquire the status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and the acquired status data is normalized. The spatiotemporal feature extraction module is used to input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between the computing nodes. The burst load prediction module is used to calculate the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window based on the time-series load characteristics and using the Gumbel Copula function, and generate burst load prediction results. The penalty queue construction module is used to construct a virtual penalty queue representing the delay deviation state based on the burst load prediction result and the delay constraint parameter in the parameters of the task to be processed, and to obtain the current queue length of the virtual penalty queue. The optimization scheduling solution module is used to define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time, and generate a resource scheduling strategy by solving the optimization model; The resource and energy efficiency execution module is used to perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the amount of tasks allocated to each computing node.

[0015] The present invention has the following beneficial effects: 1. In this invention, by introducing a spatiotemporal graph attention network and the Gumbel Copula function to extract the spatiotemporal features and tail extreme value correlation between nodes, it is possible to predict the cascading diffusion risk caused by extreme concurrent loads. This changes the lag situation of passive response of traditional scheduling schemes when facing sudden high pressure, thereby effectively avoiding the risk of server cluster resource exhaustion and downtime.

[0016] 2. In this invention, a single-step deterministic solution model based on a virtual penalty queue is constructed using Lyapunov optimization theory. This transforms the complex long-term queuing delay constraint into an extreme value solution problem for each scheduling time slot, providing a strict upper bound guarantee for task response delay and realizing the optimal dynamic trade-off between service level protocol and underlying hardware power consumption.

[0017] 3. In this invention, the upper-layer resource scheduling strategy is accurately mapped to the dynamic voltage and frequency adjustment instructions of the lower-layer processor. Combined with the phase-locked loop circuit and the temperature negative feedback closed-loop control mechanism, the overall physical operating energy consumption of the computing server cluster is reduced while smoothly switching the hardware power consumption state to avoid forced shutdown due to local overheating. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for intelligent allocation of computing server resources and energy efficiency optimization proposed in this invention; Figure 2 This is an architecture diagram of a computing server resource intelligent allocation and energy efficiency optimization system proposed in this invention. Detailed Implementation

[0019] 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 embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 In the first embodiment of the present invention, the present invention provides a method for intelligent allocation of computing server resources and energy efficiency optimization, such as... Figure 1 As shown, it includes: S1. Collect status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and normalize the collected status data. Further, in step S1, the step of normalizing the collected state data includes: using a time window interpolation algorithm to timestamp-align the hardware timing data and network bandwidth to compensate for the phase deviation of asynchronous acquisition; identifying and removing hardware noise and network transmission spikes in the state data based on an anomaly detection model to obtain a cleaned data sequence; and using an extreme value mapping function to perform spatial dimension scaling projection on the cleaned data sequence to generate a dimensionless feature matrix that meets the network input standard, which serves as the normalized state data.

[0021] Specifically, agent programs deployed on each compute node read the underlying hardware interface and operating system kernel status in real time to obtain CPU utilization, memory usage, GPU utilization, and real-time motherboard power consumption as hardware timing data. Network traffic monitoring tools are used to intercept the real-time throughput of the physical network cards entering and leaving the compute nodes as network bandwidth. The number of currently queued compute tasks and their instruction execution lengths are extracted from the cache queue of the cluster's global task scheduler as parameters for tasks awaiting processing.

[0022] Because the sampling clock crystals of the various monitoring agents in the cluster have physical deviations, the directly acquired status data is an asynchronous data sequence that is not aligned on the time axis. At the same time, the ultra-high frequency sampling of physical sensors is prone to interference with glitches and noise. Therefore, the acquired status data must be normalized.

[0023] A time window interpolation algorithm is used to align hardware timing data with network bandwidth timestamps to compensate for phase deviations in asynchronous acquisition. A globally unified discrete time reference sequence is established. For a target timestamp in the global time reference sequence, two adjacent sampling times are found in the original timestamp sequence. The aligned state data value is calculated using a linear time window interpolation algorithm, and the calculation formula is as follows: ; In the formula, This represents the target timestamp that needs to be aligned under a globally unified benchmark. This represents the underlying hardware status or network throughput value aligned to the target timestamp. and These represent the two actual physical sampling moments located before and after the target timestamp in the original asynchronous sampling sequence. and These represent the times at which physical sampling occurs. and The collected actual hardware operating status values ​​are then calculated point-by-point using this formula, mapping the asynchronously collected raw data stream to a fully synchronized global time scale.

[0024] An anomaly detection model is used to identify and remove hardware noise and network transmission spikes from state data, resulting in a cleaned data sequence. During extremely high-concurrency server operation, extreme data spikes may occur due to transient voltage fluctuations or network packet loss and retransmission. An anomaly detection algorithm based on absolute median difference is used to evaluate the timestamp-aligned data. A time sliding window is set around the target timestamp, and the data corresponding to all timestamps within the window are extracted to form a local sequence set. The median of this local sequence set is then calculated. The anomaly detection scoring formula is defined as follows: ; In the formula, This represents the current hardware status data or network bandwidth data. Abnormal score values, This represents the median of the corresponding state data within the local sequence set. This represents the median, which is the absolute deviation of all state data within a local sequence set from the median. It is also the anomaly score for a given state data point. When the value exceeds a preset threshold, it is identified as physical hardware noise or network transmission spikes and removed. At the same time, normal state data from the preceding time step is used to fill the data gap corresponding to the timestamp, and finally a smooth and clean data sequence without sudden spikes is output.

