A server resource dynamic allocation control method based on load prediction

Abstract

This invention relates to the field of server resource allocation technology and discloses a dynamic server resource allocation control method based on load prediction. The method includes: collecting and preprocessing multi-dimensional server physical state data to construct a smooth multi-dimensional resource state matrix; performing cross-resource coupling trend analysis and dynamically updating the cross-resource coupling coefficient; predicting the target resource load arrival rate at the next moment using a machine learning model combined with the cross-resource coupling coefficient; constructing a cost-sensitive objective function and solving for the optimal resource allocation; finally, distributing resource capacity for dynamic hot-plugging, and fine-tuning the network parameters of the machine learning model online based on actual response latency feedback to form an adaptive control closed loop. This invention overcomes the shortcomings of independent prediction and static scheduling, achieving an optimal balance between preventing sudden traffic risks and considering operating costs, improving global resource utilization, and reducing system response latency.

Description

Technical Field

[0001] This invention relates to the field of server resource allocation technology, specifically to a dynamic allocation and control method for server resources based on load prediction. Background Technology

[0002] With the widespread deployment of cloud computing and microservice architecture, the operation of large-scale cloud services requires real-time dynamic adjustment of CPU quotas, memory limits, and network bandwidth for each node to ensure service stability while maintaining the operating costs and service quality of the underlying hardware infrastructure.

[0003] In current practical application scenarios, existing server resource dynamic allocation technologies often use a single-dimensional independent prediction model to assess future system load pressure. When performing specific resource scheduling and quota allocation, they mostly use heuristic allocation strategies based on fixed thresholds for management.

[0004] However, this conventional allocation control scheme has some shortcomings: First, the single-dimensional independent prediction mechanism ignores the physical and logical coupling relationship between CPU, memory, and network bandwidth. Second, the heuristic scheduling strategy based on fixed thresholds lacks a global optimization mechanism and cannot dynamically adapt to complex business scenarios. This can lead to a mismatch between the underlying resource supply and the actual business needs when the system is dealing with sudden large-scale concurrent loads, which not only reduces server resource utilization but also increases the overall system response latency. Therefore, a dynamic server resource allocation control method based on load prediction is urgently needed to solve these problems. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a dynamic allocation and control method for server resources based on load prediction, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: In a first aspect, embodiments of the present invention provide a server resource dynamic allocation control method based on load prediction, specifically including: collecting physical state data of the server cluster and performing sliding window low-pass filtering to construct a smooth multidimensional resource state matrix; calculating the load change rate of each resource based on the multidimensional resource state matrix, and dynamically updating the cross-resource coupling coefficient to quantify the causal delay and mapping relationship caused by the surge in source request resources to downstream computing and storage resources; constructing a machine learning model and combining the cross-resource coupling coefficient and the load change rate of each resource to predict the target resource load arrival rate at the next moment; constructing a cost-sensitive objective function based on the target resource load arrival rate at the next moment and solving it to obtain the optimal resource allocation amount for execution; performing dynamic hot-plugging and feedback correction of resources, issuing the resource capacity finally decided to be delivered to the cluster according to the optimal resource allocation amount, collecting the actual response latency data after actual effectiveness for comparison, and performing online incremental updates to the network parameters of the machine learning model to form an adaptive control closed loop.

[0007] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the sampling frequency of the multi-dimensional physical state data is determined using an adaptive sampling strategy, specifically including: a comprehensive load fluctuation index based on the previous time window. Adaptively calculate the desired sampling frequency at the current moment. Real-time measurement of the host machine CPU utilization consumed by the probe itself. ,when When the preset probe overhead limit is exceeded, the circuit breaker mechanism is triggered, and the actual sampling frequency is reduced to a security-degraded frequency. Otherwise, the sampling frequency of the multidimensional physical state data shall be the desired sampling frequency. .

[0008] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the dynamic updating of the cross-resource coupling coefficient specifically includes: tracing the call chain topology of the microservice API gateway, extracting the actual physical time difference between resource consumption events, and calculating its statistical expected value as a causal delay constant. Extract the 99th percentile of the end-to-end response time of the microservice cluster within the past business scheduling cycle, and use it as the sliding time window for data sampling. ; in the sliding time window Inside, using a delay offset The sample covariance and sample variance are used to calculate the projection mapping relationship between the source resource load and the target resource load, thus obtaining the dynamic cross-resource coupling coefficient. Its expression is: ; In the formula, For the current moment Target resources Smooth load conditions, This indicates that after experiencing causal delay Previously, related resources Smooth load conditions, and These represent the sample covariance and sample variance, respectively.

[0009] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the offline training process of the machine learning model specifically includes: A spatiotemporal graph convolutional neural network incorporating an attention mechanism is constructed as the core architecture, utilizing the aforementioned cross-resource coupling coefficient. Construct a dynamic directed adjacency matrix with physical causal relationships. ; The smoothed multidimensional resource state sequence is input frame by frame into the graph convolutional network, combined with the dynamic directed adjacency matrix. Extract convolutional features from multi-scale spatial graphs and use a nonlinear activation function to force truncation of negative values ​​to ensure that resource features always maintain non-negative physical meaning in the latent space; The sequence containing topological coupling information after spatial feature extraction is input into a long short-term memory network to capture the long-term periodicity and short-term burstiness dependence of the load fluctuating over time, thereby obtaining the hidden state of the cumulative spatiotemporal joint evolution law. ; Using a global attention mechanism to study the hidden state The global spatiotemporal context vector is obtained by weighted aggregation. It then reduces the dimensionality of the data through a fully connected layer and maps it back to the real physical dimension space to output a prediction term for residual compensation. ; Asymmetric penalty loss function is used The network model parameters are trained by backpropagation, and gradient penalties are applied to cases where the predicted value is lower than the actual requirement, forcing the model to learn the physical safety baseline.

[0010] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the formula for calculating the target resource load arrival rate at the next moment is: ; In the formula, For prediction Class resources in Load arrival rate at any given time For the present time The actual smoothed load arrival rate of the resource class For related resources The rate of change of load, For time step, This refers to the nonlinear residual prediction term output by the spatiotemporal graph convolutional network. These represent the resource types: CPU, memory, and bandwidth, respectively.

[0011] In a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the expression of the cost-sensitive objective function is as follows: ; In the formula, For cost-sensitive objective function, For the system at time The final decision is made regarding the amount of resources to be delivered to the cluster. This is the resource supply cost coefficient. This is the SLA delay penalty coefficient. It represents three resource types: CPU, memory, and network bandwidth.

[0012] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the step of obtaining the optimal resource allocation amount for execution specifically includes: when the optimization problem can be completely decomposed by resources, obtaining the optimal allocation amount in closed form that can be solved independently for each resource item by taking the derivative and setting the derivative to zero. When cross-resource coupling constraints exist, the optimal allocation amount in the closed form is determined. As an initial guess, it is input into the real-time optimizer to perform numerical solutions using the interior-point method, and a usable solution is obtained by projection iteration on the constraint set.

[0013] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the step of collecting and comparing actual response latency data after the actual effect is implemented, and performing online incremental updates to the network parameters of the machine learning model, specifically includes: Calculate the theoretical expected delay based on a queuing-theoretic continuous approximation model under the current physical resource allocation state. ; Collect the actual average end-to-end delay after it takes effect. We construct an online incremental feedback loss function that integrates the error of pure numerical prediction with the penalty for the deterioration of macroscopic queuing delay in the system. ; The online incremental feedback loss function is calculated using an adaptive learning rate optimization algorithm. About machine learning models The gradient of the current network parameters is used to perform parameter fine-tuning and updates.

[0014] As a preferred embodiment of the server resource dynamic allocation control method based on load prediction described in this invention, the desired sampling frequency... The calculation formula is: ; In the formula, To provide a comprehensive load fluctuation index, To ensure a minimum sampling frequency during the stable period, This is the highest sampling frequency during the outbreak period. This is the fluctuation sensitivity coefficient.

[0015] Secondly, embodiments of the present invention provide a server resource dynamic allocation control system based on load prediction, comprising: a data acquisition and preprocessing module for performing multi-dimensional server physical state data acquisition and preprocessing and constructing a smooth multi-dimensional resource state matrix; a coupling analysis module for performing cross-resource coupling trend analysis and dynamically updating cross-resource coupling coefficients; a load prediction module for predicting the target resource load arrival rate at the next moment by combining a machine learning model with cross-resource coupling coefficients; an optimal allocation module for constructing a cost-sensitive objective function and solving for the optimal resource allocation amount; and a feedback adaptive module for performing dynamic hot-plugging of resources and updating model parameters online by comparing latency data.

[0016] The present invention has the following beneficial effects: 1. This invention extracts the real physical time difference between various physical resource consumption actions by tracing the microservice call chain, and dynamically calculates the coupling mapping relationship of multi-dimensional resources at the underlying level. It can restore isolated hardware monitoring indicators into a business logic chain with causal relationship. When the source network requests surge, the system does not need to wait for the downstream memory or CPU to actually spike, but can quantify and perceive the potential incremental pressure in advance, thereby eliminating the prediction delay problem caused by the fragmentation of resource status in large-scale microservice architecture and reducing the overall response latency when dealing with sudden loads.

[0017] 2. In the prediction stage, this invention adopts a hybrid architecture of gradient prediction and deep learning nonlinear residual compensation, which endows the prediction model with instantaneous response capability and implicit feature capture capability when dealing with sudden traffic surges. In the allocation stage, the resource quota allocation is transformed into an optimization problem of seeking the optimal amount of redundancy between the resource idle cost under normal operation and the service delay default penalty under extreme surges. This breaks the traditional mapping logic of allocating as much as predicted. While avoiding the waste caused by blindly reserving the underlying physical computing power, it also avoids the risk of system crash under sudden traffic surges, thereby improving the global resource utilization of the cluster server.

[0018] 3. This invention utilizes the underlying interface to perform resource hot-plugging without interrupting service, and uses the deviation between the actual end-to-end response delay after allocation and the expected delay in queuing theory as a signal for online incremental feedback. When the system faces unknown burst traffic characteristics, it triggers online fine-tuning and updating of the machine learning model, thereby enabling the system to complete self-correction and performance recovery. Furthermore, by setting a limited learning rate and bias penalty, it distinguishes between normal load jitter and actual prediction failure, achieving safe and continuous evolution of the model and avoiding catastrophic forgetting.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The present invention provides a flowchart of a server resource dynamic allocation control method based on load prediction.

[0022] Figure 2 This is a schematic diagram of the S4 process provided by the present invention.

[0023] Figure 3 This is a schematic diagram of a server resource dynamic allocation control system based on load prediction provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 Existing server resource dynamic allocation technologies often employ independent predictions in a single dimension, ignoring the physical and logical coupling between CPU, memory, and bandwidth. Furthermore, they mostly adopt heuristic allocation strategies based on fixed thresholds, lacking a global optimization mechanism. When dealing with sudden loads, this can easily lead to a mismatch between resource supply and actual demand, reducing resource utilization and causing an overall deterioration in response latency.

[0026] To solve the above technical problems, such as Figure 1As shown, Embodiment 1 of the present invention provides a method for dynamic allocation and control of server resources based on load prediction. Specifically, Embodiment 1 takes a high-concurrency video streaming microservice cluster of a large cloud service provider as an example: the cluster is deployed with a Kubernetes-based container orchestration system, which needs to dynamically adjust the CPU quota, memory limit and network bandwidth of each node in real time.

[0027] In the specific implementation of Example 1: First, multi-dimensional server status observation and preprocessing is performed. Physical status data of the server cluster is collected and subjected to sliding window low-pass filtering to construct a smooth multi-dimensional resource status matrix. This method eliminates transient hardware sampling noise through low-pass filtering, ensuring the true upward trend of business load and guaranteeing the security of the monitoring probe. Second, cross-resource coupling trend analysis is performed. Based on the multi-dimensional resource status matrix, the load change rate of each resource is calculated, and the cross-resource coupling coefficient is dynamically updated to quantify the causal delay and mapping relationship caused by the surge in source request resources on downstream computing and storage resources. This method breaks the information silos of traditional single-dimensional independent physical resource monitoring, enabling early perception and quantification of potential incremental pressure before physical bottlenecks occur in downstream resources, avoiding prediction delays caused by fragmented resource status in large-scale microservice architectures. Then, a machine learning model is constructed and combined with the cross-resource coupling coefficient and the load change rate of each resource to predict the target resource load arrival rate at the next moment. This method not only endows the prediction model with instantaneous response capability in the face of sudden traffic surges, but also retains the advantage of deep neural networks in accurately capturing long-term periodicity and implicit features. Thus, under the premise of taking into account physical constraints and system stability, it achieves advanced prediction and response to multi-dimensional resource load of the server, improving response efficiency. Next, a cost-sensitive objective function is constructed based on the target resource load arrival rate at the next moment and solved to obtain the optimal resource allocation for execution. This method transforms the allocation of underlying physical resources into a mathematical optimization problem that takes into account economic operating costs and service quality, avoiding the waste of physical computing power caused by static reservation based on experience, and preventing the risk of system collapse under sudden traffic surges, achieving global optimization of underlying infrastructure stability and cloud service operation economic benefits. Finally, dynamic hot-plugging and feedback correction of resources are performed. Based on the optimal resource allocation, the resource capacity to be delivered to the cluster is distributed to the cluster, and the actual response latency data after the actual effect is collected for comparison. The network parameters of the machine learning model are updated online incrementally to form an adaptive control closed loop. This method not only ensures that the model can continuously evolve online when facing unknown burst traffic characteristics and avoids catastrophic forgetting, but also applies gradient penalties only when the actual system latency deteriorates, thereby maintaining the long-term high reliability and robustness of the resource allocation strategy.

[0028] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a detailed description of the server resource dynamic allocation control method based on load prediction is provided, specifically including the following: S1. Perform multi-dimensional server status observation and preprocessing, and collect physical status data of the server cluster. This includes the following sub-steps: S11. Using probes deployed in the kernel layer (such as eBPF technology) at a sampling frequency Collect physical status data of the server cluster, specifically including: CPU context switching frequency, memory page fault rate, and network I / O throughput.

[0029] In this embodiment, to avoid the backlash from probe system overhead caused by fixed high-frequency sampling, the sampling frequency is... An adaptive sampling strategy is used to determine the method, and the specific implementation steps are as follows: S111. The system adaptively calculates the current time based on the load fluctuation characteristics of the previous time window. Desired sampling frequency : ; ; In the formula, This is a comprehensive load fluctuation index; This represents three resources: CPU, memory, and network bandwidth. These are the normalized weighting coefficients; Within the previous time window The rate of change of load on the resource class; To ensure a minimum sampling frequency during the stable period; This is the highest sampling frequency during the outbreak period; This is the fluctuation sensitivity coefficient.

[0030] For example, normalized weighting coefficients To uniformly eliminate scaling differences between different physical resource dimensions, the total physical capacity of the resources corresponding to the server node is taken. The reciprocal, These are the physical limits of the underlying hardware, such as the total number of CPU cores and the total physical memory capacity.

[0031] For example, the minimum sampling frequency during the stable period Used to maintain basic heartbeat and smoothing state monitoring, the maximum permissible scheduling delay time constrained by the Service Level Agreement (SLA) signed between the cloud service provider and the tenant is taken. The reciprocal of.