[0025] An extremum mapping function is used to scale and project the spatial dimensions of the cleaned data sequence, generating a dimensionless feature matrix that meets the network input criteria, which serves as the normalized state data. Since CPU utilization is measured as a percentage, memory usage as bytes, and motherboard power consumption as watts, state features of different dimensions not only have different physical dimensions but also differ in magnitude by tens of thousands of times. Directly inputting these features into the subsequent graph network model would lead to a severe skew in the backpropagation gradient update. An extremum mapping algorithm is used to scale each type of state feature in the cleaned data sequence proportionally to a preset continuous real number domain of zeros and ones. The extremum scaling projection formula is defined as follows: ; In the formula, This represents the dimensionless characteristic value generated after normalization. This represents the current true physical quantity value in a certain dimension of the cleaned data sequence. and These represent the physical minimum and physical maximum values ​​recorded for the state data of this dimension within a complete historical statistical period, respectively. After independently scaling all monitoring dimensions across all computing nodes in the cluster, the dimensionless values ​​generated for each dimension are concatenated according to the cluster spatial topology to generate a dimensionless feature matrix that can be directly used by the subsequent network model.

[0026] This step calibrated the timeline misalignment caused by asynchronous monitoring of multiple nodes in the cluster, eliminated abnormal spikes caused by sensor acquisition and network fluctuations, and smoothed out the differences in physical dimensions of various hardware operating indicators, thus building a unified and stable underlying data foundation for subsequent spatiotemporal prediction models.

[0027] S2. Input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between computing nodes; Further, in step S2, the steps of extracting the topological connections and temporal load features between computing nodes include: constructing a heterogeneous topology graph representing the architecture of the computing server cluster with computing nodes as vertices and network bandwidth and physical links as edges; performing neighborhood aggregation on the heterogeneous topology graph using graph convolution kernels to extract topological connection features representing resource competition and communication congestion radius; and using gated cyclic unit networks to process node state sequences and extracting temporal load features that capture load evolution trends by adaptively learning the decay weights of historical states.

[0028] Specifically, a heterogeneous topology graph representing the architecture of a computing server cluster is constructed, using computing nodes as vertices and network bandwidth and physical links as edges. Within the server cluster, an undirected graph is defined to map the underlying physical network layers. The set of vertices represents all independent computing nodes within the cluster, and the set of edges represents the physical communication links between nodes. An adjacency matrix is ​​constructed using normalized real-time network bandwidth as edge weights. The formula for calculating the elements of the adjacency matrix is ​​defined as follows: ; In the formula, The adjacency matrix of the network architecture represents the first... The computing node and the first Spatial association weights between computing nodes This represents the normalized, real-time available network throughput bandwidth data between nodes, passed from the preceding steps. By establishing this heterogeneous topology graph, the hardware layout of the physical data center is transformed into a graph-structured mathematical expression that can be read by the algorithm.

[0029] A graph convolution kernel is used to perform neighborhood aggregation on a heterogeneous topology graph, extracting topological connectivity features that characterize resource contention and communication congestion radii. Normalized hardware time-series data of all computing nodes at the current moment and parameters of the task to be processed are used as the initial feature inputs for each node. A spectral domain graph convolution algorithm is used to perform feature propagation and aggregation in the spatial dimension, enabling each node to perceive the resource occupancy of its connected neighbors. The graph convolution aggregation propagation formula is defined as follows: ; In the formula, Indicates the time generated after spatial neighborhood aggregation. The topology connectivity feature matrix maps the congestion diffusion intensity of multiple servers competing for network resources within a local area. This represents a physical link adjacency matrix with added self-loops, allowing nodes to retain their original computational state. This represents the degree matrix corresponding to the adjacency matrix, used to measure and average the node features. Indicates time The input is a normalized dimensionless feature matrix containing processor utilization and memory usage characteristics. This represents the trainable spatial feature weight parameter matrix of a graph convolutional network layer. This represents a non-linear activation function.

[0030] A gated recurrent unit (GRU) network is used to process node state sequences. By adaptively learning the decay weights of historical states, temporal load features capturing load evolution trends are extracted. Topological connectivity features containing spatial interaction information are sequentially input into the network according to time steps. The network uses a reset gate mechanism to determine the proportion of high-load server states accumulated from historical moments to be retained at the current moment. The temporal evolution state update formula is defined as follows: ; ; In the formula, The output vector of the reset gate represents the time decay weights adaptively learned for the server's historical overload or idle states. This represents the topological connectivity feature matrix input at the current time. This indicates the historical timing load characteristics of the output from the previous scheduling slot. This represents the update gate vector, used to control the fusion ratio of historical computing load and the currently newly introduced load. and as well as and These are all trainable weight matrices used in the network for extracting time-series information. and This is the bias coefficient. and This is the activation function. This represents the Hadamard product operation. This represents the time-series load characteristics of the network's final output, which include historical trends and the current state.