[0032] For example, the highest sampling frequency during a burst. Used to capture fine-grained features under drastic load fluctuations, its limitations are imposed by the host machine's computing power protection mechanisms. The probe execution overhead can be determined through micro-benchmark testing, such as: ;in, A globally mandated limit on probe CPU resource usage for the operation and maintenance system (e.g., a limit of 0.01%, or 1%). This is the CPU time required for a single eBPF probe to complete execution in kernel space, which is measured offline using the kernel's built-in bpf_prog_test_run interface.

[0033] For example, the fluctuation sensitivity coefficient The convergence speed used to control the frequency as volatility increases is determined by the statistical characteristics of offline historical data in an unsupervised state. The specific method is as follows: The system extracts the historical comprehensive load volatility index sequence within a complete past business cycle (e.g., 30 days). And take its 99th percentile (P99) as the saturation reference point. ,calculate It represents the volatility index at the current moment. Reaching a historically high level At that time, the exponent term Become (Approximately equal to 0.368), at which point the adaptive sampling frequency will smoothly approach 63.2% of the extreme range, which not only ensures the sensitivity of frequency increase, but also uses the mathematical characteristics of exponential asymptoticity to prevent premature frequency saturation in non-extreme states.

[0034] S112. Implement probe overhead limit guarantee and circuit breaker mechanism, specifically: The system runs a lightweight overhead monitoring daemon in parallel in the eBPF probe runtime space to measure the host CPU utilization consumed by the probe itself in real time. The desired sampling frequency calculated above. This determines the final actual sampling frequency. : ; In the formula, This represents the host CPU utilization consumed by the eBPF probe program itself in the previous cycle. The preset probe overhead upper limit and circuit breaker threshold; To determine the frequency of safety degradation after a circuit breaker trips, a setting is typically used. When the probe's own overhead exceeds the limit, the adaptive logic is forcibly blocked and the frequency is reduced, thus ensuring the security of the monitoring probe from a physical mechanism.

[0035] For example, the probe overhead upper limit circuit breaker threshold Equivalent to the native basic scheduling overhead ratio of the operating system kernel, its calculation formula is as follows: ;in, The time taken for a single hardware context switch in the current server CPU architecture is determined by the physical instruction cycle for saving and restoring CPU registers, and is measured offline using the perfsched tool; This is the minimum task scheduling time slice allowed by the current operating system kernel (such as sched_min_granularity_ns in the Linux kernel).

[0036] For example, the frequency of security degradation after a circuit breaker is triggered The expression is determined by reverse engineering based on the actual degradation performance of the probe at the moment of triggering the circuit breaker, and is as follows: ;in, In order to be in The actual measured execution time of a single probe operation when the circuit breaker is triggered. This is based on the probe's already degraded actual execution time, and is calculated backwards to determine the force required to meet the circuit breaker's requirements. Overhead limit, the highest theoretical frequency currently allowed, and then... By taking the minimum value, it is ensured that the monitoring component can run at a safe frequency in a harsh physical kernel environment, without deadlocking the host machine, and while retaining a minimum self-rescue monitoring capability.

[0037] S113. Implement an adaptive sampling frequency recovery mechanism based on state hysteresis and smooth prediction. The specific implementation steps are as follows: During the period when the system is in the circuit breaker frequency reduction state, predict in real time if it can be directly restored to the current desired sampling frequency. The resulting theoretical probe CPU overhead : ; In the formula, To theoretically predict probe overhead; The desired sampling frequency is adaptively calculated for the current moment; This represents the actual physical measurement time taken for a single probe execution at the current moment.

[0038] To prevent probe overhead from fluctuating repeatedly at the critical point, a probe overhead recovery threshold with buffering characteristics is set. And introduce a continuous safety status assessment counter. Continuously monitor the current system status: ; ; In the formula, This is the hysteresis buffer coefficient, and its value range is... This is used to provide a safety margin below the circuit breaker threshold; The preset probe overhead upper limit and circuit breaker threshold; This is a continuous safety state assessment counter; its initial value is reset to [value] when the circuit breaker is triggered. .

[0039] When evaluating the counter Reaching or exceeding the preset continuous monitoring window length (Right now When the system determines that the underlying computing environment has stabilized and moved out of the high-load danger period, it releases the circuit breaker mechanism and remounts the actual sampling frequency for the next cycle to the desired sampling frequency. and will Reset to Otherwise, the system will continue to forcibly maintain the security degradation frequency. To ensure the stability of the underlying hardware.

[0040] For example: In this embodiment, the hysteresis buffer coefficient To address the uncertainty and fluctuations in probe execution time, and to prevent frequent activation and shutdown of the fuse mechanism near the red line due to minor noise, therefore, The size of the buffer safe zone should be determined by the statistical dispersion of the probe execution time: when the probe time is stable, the buffer safe zone can be extremely small; when the probe time fluctuates drastically, a very large buffer safe zone must be reserved. The convergence characteristics of the natural exponential function, combined with the coefficient of variation of the probe execution time, are used to determine... The expression is: ; in, This represents the statistical mean of the execution time of a single eBPF probe, as measured during the intervention-free benchmark test. The standard deviation of the execution time of a single eBPF probe, as measured during equivalent benchmark testing.

[0041] Specifically: when the probe performs stably, , This means the system does not need to reserve a buffer, as long as the estimated overhead is lower than [the threshold]. Recovery can be assessed immediately; when the underlying environment is extremely harsh and probe latency fluctuates greatly, the exponential term tends to be... , By reducing the probe's estimated overhead to an extremely low level, the system will require the probe's estimated overhead to be reduced before allowing it to enter the recovery assessment, thus ensuring the safety of probe operation.

[0042] In this embodiment, the continuous monitoring window length The number of consecutive safe sampling periods required to confirm that the system has truly moved out of the high-load danger zone depends on the complete business processing cycle of the microservice cluster. Only when the physical duration of the continuous safe state covers the longest possible end-to-end request lifecycle can it be confident that the upstream flood has completely passed. The sliding time window is determined using the 99th percentile (P99) of the end-to-end response time extracted through the microservice API gateway in Example 1. In combination with the current frequency of safety degradation under circuit breaker conditions ,definition The expression is: ; in, This represents the 99th percentile of the end-to-end response time of the microservice cluster within the past business scheduling cycle. Indicates the frequency of security degradation after a circuit breaker is triggered; This is a rounding function that ensures the output is a physically meaningful positive integer number of cycle periods.

[0043] S12. The field terminal performs sliding window low-pass filtering on the acquired raw time series data to remove transient hardware sampling noise and construct a smooth multidimensional resource state matrix. The specific implementation steps are as follows: S121, Non-equal interval time step extraction: Record the first... Timestamp of the actual sampling point And calculate the current sampling point and the previous valid sampling point. The actual physical time interval between : ; Due to sampling frequency Adaptive fluctuations It is dynamically changing.

[0044] S122. To filter out instantaneous high-frequency spikes caused by hardware microarchitecture (such as a surge in CPU cache misses) while preserving the true load trend, the system calculates the dynamic smoothing weight coefficient for the current sampling point. : ; ; In the formula, For dynamic smoothing coefficients; This is the physical time constant of the low-pass filter.

[0045] S123. Using the solved smoothing coefficient For all types of resources ( )exist Original sampled discrete values ​​at time 1 Perform filtering and updating to obtain smoothed state values. : .

[0046] S124. At macroscopic moments when model prediction or resource scheduling is being performed. The on-site terminal reads the latest updated smoothed state value of each dimension of resources. , , They are combined to construct a smooth multidimensional resource state matrix at the current moment. : .

[0047] In this embodiment 1, the system combines an adaptive sampling mechanism based on load fluctuations with probe overhead circuit breaker protection to achieve high-precision acquisition of the server's CPU, memory, and network underlying status while ensuring the stable operation of the host system. Subsequently, a dynamic smoothing weighted low-pass filter based on real physical time intervals is used to eliminate transient noise at the hardware microarchitecture level, breaking the rigid mode of traditional fixed-frequency monitoring. This avoids resource waste during stable periods and prevents the monitoring tool itself from causing system crashes, thus ensuring the security of the monitoring probe from a physical mechanism perspective. Specifically, for example, when a cloud cluster encounters a sudden surge in concurrent access during a promotional activity, the system's comprehensive load fluctuation index... The frequency increases sharply, and the adaptive module will automatically adjust the sampling frequency. The system is designed to detect micro-level resource bottlenecks. If high-frequency sampling causes the eBPF probe's CPU consumption to exceed a preset kernel meltdown threshold (e.g., 1%), the system will immediately trigger a protection mechanism, forcibly reducing the sampling frequency. Reduced to the safety baseline To prevent the monitoring process from crashing the server, during the data preprocessing stage, if the underlying data captures extreme spikes caused by a single CPU cache miss or network jitter, the low-pass filter will utilize a dynamic smoothing coefficient. Smooth it out to ensure that the state matrix output to the model is accurate. It presents the true upward trend of business load, rather than random noise from the underlying hardware.

[0048] S2. Perform cross-resource coupling trend analysis and dynamically update the cross-resource coupling coefficient, which includes the following sub-steps: S21. Based on the above multidimensional resource state matrix Calculate various resources (CPU) ,Memory ,bandwidth In the current First derivative of load at time t It is used to characterize the rate of change of load for various resources.

[0049] S22. By tracing the call chain topology of the microservice API gateway, calculate and dynamically update the cross-resource coupling coefficient. The specific implementation steps are as follows: S221. In a microservice architecture, resource consumption exhibits a waterfall-like delayed delivery characteristic. For example, the gateway first receives network I / O packets, and after a period of queuing and interrupt response time, it triggers a CPU context switch, which in turn causes memory allocation. The system extracts resources by parsing the Span tracing logs of all requests within the current period in the API gateway. Consume events to resources The actual physical time difference between consumption events is used to calculate its statistical expectation, which is then used as a causal delay constant. : ; In the formula, This represents the total number of call chain samples within the current evaluation period. and The first In the call chain, trigger Class Resources and Hardware timestamps of core resource-consuming actions.

[0050] S222. The system extracts the 99th percentile (P99) of the end-to-end response time of the microservice cluster in the API gateway within the past business scheduling cycle as the sliding time window for data sampling. : ; In the formula, This refers to the complete lifecycle time from request entry into the gateway to the completion of the full response. It uses the actual P99 response time of the business as the statistical window. This ensures that the sample data used to calculate the coupling coefficient covers exactly one complete request lifecycle, guaranteeing data integrity and timeliness.

[0051] S223, within the established characteristic time window Inside, the smoothed state value output in step S12 is used. Computing resources rate of change of resources The projection mapping relationship is used to obtain the cross-resource coupling coefficient. : ; ; In the formula, For the current moment Target resources Smooth load conditions; This indicates that after experiencing causal delay Previously, source resources Smooth load conditions; For time window The total number of valid sampling points within the area; and These represent the sample covariance and sample variance, respectively.

[0052] It should be noted that when the source resources Extremely stable during the window period (i.e.) When the denominator approaches 0, the system automatically triggers a calculation bypass according to the limit arithmetic rules, letting... To avoid system computational crashes.

[0053] In this embodiment 1, the system extracts the actual physical time difference between different physical resource consumption actions, i.e., the causal delay constant, by tracing the Span call chain of the microservice API gateway. And establish a sliding time window based on the actual P99 end-to-end response time in the business; By utilizing the mathematical expectation of the sample covariance and variance with delay offset, the cross-resource coupling coefficient reflecting the mapping strength is dynamically calculated. This breaks down the information silos of traditional single-dimensional, independent physical resource monitoring, restoring isolated hardware-level monitoring metrics into a waterfall-style business logic chain with temporal causal relationships. This allows the system to accurately quantify the response latency caused by a surge in source resource requests to downstream computing and storage resources. Thus, it can detect and quantify potential incremental pressure before downstream resources actually experience physical bottlenecks, eliminating the prediction lag caused by fragmented resource states in large-scale microservice architectures. Specifically, for example, when a high-concurrency video streaming microservice cluster faces a sudden traffic surge, the system detects a sharp increase in network requests received by the gateway, i.e., network bandwidth... Load change rate The system showed a significant increase; through continuous analysis of the call chain logs, it was discovered that there was a physical causal delay of approximately 50 milliseconds between the arrival of network I / O packets and the memory allocation operations required for subsequent video transcoding and unpacking (i.e., (equal to 50 milliseconds); within the current dynamic time window, the algorithm calculates the cross-resource coupling coefficient of network bandwidth mutation on memory consumption. The value is 0.5. This indicates that if the network bandwidth load suddenly surges by 800 Mbps / s at the current moment, the system can instantly infer that the memory load will experience a compensatory surge of about 400 MB / s after 50 milliseconds without waiting for the memory index to deteriorate. Without this cross-resource coupling analysis mechanism, the system can only react passively when the memory actually starts to spike after 50 milliseconds, which can easily lead to OOM failures due to memory exhaustion or severe long-tail response delays.

[0054] S3, Building Machine Learning Models By combining the cross-resource coupling coefficient with the load change rate of each resource, the target resource load arrival rate at the next moment is predicted, which specifically includes the following sub-steps: S31. Using a Spatiotemporal Graph Convolutional Neural Network (STGCN) with an attention mechanism as the core architecture, construct a machine learning model. Collect a smoothed, multi-dimensional resource state matrix over the historical runtime of the server cluster. Given a historical observation time window Extract the aligned smooth multidimensional resource state sequence , at any time Input tensor Representing the current load values ​​for CPU, memory, and bandwidth, listed in chronological order. The dataset is divided into training, validation, and test sets. It should be noted that, to ensure stable gradient propagation in the latent space of the neural network while preserving its physical meaning, the input multidimensional resource state matrix is ​​processed using... Standardized processing.

[0055] In this embodiment, the specific layer-by-layer processing steps of the spatiotemporal graph convolutional neural network are as follows: S311. Before performing spatial graph convolution, the system utilizes the cross-resource coupling coefficients obtained in step S22. Construct a dynamic directed adjacency matrix with physical causal relationships. .

[0056] Matrix elements are defined as: when hour, ;when (When it is its own node) .

[0057] It should be noted that, due to The dimensions are When the adjacency matrix Input features Perform matrix multiplication At that time, the system automatically completes cross-dimensional conversion and accumulation of different physical resource dimensions to ensure the physical consistency of the input layer.

[0058] S312. Extracting Multi-Scale Spatial Graph Convolutional Features: To extract the nonlinear spatial dependencies of multi-dimensional resources at the same time, the state sequence is input frame by frame. Layered Graph Convolutional Network (GCN). Hidden state tensors of the layer The update formula is: ; In the formula, The node features output from the previous layer (when) hour, ); For the first The learnable weight matrix of the layer graph convolution is used to map physical features to a high-dimensional latent space; It is a spatial bias vector; It is a nonlinear activation function, and in extreme cases, such as when the load suddenly drops to near 0 or when a negative fluctuation is calculated, the activation function forcibly truncates negative values ​​to ensure that the resource characteristics always maintain a non-negative physical meaning in the latent space.