[0031] This step successfully maps the independent and fragmented physical operating states of each node into high-dimensional comprehensive variable characteristics that combine the spatial interference network of the cluster with the temporal evolution law of computing load.

[0032] S3. Based on the time-series load characteristics, use the Gumbel Copula function to calculate the joint probability distribution of the load of different computing nodes exceeding the preset threshold within a preset time window, and generate the burst load prediction result. Further, in step S3, the step of calculating the joint probability distribution of different computing nodes exceeding a preset threshold within a preset time window includes: fitting the temporal load characteristics of each computing node using kernel density estimation to construct an edge cumulative distribution function characterizing the fluctuation of resource utilization of a single node; using the maximum likelihood estimation method combined with historical samples of concurrent loads exceeding the preset threshold to solve for the correlation parameters of the Gumbel Copula function characterizing the correlation strength of the tail extreme values ​​of multiple nodes; substituting the edge cumulative distribution function into the parameter-calibrated Gumbel Copula function, and solving for the joint probability distribution of multiple computing nodes simultaneously exceeding the preset threshold through numerical integration.

[0033] Further, in step S3, the step of generating the burst load prediction result includes: comparing and verifying the joint probability distribution with the preset probability threshold, and screening out the overload risk node subset with concurrent overload risk; performing connectivity analysis on the overload risk node subset using a graph clustering algorithm to depict the cascading diffusion path that is expected to cause resource depletion; and encapsulating the overload risk node subset, the cascading diffusion path, and the corresponding occurrence probability into a risk feature vector, which is then output as the burst load prediction result.

[0034] Specifically, firstly, kernel density estimation is used to fit the temporal load characteristics of each computing node, constructing an edge cumulative distribution function characterizing the fluctuation of resource utilization of a single node. The combined CPU and memory load rates from the temporal feature matrix output in the previous stage are extracted as analysis variables. The edge cumulative probability evaluation formula is defined as follows: ; In the formula, Indicates the first The overall resource load rate of each computing node is less than or equal to a given assessment level. The marginal cumulative distribution probability. This represents the total number of time-series load feature samples extracted from the historical time series. The bandwidth parameter represents the degree of smoothness of the probability density curve. This represents the standard Gaussian kernel function. Represents a node In the The actual overall load rate value recorded in each historical sampling time slot. This represents a continuously integral variable. Using this formula can smooth out statistical abrupt changes caused by discrete sampling from a single monitoring probe, mapping discrete computing load data into a smooth and continuous probability distribution.

[0035] By employing maximum likelihood estimation combined with samples of historical concurrent loads exceeding a preset threshold, the correlation parameters of the Gumbel Copula function, characterizing the strength of the tail extreme value correlation among multiple nodes, are calculated. The server cluster's physical monitoring logs are retrieved, and extreme high-pressure sample pairs of any two computing nodes with physical communication links simultaneously experiencing resource request queuing congestion are extracted to form an extreme value training set. The two-dimensional extreme value correlation mapping formula is constructed as follows: ; In the formula, It represents the joint distribution probability of two computing nodes experiencing a specific extreme load state under the same preset time dimension. and These represent the independent probabilities of the load levels of the two nodes, calculated from the aforementioned edge cumulative distribution function. This represents the core correlation parameter specifically designed to capture the right-tail correlation of multi-node extreme value distributions. The extracted historical high-pressure samples are substituted into the maximum likelihood estimation objective function for gradient optimization to find the solution that maximizes the probability of historical extreme congestion events. The numerical value is used to quantify the physical synchronization transmission strength when the local load in the computing network spikes.

[0036] Substituting the edge cumulative distribution function into the parameter-calibrated Gumbel Copula function, the joint probability distribution of multiple computing nodes simultaneously exceeding a preset threshold is solved through numerical integration. The highest alert load threshold for ensuring the safe operation of the physical server is set as follows: The formula for predicting the probability of node concurrency overload is as follows: ; In the formula, Represents a computing node With computing nodes The computing load will simultaneously exceed the safety threshold within the future prediction time window. The absolute probability of concurrent overload. and These represent the unilateral limit cumulative probability that the loads of the two nodes are below the safety warning line.

[0037] The specific implementation process for generating burst load prediction results includes dimensionality reduction and restructuring of the probability matrix calculated above. This involves combining the joint probability distribution of any adjacent nodes globally. The probability is compared step-by-step with the preset probability threshold set during system initialization. When the joint probability of a computing node and its communicating neighbor nodes experiencing concurrent overload exceeds the preset probability threshold, it is removed from the global graph nodes and stored in a subset of overload risk nodes representing dangerous states. A graph connectivity analysis is performed on this subset of overload risk nodes using a graph clustering algorithm. The spatial topology subgraph corresponding to this risk subset in the physical cluster is retrieved, and the longest physical node connection branch directly connected by high-speed switches with a joint overload probability continuously exceeding the threshold is found. The topology of this branch is depicted as a cascading diffusion path that predicts resource depletion. Finally, the absolute occurrence probabilities of the selected overload risk node subset, cascading diffusion path, and corresponding node indices are sequentially concatenated and encapsulated into a one-dimensional risk feature vector, which is output as the burst load prediction result to the subsequent reinforcement learning queue control environment.

[0038] This step accurately captures the risk of cluster overload of nodes caused by sudden surges in traffic under the microservice architecture from a deep statistical perspective, providing a solid digital basis for the system to shift from passive response to proactive defensive resource scheduling.

[0039] S4. Based on the burst load prediction results and the delay constraint parameters in the parameters of the task to be processed, construct a virtual penalty queue that represents the delay deviation state, and obtain the current queue length of the virtual penalty queue. Further, in step S4, the step of constructing a virtual penalty queue representing the delay deviation state includes: parsing the delay constraint parameters in the parameters of the task to be processed, extracting the preset service time limit and maximum default tolerance rate of different priority business flows; instantiating an independent state machine for each priority business flow, and establishing a discrete-time queuing update equation driven by the task arrival rate and the actual processing rate; using the queuing update equation to accumulate the task timeout penalty term in the scheduling time slot, and generating a virtual penalty queue representing the degree of service level agreement violation.

[0040] Specifically, the latency constraint parameters in the task parameters are parsed to extract the preset service time limits and maximum default tolerance rates for different priority business flows. When the server cluster receives a computation request forwarded by the front-end gateway, the business metadata in the request data packet header file is read. Different priority business flows are divided according to the business type tag; for example, financial transaction business flows are classified as high priority, and offline data cleaning business flows are classified as low priority. The preset service time limits and the allowed maximum default tolerance rates agreed upon for each priority business flow are extracted from the metadata. The preset service time limit represents the maximum physical time delay allowed from task submission to return a result, and the maximum default tolerance rate represents the upper limit of the percentage of tasks that are allowed to time out within a statistical period.

[0041] Independent state machines are instantiated for each priority business flow, and a discrete-time queuing update equation driven by task arrival rate and actual processing rate is established. An independent data structure is allocated in the memory space of the central scheduler to instantiate the state machine for each priority business flow. The number of newly arriving computing tasks in the current scheduling time slot is collected in real time through network port monitoring as task arrival rate data, and converted into actual processing rate data based on the processor instruction throughput capacity of the underlying computing nodes at the current operating frequency. The discrete-time queuing update equation is defined as follows: ; In the formula, Indicates priority is marked as The service flow in the next scheduling slot The physical queuing workload. This indicates that the service flow is in the current scheduling time slot. The initial physical queuing task volume. This indicates the task arrival rate data corresponding to this service flow within the current scheduling time slot. This represents the actual processing rate data allocated by the system to the service flow of this priority. This equation can accurately track the physical accumulation evolution of tasks of each priority in the compute node cache.

[0042] By accumulating task timeout penalties within scheduling slots using a queuing update equation, a virtual penalty queue representing the degree of service level agreement (SSPA) violation is generated. Combined with the burst load prediction results from previous steps, when the system anticipates that a specific node is about to experience resource exhaustion congestion, the backlog of tasks in the physical queue will cause a surge in response latency. Mapping the task timeouts caused by physical backlog to violation penalties, the dynamic update formula for the virtual penalty queue is constructed as follows: ; In the formula, This indicates the current queue length of the updated virtual penalty queue. This indicates the initial accumulated volumetric length of the virtual penalty queue in the current time slot. This represents the penalty value for the task timeout, calculated according to the queuing update equation, when the physical queuing delay of the service flow exceeds the preset service time limit within the current scheduling time slot. This represents the maximum default tolerance constant extracted from business metadata.

[0043] This step abstracts the complex task delay constraints into a rigorous dynamic queuing mathematical model, and intuitively quantifies the degree of deviation of the system's current default risk by using the length of the virtual penalty queue.

[0044] S5. Define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time, and generate a resource scheduling strategy by solving the optimization model. Further, in step S5, the step of establishing an optimization model with the objective of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current moment includes: defining a Lyapunov function representing the queuing congestion state of the computing server cluster using half the sum of the squares of the current queue length of the virtual penalty queue; calculating the conditional expectation change of the Lyapunov function between adjacent scheduling time slots, obtaining the upper bound term of the drift, and introducing energy efficiency weight parameters to perform a product operation with the energy consumption penalty at the current moment; and linearly combining the upper bound term of the drift with the energy consumption penalty at the current moment after the product operation to construct an optimization model that transforms long-term stochastic optimization into a single-step deterministic extreme value solution.

[0045] Further, in step S5, the step of generating a resource scheduling strategy by solving the optimization model includes: introducing Lagrange multipliers to relax the resource capacity constraints in the optimization model to the objective function, decomposing the original problem into multiple sub-problems that can be computed in parallel; using an alternating direction multiplier algorithm to perform iterative communication between the central controller and each computing node, synchronously updating the resource allocation variables until the two-sided gradient converges; and extracting the task routing forwarding matrix and the node power budget allocation vector based on the converged variable matrix, integrating them to form a resource scheduling strategy that indicates the virtual machine location and hardware energy consumption limit.