[0059] S313. After spatial feature extraction, a sequence containing topological coupling information is obtained. Input it into a Long Short-Term Memory network (hidden state dimension). Set to 64 or 128 to capture long-term periodicity and short-term burstiness of load fluctuations over time. Hidden state Controlled by a forget gate, input gate, and output gate: ; In the formula, ( (For hidden layer dimensions), which records up to The spatiotemporal joint evolution law accumulated over time.

[0060] S314, Global Attention Mechanism and Residual Output Mapping: Considering the historical window Within this system, not all moments contribute equally to the prediction of the next moment; therefore, a global attention mechanism is introduced. ; ; ; In the formula, These are learnable parameters in the attention mechanism; For attention weight scalars, satisfying ; This is the weighted aggregated global spatiotemporal context vector.

[0061] Finally, the high-dimensional context vector is processed through a fully connected layer. Dimensionality reduction and mapping back to the true physical dimension space are used to output prediction terms for residual compensation: ; In the formula, To output the mapping weights, For output bias, the model output In terms of units, it is reduced to the task arrival rate or byte change rate of CPU, memory, and bandwidth.

[0062] S32. Because the risk of downtime due to under-allocation of resources in server resource scheduling is far greater than the cost waste caused by over-allocation of resources, the system adopts an asymmetric penalty loss function. For network model parameters Perform backpropagation training: ; In the formula, This is the total loss evaluation value for the current training batch of the model; This represents the total number of samples in the training batch. The type of resource that represents the target prediction; For the sample middle The actual load reach rate of the resource class; For forward propagation networks based on input feature vectors Output predicted load factor; As the weighting coefficient for underprediction penalty, set Apply a higher gradient penalty to cases where the predicted value is lower than the actual demand, forcing the model to learn the physical safety baseline; To set the overprediction penalty weighting coefficient, set This allows the model to make moderately higher predictions when faced with traffic uncertainty, in exchange for system stability; This is the L2 regularization term, used to constrain parameter size and prevent overfitting.

[0063] S33. After completing offline training of the model and parameters After solidification, it is deployed in real time on online inference nodes, and the real-time cross-resource coupling coefficient calculated in step S22 is used. with the first derivative of the load and machine learning models Combine and calculate the next moment. Predicted target resource load reach rate : ; In the formula, For prediction Class resources in Load arrival rate at any given time; For the present time The actual smoothed load arrival rate of the resource class; For related resources The rate of change of load; For time step; Spatiotemporal graph convolutional networks based on historical features The output nonlinear residual prediction term.

[0064] In this embodiment 1, by constructing a spatiotemporal graph convolutional neural network that integrates cross-resource coupling topology, the causal relationships of the physical world are injected into the deep learning model. In the latent space feature extraction, and using an asymmetric penalty loss function The model is guided to establish a bottom-line awareness of resource security. In the final output stage, a hybrid architecture combining physical kinematic gradient prediction and deep learning nonlinear residual compensation is adopted. This combines physically meaningful deterministic evolutionary trends with complex nonlinear biases in machine learning predictions. This approach endows the prediction model with the ability to respond instantly to sudden traffic surges while retaining the advantage of deep neural networks in accurately capturing long-term periodicity and implicit features. Thus, while considering physical constraints and system stability, it achieves proactive prediction and response to multidimensional server resource loads, improving response efficiency. Specifically, for example, this involves setting a time step. The target predicted resource is memory. ,current Actual memory consumption at any given moment At this point, the system detected network bandwidth. Due to a sudden surge in traffic, its rate of change The coupling coefficient of the network to memory Meanwhile, the CPU rate of change is gradual, and its impact on memory is negligible. Machine learning model After assessing the historical cycle characteristics, the output residual compensation term is: Substitute into the formula to calculate This result indicates that the system predicts the memory load will reach [a certain level] in the next second. If predictions are made solely based on historical data, the strong coupling impact caused by the surge in bandwidth will be ignored, leading to prediction delays and reduced response efficiency.

[0065] S4, Based on predicted load arrival rate Construct a cost-sensitive objective function And solve for the optimal resource allocation for execution, referring to... Figure 2 As shown, the specific steps are as follows.

[0066] S41. For each type of resource Input Predicted Arrival Rate and resource supply cost coefficient SLA delay penalty coefficient and the physical limit of this type of resource. .

[0067] Specifically: the physical upper limit of this type of resource. The physical upper limit of the system's current resources can be calculated by accessing the operating system kernel to obtain the factory-set maximum values ​​of the underlying hardware (such as the total number of logical cores or the total amount of physical memory), and then subtracting the non-allocatable reserved space that the operating system forcibly locks to ensure basic operation (such as interrupt handling core binding or kernel reserved memory). .

[0068] Specifically: resource supply cost coefficient The actual purchase price of this type of hardware asset can be amortized over a unit of time according to the statutory depreciation period, and then multiplied by the product of the rated physical energy consumption per unit of resource obtained by actual measurement through the motherboard sensor and the standard grid electricity price. This method transforms the theoretical optimization weight into the economic cost generated by providing each unit of resource in reality.

[0069] Specifically: SLA latency penalty coefficient The true second-level cost of service interruption can be calculated by dividing the maximum service penalty stipulated in the contract by the corresponding service unavailability time threshold. Then, the cost can be determined by the proportion of the hardware procurement cost of the target resource in the total physical hardware investment of the cluster. This method decomposes the global penalty proportionally and distributes it to a single resource dimension, ensuring that the constraint of the optimization target has commercial and legal support.

[0070] S42. Represent the resource allocation problem as an optimization problem that sums the costs of each type of resource, and construct a cost-sensitive objective function. : ; The constraints are: ; In the formula, For the system at time The final decision is made regarding the amount of resources to be delivered to the cluster. S3 represents the predicted arrival rate. For linear cost items of resource supply, It represents the average time cost or rental cost per unit of resource.

[0071] S43. When the optimization problem can be completely decomposed according to resources, i.e., there are no cross-resource coupling constraints, the cost-sensitive objective function... For each resource Solvable independently; for a single resource item, specify: ; right Taking the derivative and setting it to zero yields the analytical optimal value: ; Solving the above equation, we get: ; Therefore, the optimal allocation in closed form is obtained: ; Require , , Perform a second derivative test on the solution: .

[0072] S44. If there are cross-resource coupling constraints or if all resources need to satisfy a certain node's total quantity constraint, then the problem cannot be solved using resource-independent closed-form solutions. Therefore, this embodiment adopts the following numerical solution process: The objective function and constraints are input into the real-time optimizer, and the interior-point method is preferred as the solver to ensure convergence speed. In soft real-time scenarios, the finite-step projection gradient method can be used and approximate solutions can be accepted.

[0073] In the solution process, closed-form solutions are used first. As an initial guess, projection iterations are then performed on the constraint set, which usually yields a usable solution that satisfies the constraints within a finite number of iterations.

[0074] If discretization is required, such as CPU granularity of 0.25 cores or memory quantization of 1MB, the results are rounded according to granularity and the constraints are re-verified after the numerical solution converges; if the hard constraints are violated after rounding, the process is rolled back to the projection step to continue fine-tuning.

[0075] If a certain Then: Set as The remaining unmet load Mark as shortage; attempt to compensate with other resources, such as increasing bandwidth or CPU, to alleviate memory pressure according to the cross-resource coupling strategy, based on the coupling coefficient. If the shortage cannot be eliminated through compensation, an overload protection strategy, such as rate limiting, degradation, or migration request, is triggered, and an alarm is generated at the return layer.

[0076] To ensure the real-time performance of the control loop, the optimizer should be set with a maximum solution time threshold. (For example, 50–200 milliseconds, depending on cluster size and control cycle). If the numerical solution is not in If convergence occurs, then revert to using the closed-form solution mentioned above or the allocation from the previous time step as a safe solution and log it for offline analysis.