[0046] Specifically, the current length of the virtual penalty queue for each priority business flow, representing the degree of service level agreement violation generated in the previous steps, is extracted. A Lyapunov function representing the queuing congestion state of the computing server cluster is defined as half the sum of the squares of the current queue lengths of the virtual penalty queues. The formula for the physical congestion scalar of the Lyapunov function is constructed as follows: ; In the formula, Indicates the current scheduling time slot The system scalar state representing the severity of global task congestion in the computing server cluster. Indicates priority is marked as Business flows in scheduling time slots The current backlog length of the virtual penalty queue. This indicates the total number of service flow types carried by the server cluster.

[0047] An optimization model is established with the objective of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time step. The conditional expectation change of the Lyapunov function between adjacent scheduling slots is calculated, the upper bound of the drift is obtained, and an energy efficiency weight parameter is introduced and multiplied with the energy consumption penalty at the current time step. The upper bound of the drift and the energy consumption penalty at the current time step after the product are linearly combined to construct an optimization model that transforms long-term stochastic optimization into a single-step deterministic extremum solution. The objective function for the single-step extremum solution is as follows: ; In the formula, This represents a constant upper bound for drift, determined by the system's physical processing capacity, used to limit the rate of divergence in physical queue length. This indicates the operation of obtaining mathematical expectation. Indicates the service flow within the scheduling time slot The timeout penalty value for the task. This represents the maximum tolerance rate constant for default. This represents the set energy efficiency weight parameter, used to measure the strength of the system's subjective willingness to save power consumption of the underlying server. This represents the total real-time physical energy consumption of each computing node in the computing cluster during the current scheduling time slot, as it processes all assigned tasks.

[0048] Resource scheduling strategies are generated by solving an optimization model. Lagrange multipliers are introduced to relax the resource capacity constraints in the optimization model to the objective function, decomposing the original problem into multiple parallelizable subproblems. CPU core utilization thresholds and total physical memory capacity are set as hardware resource capacity constraints for each computing node. An augmented Lagrange function is used to transform the original global scheduling model with hard hardware constraints into an unconstrained form. The objective relaxation function is defined as follows: ; In the formula, This represents the matrix of server node computation task allocation ratios that needs to be solved. This represents the corresponding server motherboard operating power allocation variable matrix. and These represent the task latency cost function and the physical energy consumption cost function, respectively, from the perspective of a single computing node. Let represent the dual variable vector of the Lagrange multipliers. and These represent the task routing allocation constraint matrix and the motherboard power consumption constraint matrix, respectively. This represents the physical boundary vector of the actual available processor and memory resources of a node. This represents the preset penalty parameter used to control the approximation speed of the algorithm.

[0049] An alternating direction multiplier algorithm is used for iterative communication between the central controller and each computing node, synchronously updating resource allocation variables until the bidirectional gradient converges. The central scheduler distributes global multiplier variables to each computing node. Each node solves for the optimal values ​​of its own task allocation variables and power allocation variables in parallel based on its local underlying physical state, and sends the updated calculation results back to the central scheduler. The central scheduler synchronously updates the global communication multiplier variables based on the aggregated data from each node, repeatedly executing the above bidirectional network communication process until the L2 norm of the difference between variables in two adjacent iterations is less than a preset convergence threshold.

[0050] Based on the converged variable matrix, the task routing and forwarding matrix and the node power budget allocation vector are extracted and integrated to form a resource scheduling strategy that indicates the virtual machine location and hardware power consumption limits. The converged task allocation variable matrix is ​​then extracted. The non-zero elements are mapped to the task routing and forwarding rules from the front-end gateway to each back-end server node. The converged power allocation variable matrix is ​​extracted. This generates a node power budget allocation vector that guides fine-tuning of the motherboard's underlying structure. The two are then combined to generate physical execution strategy instructions that are sent to the cluster control engine.

[0051] This step transforms the long-term congestion control requirement into a fast computation problem independent of each time slot, effectively reducing the overall physical energy consumption of the system while strictly limiting the task processing latency boundary.

[0052] S6. Perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the workload allocated to each computing node.

[0053] Furthermore, in step S6, the steps of adjusting the operating frequency and voltage of the processors in each computing node include: parsing the node power budget allocation vector in the resource scheduling strategy and mapping it to the underlying physical state instruction set of the processor in each computing node; sending the physical state instruction set to the dynamic adjustment module through the baseboard management controller, and using a phase-locked loop circuit to smoothly switch the processor's power supply voltage and operating frequency; establishing a temperature monitoring closed loop based on a digital thermal sensor, and triggering thermal throttling protection when the core junction temperature reaches a preset core temperature threshold to execute the workload.

[0054] Specifically, the node power budget allocation vector in the resource scheduling strategy is parsed and mapped to the underlying physical state instruction set of the processors in each computing node. The target power upper limit value is extracted from the node power budget allocation vector output by the previous steps. Based on the dynamic power consumption physical characteristics of digital integrated circuits, the physical mapping equation between processor power, operating frequency, and supply voltage is constructed as follows: ; In the formula, This represents the target power ceiling value issued to a specific computing node by the node power budget allocation vector. This represents the processor's static leakage current and physical power consumption. The average gate toggling activity factor represents the amount of work performed by the processor in executing the current task. This represents the equivalent load capacitance of the processor's internal circuitry. This indicates the power supply voltage to the processor core. This indicates the operating frequency of the processor core. Based on this physical mapping equation, a greedy search algorithm is used to select the state level with the highest operating frequency and full-load power consumption (less than the target power limit) from multiple discrete performance state ladders preset by the advanced configuration and power interface standard. Its hexadecimal hardware control code is then used as the underlying physical state instruction set.