[0077] It is important to note that during the vacuum period when global concept drift is triggered and the new model has not yet been fully trained, the nonlinear predictions of the original machine learning model (i.e., in S33) will be affected. The model is no longer reliable and a resource allocation degradation strategy needs to be implemented to prevent system crashes. The specific implementation steps are as follows: When in a model drift vacuum period, the system forcibly masks the nonlinear residual prediction term output by the neural network (i.e., makes...). The prediction of resource load arrival rate is completely degraded to a pure kinematic gradient calculation based on the underlying physical causality: ; Meanwhile, in constructing a cost-sensitive objective function At this time, the system forcibly injects a penalty multiplier into the resource pool, increasing the SLA latency penalty coefficient. Magnified to the original value times ( This strategy sacrifices short-term operating costs (i.e., resource supply cost coefficient). The weight of the weight is relatively reduced, which forces the optimization solver to find the optimal allocation with high safety redundancy, ensuring the stability of the underlying cluster during business mode switching.

[0078] In this embodiment, based on the predicted load arrival rate Construct a cost-sensitive objective function And in guarantee Solving for optimal allocation under hard constraints improves resource utilization under normal conditions. Specifically, for example, taking memory as an example, assuming... MB / s, Yuan / MB, Yuan / second, calculated according to the analysis of S43: This result indicates that, in order to avoid incurring high SLA default and delay compensation, facing For the projected load, the resource surplus margin that best aligns with economic efficiency should be set at approximately [value missing]. It breaks away from the traditional resource scheduling logic of predicting and allocating resources, transforming the physical resource allocation problem into a mathematical optimization problem that balances economic operating costs and service quality by introducing a cost-sensitive objective function. By dynamically calculating the optimal redundancy between the cost of idle resources during normal operation and the penalty for default due to extreme delays, the waste of physical computing power caused by static reservation based on experience is avoided while ensuring the service level agreement of the microservice cluster. It also avoids the risk of system crash under sudden traffic surges, and ultimately achieves the global optimization of the stability of the underlying infrastructure and the economic benefits of cloud service operation.

[0079] S5. Dynamic hot-swapping and feedback correction of resources, specifically including the following steps: S51. The system uses the underlying container interface of Kubernetes CRI / CNI to ultimately determine the resource capacity delivered to the cluster. The system allows for dynamic hot-swapping of CPU and memory without service interruption via cgroups.

[0080] It is important to note that in real-world cloud-native environments, hot-plugging of underlying container resources faces uncertainties. Furthermore, due to runtime state isolation, the application layer often cannot perceive instantaneous changes in the underlying physical capacity, potentially causing microservices to freeze or experience memory overflows. In such cases, a hot-plugging failure rollback mechanism must be implemented. The specific implementation steps are as follows: S511, the system uses the Kubernetes Container Runtime Interface (CRI) to ultimately determine the resource capacity delivered to the cluster. The command is sent to the Kubelet component on the host machine and leverages the dynamic adjustment capabilities of Linux cgroups to modify the target container's resource limit files in real time (e.g., modifying the cpu.max and memory.max parameters for cgroups v2 architecture) to perform physical scaling up or down without interrupting services. After the command is sent, the system sets an asynchronous confirmation delay window (e.g., 500 milliseconds). Upon expiration, the system reads the actual parameter values ​​of the cgroups mount point on the kernel side to perform a hard verification of whether the physical configuration has been successfully written to disk.

[0081] S512. Considering that most microservice applications built on the JVM (Java Virtual Machine) or Golang statically read and lock the system's resource limits during startup, direct modification of the underlying cgroups quotas can lead to a misalignment between the application layer and the physical layer in resource perception (e.g., JVM heap memory not shrinking proportionally after memory scaling down, resulting in direct killing by the kernel's OOMKiller, or application thread pool capacity not increasing accordingly after CPU scaling up, leading to idle computing power). To eliminate this misalignment, after successful verification in S511, the system sends a custom signal (such as SIGUSR1 or SIGUSR2) to the target container's main process via inter-process communication, or calls the runtime reload webhook pre-registered by the microservice. Upon receiving this signal or call, the microservice dynamically resets its internal garbage collector trigger threshold and worker thread pool capacity, forcing the application layer logic to align with the latest underlying physical resources. Maintain consistency.

[0082] S513. If the hard check of the status in S511 fails (usually due to the physical kernel refusing to allocate resources because of host machine resource fragmentation), or if the application layer in S512 fails to return a confirmation of successful reload within the preset timeout window, the system immediately blocks subsequent logic and triggers the following failure rollback process: In-place rollback: Forcefully terminates the current hot-plug operation and safely rolls back the cgroups resource limit parameters of the target container to the safe allocation amount of the previous stable scheduling cycle. .

[0083] Global scheduling intervention: Mark the node's resource shortage or hot update anomaly as a taint and report it to the Kubernetes global scheduling control plane, triggering the corresponding level of system alarm.

[0084] Space for safety: If the current period is a sudden traffic surge and the local physical resources of the node are indeed unable to meet the predicted load demand, the system will coordinate with the microservice API gateway to start cross-node traffic redirection, or directly trigger automatic horizontal scaling of Pods based on the nearest healthy node. By horizontally increasing the number of instances, the predicted incremental load can be digested, ensuring that the overall service of the cluster is not interrupted.

[0085] S52. Collect actual server response latency data after implementation, compare it with the expected latency of the queuing theory model, calculate the loss function, and use the backpropagation algorithm to improve the machine learning model. The network parameters are updated incrementally online to form a continuously evolving adaptive closed loop. The specific implementation steps are as follows: S521. Collect the actual average end-to-end latency of the server after it takes effect within the current control period. and the actual consumption of various resources Simultaneously, based on a continuous approximation model using queuing theory, the theoretically expected delay under the current physical resource allocation state is calculated. : ; In the formula, For resources The physical processing time mapping constant is used to unify the calculation dimensions of overall latency across various resource dimensions; This indicates the actual physical redundancy of various resources in the current period; For a very small positive number (such as This is used to prevent resource redundancy from approaching its limit when the system is at full load, i.e., when the actual consumption approaches the allocated amount. A division-by-zero overflow occurs, causing mathematical calculations to crash.

[0086] For example, resources Physical processing time mapping constant With sliding time window As a statistical period, and combined with the actual observations of the kernel eBPF probe, it is dynamically calculated and can be expressed by the following formula: ; In the formula, For the current moment Regarding resources Calculated physical processing time mapping constant; For sliding time windows; In order to be in The total number of valid discrete sampling points actually captured by the kernel eBPF probe within the window period; To be at the sampling time The system The actual load sampling arrival rate of the resource class; For the first The second sampling and the first The physical time interval between samples; Indicates in During the window period, the microservice API gateway records and successfully processes the total number of request call chains.

[0087] S522. Construct an online incremental feedback loss function that integrates the realistic representation of physical systems. The pure numerical prediction error of the model feedforward is combined with the penalty for the deterioration of the system's macroscopic queuing delay: ; In the formula, For machine learning models The predicted load in the output; For the current moment The delay deviation penalty scaling conversion factor is used to balance the difference in physical dimensions and orders of magnitude between the square of the load capacity and the square of the time delay; The function is used to ensure that the model prediction is severely underfitted only when the actual delay deteriorates and exceeds the expectations of queuing theory, and then additional physical inverse penalties are imposed, while no negative interference is imposed on the model gradient when the system performance is better than expected.

[0088] For example, delay bias penalty scaling conversion factor With sliding time window As a statistical period, in each control period The variance of the actual load in each dimension within the window is dynamically calculated to be the ratio of the sample variance of the actual latency, which can be expressed by the following formula: ; In the formula, For the current moment Delay bias penalty scaling conversion factor; Indicates within the time window Inside, the actual data collected by the system's bottom-level probes Statistical sample variance of physical load arrival rate for resource categories; This represents the statistical sample variance of the actual end-to-end latency recorded by the gateway within the same time window. This represents the minimum observation variance baseline at the hardware level. This represents the time taken for a single hardware context switch in the current server CPU architecture.