[0055] The physical state instruction set is sent from the baseboard management controller to the dynamic adjustment module, which uses a phase-locked loop (PLL) circuit to smoothly switch the processor's power supply voltage and operating frequency. The central scheduler sends intelligent platform management interface messages containing the physical state instruction set to the baseboard management controller on each computing node's motherboard via the out-of-band management network. After parsing the network message, the baseboard management controller writes hexadecimal hardware control code into the processor's internal machine-specific registers via the integrated circuit's built-in bus. The processor's built-in power management unit reads the register values ​​and sends a serial voltage flag instruction to the voltage regulation module on the motherboard to adjust the core power supply voltage, while simultaneously sending frequency division and multiplication instructions to the motherboard clock generator. The PLL circuit gradually adjusts the voltage-controlled oscillator's output clock frequency on a microsecond timescale, ensuring a smooth transition of the processor's power supply voltage and operating frequency to the target state, preventing sudden voltage drops from causing direct system crashes.

[0056] A closed-loop temperature monitoring system based on digital thermal sensors is established. When the core junction temperature reaches a preset core temperature threshold, thermal throttling protection is triggered to execute the task. Data from the digital thermal sensor array distributed across the processor's internal die is read in real-time, and the highest junction temperature of all physical cores is calculated as the current physical temperature state. A preset core temperature threshold is set to ensure the lifespan of the silicon-based chip. The temperature negative feedback control equation based on the proportional-integral-differential algorithm is constructed as follows: ; In the formula, This indicates the relationship between the preset core temperature threshold and the current physical temperature state during the current control cycle. The error value, This represents the amount of processor frequency compensation required to forcibly reduce the frequency in order to suppress a continuous rise in junction temperature. This represents the proportional gain coefficient that controls the current junction temperature error. This represents the integral gain coefficient used to eliminate steady-state junction temperature error. The differential gain coefficient represents the predicted junction temperature trend. This represents a continuously integral variable. When the core junction temperature reaches a preset core temperature threshold, the underlying thermal throttling protection mechanism is activated. Using this temperature negative feedback control equation, the frequency reduction magnitude is dynamically calculated based on the degree of junction temperature exceeding the limit, forcing the motherboard clock generator to reduce the clock frequency, so that the computing node can stably execute the allocated workload within the safe heat dissipation envelope.

[0057] This step achieves a precise mapping of computing power scheduling strategies to underlying physical hardware control, significantly reducing the actual physical energy consumption of computing nodes while ensuring continuous task execution without overheating and system crashes.

[0058] Example 2 Conventional resource scheduling schemes mostly rely on the assumption of stable distribution, employing passive threshold alarms and delayed capacity expansion. When faced with sudden, extreme concurrent traffic surges, they struggle to predict the tail-end correlation of congestion across multiple underlying nodes, easily leading to cascading resource exhaustion and partial cluster outages due to delayed responses. To address these issues, this invention provides an intelligent resource allocation and energy efficiency optimization system for computing servers, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows: The data acquisition and normalization module is used to acquire the status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and the acquired status data is normalized. The spatiotemporal feature extraction module is used to input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between the computing nodes. The burst load prediction module is used to calculate the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window based on the time-series load characteristics and using the Gumbel Copula function, and generate burst load prediction results. The penalty queue construction module is used to construct a virtual penalty queue representing the delay deviation state based on the burst load prediction result and the delay constraint parameter in the parameters of the task to be processed, and to obtain the current queue length of the virtual penalty queue. The optimization scheduling solution module is used to define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time, and generate a resource scheduling strategy by solving the optimization model; The resource and energy efficiency execution module is used to perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the amount of tasks allocated to each computing node.

[0059] Specifically, the data acquisition and normalization module first uses a time window interpolation algorithm to timestamp-align hardware time-series data and network bandwidth to compensate for asynchronous deviations. After removing glitches using an anomaly detection model, it generates a dimensionless feature matrix using an extreme value mapping function. Subsequently, the spatiotemporal feature extraction module constructs a heterogeneous topology graph using computing nodes and physical links. Combining graph convolution kernels and gated recurrent unit networks, it adaptively extracts topological connectivity features representing resource competition and time-series load features capturing load evolution trends. Next, the burst load prediction module uses kernel density estimation to fit the marginal cumulative distribution function and solves the Gumbel equation using historical samples. The Copula function's correlation parameters are used to calculate the joint probability distribution of multiple nodes simultaneously exceeding a preset threshold. Combined with a graph clustering algorithm, this outputs the estimated diffusion path leading to cascading resource exhaustion. Simultaneously, the penalty queue construction module analyzes the preset service time limits and default tolerance rates of each priority business flow, establishing a discrete-time queuing update equation to accumulate timeout penalty terms, thereby dynamically obtaining the virtual penalty queue length reflecting the degree of service level agreement violation. Based on this, the optimization scheduling solution module defines the system Lyapunov function as half the sum of the squares of the current queue length, introducing Lagrange multipliers and alternating direction multiplication. The sub-algorithm iteratively communicates between the central controller and the nodes to solve a single-step extreme model that minimizes drift plus energy consumption penalty, and integrates to generate a resource scheduling strategy that includes task routing and power budget. Finally, the resource and energy efficiency execution module parses the strategy and sends instructions to the dynamic adjustment module through the baseboard management controller. It uses a phase-locked loop circuit to smoothly switch the processor's power supply voltage and operating frequency, and establishes a monitoring closed loop with a digital thermal sensor. When the junction temperature reaches a preset threshold, it triggers thermal throttling protection, thereby achieving the ultimate voltage reduction of the underlying physical energy consumption while avoiding the risk of concurrent downtime and strictly ensuring task delay.