[0089] S523. Employ the adaptive learning rate optimization algorithm with momentum to calculate the online incremental feedback loss function. About machine learning models Current network parameters The gradient is used to perform online fine-tuning updates of the parameters: ; In the formula, These are the updated model parameters; For the current moment The restricted base learning rate for online fine-tuning is usually set to a very small percentage of the offline training learning rate. This is intended to ensure that the model evolves robustly on new data while avoiding catastrophic forgetting that could cause the original spatiotemporal graph topological feature extraction capability to collapse. This is a second-moment moving average estimate based on historical gradients; To prevent the minimum value of the denominator being zero (such as...) ).

[0090] For example, the system in each control cycle By extracting the physical capacity limit of the underlying hardware and the hard timeout limit of the network protocol, the maximum physical limit of the computing system's theoretically possible crash loss at the current moment is determined. and the current online actual losses In comparison, calculate the current moment. Limited base learning rate for online fine-tuning : ; Among them, the theoretical maximum physical limit loss The calculation formula is: ; In the formula, For the current moment Limited base learning rate for online fine-tuning; The current real system feedback loss calculated in step S522; These are the physical limits of the server's underlying hardware, such as total physical memory capacity and maximum line speed of the network card. They are used to measure the maximum theoretical error tolerance when load prediction fails completely (i.e., the predicted value is 0, but the actual load is full, or vice versa). Configure a global forced disconnection timeout threshold (e.g., the default 60 seconds) for the microservice API gateway at the network protocol layer.

[0091] S53. To compensate for the limitations of online fine-tuning, an exponential moving average mechanism is used to monitor online feedback loss over a long period, and a global model drift trigger is constructed. The specific steps are as follows: The system maintains a global concept drift indicator in real time. The online incremental feedback loss function calculated in step S522 Perform smooth tracking: ; In the formula, This serves as the model drift indicator for the current moment. This is the historical decay smoothing coefficient, with a value range of [value range missing]. It is used to filter out instantaneous loss jitter caused by a single burst and extract trend errors; The actual system feedback loss at the current moment.

[0092] Retraining trigger condition: Set the average loss on the offline validation set as the baseline value. When continuous Within each control cycle, satisfy At that time, the system determines that a global concept drift has occurred; among which Set a drift tolerance multiplier, for example, 3.0. Then, trigger an offline asynchronous retraining process, overwriting the oldest data with the latest collected data stream containing the newest patterns, reinitializing and retraining the model parameters. .

[0093] In this embodiment 1, the system implements seamless hot-plugging of resources through the Kubernetes underlying interface, and uses the actual end-to-end latency of the system after allocation as the key feedback signal to construct an online incremental feedback loss function that integrates the expected and actual deviations in physical queuing. And combined with the limited base learning rate Fine-tuning the model to form an adaptive closed loop of prediction-execution-observation-self-healing not only ensures the model can safely and continuously evolve online when facing unknown burst traffic characteristics, avoiding catastrophic forgetting, but also distinguishes between regular load jitter and true prediction failure, applying gradient penalties only when the system's actual latency deteriorates, thereby maintaining the long-term high reliability and robustness of the resource allocation strategy. Specifically, for example, when the streaming microservice cluster launched a new video interactive function, it generated high CPU and high network coupling traffic characteristics never seen in the offline training data. In the first control cycle, the deep learning model... The predicted load is conservative, resulting in the gateway actually detecting the average end-to-end latency after resource allocation. The latency spiked to 150 milliseconds, far exceeding the expected queuing latency calculated based on the current resource redundancy (assumed to be 50 milliseconds); at this point, the online feedback mechanism was immediately triggered, and the loss function... The system captures the difference within 100 milliseconds and generates a targeted backward penalty gradient; simultaneously, it calculates the constrained base learning rate online. The model parameters are fine-tuned. In the next scheduling cycle, when faced with the same interactive traffic, the model immediately corrects itself and outputs a prediction result that increases the resource margin, so that the actual response delay quickly falls back to the safe range, completes the system's self-correction and performance recovery, and improves the response efficiency of resource allocation.

[0094] Example 2 As a second embodiment of the present invention, such as Figure 3 As shown, based on Embodiment 1, this embodiment also discloses a server resource dynamic allocation control system based on load prediction. This system adopts the server resource dynamic allocation control method based on load prediction described in Embodiment 1, and specifically includes the following modules: The data acquisition and preprocessing module is used to perform multi-dimensional server physical status data acquisition and preprocessing and construct a smooth multi-dimensional resource status matrix. In practice, this module deploys a probe component based on the system kernel layer, capable of assessing comprehensive load fluctuation indicators. Adaptive adjustment of sampling frequency It collects physical status data such as server load (CPU, memory, and network bandwidth); simultaneously, the module has a built-in circuit breaker mechanism to ensure the security of the monitoring probe: when the host machine's CPU utilization is affected by the measurement probe's own system load... The circuit breaker threshold exceeds the preset probe overhead limit. Time-triggered frequency reduction protection; and employs a dynamic smoothing weighting coefficient-based approach. The sliding window low-pass filter removes transient hardware sampling noise, outputting a smooth multidimensional resource state matrix. This module breaks away from the rigid mode of traditional fixed-frequency monitoring, avoiding resource waste during stable periods and ensuring the security of monitoring probes.

[0095] The coupling analysis module is used to perform cross-resource coupling trend analysis and dynamically update the cross-resource coupling coefficients. In practice, this module communicates with the microservice API gateway, extracts the actual physical time difference between different physical resource consumption actions by parsing the call chain tracing logs, and calculates its statistical expected value as a causal delay constant. The 99th percentile of the end-to-end response time of the microservice cluster in the API gateway over the past business scheduling cycle is extracted as the sliding time window for data sampling. Calculate and output the load change rate of the source resources. The mapping and projection relationship of downstream resource load, i.e., the cross-resource coupling coefficient. This module can detect and quantify the potential incremental pressure on downstream resources caused by a surge in source load in advance, eliminating the prediction lag caused by the fragmentation of resource states in large-scale microservice architectures, thereby reducing response latency when dealing with sudden traffic surges.

[0096] The load forecasting module is used to predict loads using machine learning models. Output the next time step by combining real-time differential trend. Predicted target resource load reach rate In practical implementation, this module uses a spatiotemporal graph convolutional neural network with a global attention mechanism as its core architecture, and receives the cross-resource coupling coefficients output by the coupling analysis module. The constructed dynamic directed adjacency matrix It also receives the smoothed resource state matrix output by the data acquisition and preprocessing module. Through the asymmetric penalty loss function Guided parameter space, calculate and output the target resource load arrival rate that takes into account both physical causal constraints and complex nonlinear residual compensation. This module enables the system to respond instantly to sudden traffic surges while retaining its advantage of accurately capturing long-term implicit characteristics, reducing response latency, and avoiding blind over-provisioning based on experience, thus improving resource utilization.

[0097] Find an allocation module to construct a cost-sensitive objective function. And solve for the resource capacity that will ultimately be delivered to the cluster. In practical implementation, this module receives the target resource load arrival rate output by the load prediction module. And comprehensively incorporate resource supply cost coefficients With SLA delay penalty coefficient Construct a cost-sensitive objective function The problem of allocating physical resources is transformed into a mathematical optimization problem that balances economic operating costs and service quality. Through a built-in solver, under the constraints of equal normal operating costs and avoiding extreme risks, the system calculates the resource capacity that each node ultimately decides to deliver to the cluster. This module improves resource utilization by dynamically determining the optimal redundancy ratio between resource idle costs and SLA penalties.