[0060] The following is a description with reference to specific embodiments: The application scenario is based on the core data center of a major e-commerce platform's financial payment system during the Double Eleven shopping festival. This computing server cluster contains 1,000 physical computing nodes, carrying two types of core business flows: high-priority and low-priority services. The high-priority service is real-time transaction payment, with a latency constraint of less than 50 milliseconds and a maximum default tolerance rate of 0.1%. The low-priority service is offline user behavior log analysis, with a latency constraint of less than 2,000 milliseconds.

[0061] In the lead-up to the peak traffic flow, the system of this invention performs the following operations: The system collects data on the CPU, memory, motherboard power consumption, and network throughput of each node at 100-millisecond intervals, aligning timestamps and removing transmission spikes using a time window interpolation algorithm. The cleaned data is then input into a spatiotemporal graph attention network, which identifies nodes 1 to 10 on rack A as exhibiting strong spatial resource contention due to sharing the same high-speed switch. Subsequently, using the Gumbel Copula function combined with historical peak sales data, calculations reveal that the joint probability of nodes 1 to 5 simultaneously exceeding the 90% safe load threshold within the next 3 minutes is as high as 87%, which could easily trigger network paralysis within the same rack. The system immediately marks this as an overload risk subset.

[0062] At this point, anticipating the impending congestion, the length of the virtual penalty queue for high-priority payment transactions began to rise sharply. The central scheduler quickly defined the Lyapunov function for this queue and established an optimization model with drift and energy consumption penalties. It then solved the model by communicating with each node using the alternating direction multiplier algorithm, generating the optimal strategy to avoid cascading failures. According to this strategy, one minute before the peak arrived, the system hot-migrated the low-priority log analysis virtual machines on nodes 1 to 5 to the idle rack B. Simultaneously, the baseboard management controller used a phase-locked loop circuit to increase the operating frequency of the processors on nodes 1 to 5 to the maximum level to fully handle payment tasks. Furthermore, when the core junction temperature reached 85 degrees Celsius, it triggered proportional-integral-derivative temperature negative feedback fine-tuning, ensuring that the chip did not overheat and shut down under full load.

[0063] The following is a comparison of the core cluster performance metrics during this major sales period using the traditional threshold-based reactive scheduling method and the intelligent allocation and optimization method of this invention: Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent allocation of computing power server resources and energy efficiency optimization, characterized in that, include: S1. Collect status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and normalize the collected status data. S2. Input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between the computing nodes; S3. Based on the time-series load characteristics, use the Gumbel Copula function to calculate the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window, and generate a burst load prediction result. S4. Based on the burst load prediction results and the delay constraint parameters in the parameters of the task to be processed, construct a virtual penalty queue representing the delay deviation state, and obtain the current queue length of the virtual penalty queue. S5. Define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current moment, and generate a resource scheduling strategy by solving the optimization model; S6. Perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the amount of tasks allocated to each computing node. 2.The method of claim 1, wherein, Step S1, the step of normalizing the collected state data, includes: The hardware timing data and the network bandwidth are timestamped using a time window interpolation algorithm to compensate for the phase deviation of asynchronous acquisition. Based on the anomaly detection model, hardware noise and network transmission spikes in the state data are identified and removed to obtain a cleaned data sequence. The cleaned data sequence is scaled and projected spatially using an extremum mapping function to generate a dimensionless feature matrix that meets the network input standard, which serves as the normalized state data.

3. The method of claim 1, wherein, Step S2, the step of extracting the topological connections and temporal load characteristics between the computing nodes, includes: Using the computing nodes as vertices and network bandwidth and physical links as edges, a heterogeneous topology graph representing the architecture of the computing server cluster is constructed. The heterogeneous topology graph is subjected to neighborhood aggregation using graph convolution kernels to extract the topological connectivity features that characterize resource contention and communication congestion radius; By using a gated recurrent unit network to process node state sequences and by adaptively learning the decay weights of historical states, the time-series load features that capture the load evolution trend are extracted.

4. The computing power server resource intelligent allocation and energy efficiency optimization method of claim 1, wherein, Step S3, the step of calculating the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window, includes: By using kernel density estimation to fit the temporal load characteristics of each computing node, an edge cumulative distribution function characterizing the resource utilization fluctuation of a single node is constructed. The maximum likelihood estimation method is used in conjunction with samples whose historical concurrent load exceeds the preset threshold to solve for the correlation parameters of the Gumbel Copula function, which characterizes the correlation strength of the tail extrema of multiple nodes. Substituting the edge cumulative distribution function into the parameter-calibrated Gumbel Copula function, the joint probability distribution of multiple computing nodes simultaneously exceeding the preset threshold is solved through numerical integration.

5. The computing power server resource intelligent allocation and energy efficiency optimization method of claim 1, wherein, Step S3, the step of generating the burst load prediction result, includes: The joint probability distribution is compared and verified with a preset probability threshold to filter out a subset of nodes with concurrency overload risk. By combining graph clustering algorithms to perform connectivity analysis on the subset of overload risk nodes, the cascading diffusion path of the predicted resource depletion is depicted. The overload risk node subset, the cascade diffusion path, and the corresponding occurrence probability are encapsulated into a risk feature vector, which is then output as the burst load prediction result.

6. The computing power server resource intelligent allocation and energy efficiency optimization method of claim 1, wherein, Step S4, the step of constructing the virtual penalty queue representing the delay deviation state, includes: The delay constraint parameters in the parameters of the task to be processed are analyzed to extract the preset service time limit and maximum default tolerance rate of different priority business flows; For each priority service flow, an independent state machine is instantiated, and a discrete-time queuing update equation driven by the task arrival rate and the actual processing rate is established. The virtual penalty queue, which represents the degree of service level agreement violation, is generated by accumulating the task timeout penalty term within the scheduling time slot using the queuing update equation.

7. The computing power server resource intelligent allocation and energy efficiency optimization method of claim 1, wherein, Step S5, the step of establishing an optimization model aimed at minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time, includes: The Lyapunov function characterizing the queuing congestion state of the computing server cluster is defined as half the sum of the squares of the current queue lengths of the virtual penalty queue. Calculate the conditional expectation change of the Lyapunov function between adjacent scheduling time slots, obtain the upper bound of the drift term, and perform a product operation between the energy efficiency weight parameter and the energy consumption penalty at the current time. By linearly combining the drift upper bound term with the energy consumption penalty at the current moment after the product operation, the optimization model that transforms long-term stochastic optimization into single-step deterministic extremum solution is constructed.

8. The intelligent resource allocation and energy efficiency optimization method for computing servers according to claim 1, characterized in that, Step S5, the step of generating a resource scheduling strategy by solving the optimization model, includes: By introducing Lagrange multipliers, the resource capacity constraint in the optimization model is relaxed to the objective function, and the original problem is decomposed into multiple subproblems that can be computed in parallel. An alternating direction multiplier algorithm is used to perform iterative communication between the central controller and each computing node, synchronously updating resource allocation variables until the two-sided gradient converges. Based on the converged variable matrix, the task routing and forwarding matrix and the node power budget allocation vector are extracted and integrated to form the resource scheduling strategy that indicates the virtual machine location and hardware power consumption limit.

9. The intelligent resource allocation and energy efficiency optimization method for computing servers according to claim 1, characterized in that, Step S6, the step of adjusting the operating frequency and voltage of the processors in each computing node, includes: The node power budget allocation vector in the resource scheduling strategy is parsed and mapped to the underlying physical state instruction set of the processor in each computing node; The physical state instruction set is sent to the dynamic adjustment module through the baseboard management controller, and the power supply voltage and operating frequency of the processor are smoothly switched using a phase-locked loop circuit; A temperature monitoring closed loop based on a digital thermal sensor is established. When the core junction temperature reaches a preset core temperature threshold, thermal throttling protection is triggered to execute the task.

10. A system for intelligent allocation of computing server resources and optimization of energy efficiency, characterized in that, A method for intelligent allocation of computing server resources and energy efficiency optimization as described in any one of claims 1-9, the system comprising: The data acquisition and normalization module is used to acquire the status data of each computing node in the computing server cluster. The status data includes hardware timing data, network bandwidth and parameters of the task to be processed, and the acquired status data is normalized. The spatiotemporal feature extraction module is used to input the normalized state data into the spatiotemporal graph attention network model to extract the topological connections and temporal load features between the computing nodes. The burst load prediction module is used to calculate the joint probability distribution of the load of different computing nodes exceeding a preset threshold within a preset time window based on the time-series load characteristics and using the Gumbel Copula function, and generate burst load prediction results. The penalty queue construction module is used to construct a virtual penalty queue representing the delay deviation state based on the burst load prediction result and the delay constraint parameter in the parameters of the task to be processed, and to obtain the current queue length of the virtual penalty queue. The optimization scheduling solution module is used to define the Lyapunov function of the virtual penalty queue, establish an optimization model with the goal of minimizing the drift of the Lyapunov function plus the energy consumption penalty at the current time, and generate a resource scheduling strategy by solving the optimization model; The resource and energy efficiency execution module is used to perform virtual machine migration on the computing server cluster according to the resource scheduling strategy, and adjust the operating frequency and voltage of the processor in each computing node according to the amount of tasks allocated to each computing node.

Images