[0098] The feedback adaptive module is used to perform dynamic hot-plugging of resources and update the machine learning model online based on the actual server response latency data after the changes take effect. Parameters. In practice, this module utilizes the underlying container orchestration interface to deliver the final decision output by the optimization allocation module to the cluster's resource capacity. The data is distributed to the server cluster, and the module continuously collects the actual average end-to-end latency after the data takes effect. The expected delay is consistent with the theoretical predictions derived from queuing theory. Perform deviation comparison and construct an online incremental feedback loss function. Based on the set base learning rate This triggers backpropagation and online fine-tuning of the network parameters within the load prediction module, forming an adaptive control closed loop. This module ensures that the model can safely and continuously evolve online when faced with unknown burst traffic characteristics, avoiding catastrophic forgetting. It can also distinguish between regular load jitter and true prediction failure, applying gradient penalties only when the actual system latency deteriorates, thereby maintaining the long-term high reliability and robustness of the resource allocation strategy.

[0099] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0100] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A load prediction based dynamic allocation control method of server resources, characterized in that, include: Collect physical status data of the server cluster and perform sliding window low-pass filtering to construct a smooth multidimensional resource status matrix; The load change rate of each resource is calculated based on the multidimensional resource state matrix, and the cross-resource coupling coefficient is dynamically updated to quantify the causal delay and mapping relationship caused by the surge in source request resources to downstream computing and storage resources. A machine learning model is constructed and combined with the cross-resource coupling coefficient and the load change rate of each resource to predict the target resource load arrival rate at the next moment. Based on the target resource load arrival rate at the next moment, a cost-sensitive objective function is constructed and solved to obtain the optimal resource allocation for execution. The system performs dynamic hot-plugging and feedback correction of resources. Based on the optimal resource allocation, it distributes the final resource capacity to the cluster to the cluster and collects and compares the actual response latency data after it takes effect. It then performs online incremental updates to the network parameters of the machine learning model to form an adaptive control closed loop.

2. The server resource dynamic allocation control method based on load prediction according to claim 1, characterized in that, The sampling frequency of the multidimensional physical state data is determined using an adaptive sampling strategy, specifically including: a comprehensive load fluctuation index based on the previous time window. Adaptively calculate the desired sampling frequency at the current moment. Real-time measurement of the host machine CPU utilization consumed by the probe itself. ,when When the preset probe overhead limit is exceeded, the circuit breaker mechanism is triggered, and the actual sampling frequency is reduced to a security-degraded frequency. Otherwise, the sampling frequency of the multidimensional physical state data shall be the desired sampling frequency. .

3. The server resource dynamic allocation control method based on load prediction according to claim 1, characterized in that, The dynamic updating cross-resource coupling coefficient specifically comprises: tracking a call chain topology graph of a micro-service API gateway, extracting a real physical time difference between resource consumption events, and calculating a statistical expectation value as a causal delay constant ; extracting a 99th percentile of an end-to-end response time of the micro-service cluster in a past business scheduling period as a sliding time window of data sampling ; within the sliding time window , using a sample covariance with a delay offset and a sample variance to calculate a projection mapping relationship of a source resource load to a target resource load, to obtain a dynamic cross-resource coupling coefficient , and the expression is: ； wherein is the target resource at the current time , and is the smoothed load state of the relevant resource before experiencing the causal delay , and and represent the sample covariance and sample variance, respectively.

4. The method of claim 3, wherein the load prediction-based dynamic allocation control of server resources is characterized by, The offline training process of the machine learning model specifically includes: A spatio-temporal graph convolutional neural network comprising an attention mechanism is constructed as a core architecture, and the cross-resource coupling coefficients are utilized A dynamic directed adjacency matrix with physical causal relationships is constructed ; inputting the smoothed multi-dimensional resource state sequence frame by frame into a graph convolution network, in combination with the dynamic adjacency matrix extracting multi-scale spatial graph convolution features, and using a nonlinear activation function to force truncation of negative values to ensure that resource features always maintain a non-negative physical meaning in the hidden space; The sequence containing the topological coupling information after the spatial feature extraction is input into a long short-term memory network to capture the long-term periodicity and short-term burst dependence of the load fluctuation over time, and a hidden state of the cumulative spatio-temporal joint evolution law is obtained ; The hidden states are aggregated using a global attention mechanism to obtain a global spatio-temporal context vector which is then reduced in dimension and mapped back to the real physical dimension space by a fully connected layer to output a prediction term for residual compensation ; An asymmetric penalty loss function is adopted The network model parameters are trained by back propagation, gradient penalty is applied to the case that the predicted value is lower than the real demand, and the model is forced to learn the physical safety bottom line.

5. The method of claim 4, wherein the load prediction-based dynamic allocation control of server resources is characterized by, The formula for calculating the target resource load arrival rate at the next moment is: ； wherein, is the predicted class resource at the load arrival rate at the time instant, is the current time instant the actual smoothed load arrival rate of the class resource, is the load change rate of the associated resource , is the time step, is the nonlinear residual prediction term output by the spatio-temporal graph convolutional network, respectively represent the resource types of CPU, memory, and bandwidth. 6.The load prediction based server resource dynamic allocation control method of claim 5, wherein, The expression for the cost-sensitive objective function is: ； wherein, is a cost-sensitive objective function, is the system state at time is the final decision of the resource capacity delivered to the cluster, is the resource supply cost coefficient, is the SLA delay penalty coefficient, represents CPU, memory, and network bandwidth.

7. The server resource dynamic allocation control method based on load prediction according to claim 6, characterized in that, The optimal resource allocation amount for execution is obtained, specifically comprising: when the optimization problem can be completely decomposed according to resources, a closed-form optimal allocation amount independently solved by a single resource item is obtained by derivation and setting the derivative to zero When there is a cross-resource coupling constraint, the closed-form optimal allocation amount As an initial guess, input to the real-time optimizer is numerically solved by using an interior point method, and a projection iteration is performed on the constraint set to obtain an available solution.

8. The server resource dynamic allocation control method based on load prediction according to claim 1, characterized in that, The process of collecting and comparing actual response latency data after the system takes effect, and then incrementally updating the network parameters of the machine learning model online, specifically includes: Calculate the theoretical expected delay based on a queuing-theoretic continuous approximation model under the current physical resource allocation state. ; Collect the actual average end-to-end delay after it takes effect. We construct an online incremental feedback loss function that integrates the error of pure numerical prediction with the penalty for the deterioration of macroscopic queuing delay in the system. ; The online incremental feedback loss function is calculated using an adaptive learning rate optimization algorithm. About machine learning models The gradient of the current network parameters is used to perform parameter fine-tuning and updates.

9. The server resource dynamic allocation control method based on load prediction according to claim 2, characterized in that, The desired sampling frequency The calculation formula is: ； In the formula, To provide a comprehensive load fluctuation index, To ensure a minimum sampling frequency during the stable period, This is the highest sampling frequency during the outbreak period. This represents the fluctuation sensitivity coefficient.

10. A server resource dynamic allocation control system based on load prediction, employing the server resource dynamic allocation control method based on load prediction as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and preprocessing module is used to perform multi-dimensional server physical status data acquisition and preprocessing and construct a smooth multi-dimensional resource status matrix. The coupling analysis module is used to perform cross-resource coupling trend analysis and dynamically update the cross-resource coupling coefficients. The load forecasting module is used to predict the target resource load arrival rate at the next moment by combining a machine learning model with cross-resource coupling coefficients. The optimal allocation module is used to construct a cost-sensitive objective function and solve for the optimal resource allocation. The feedback adaptive module is used to perform dynamic hot-plugging of resources and update model parameters online by comparing latency data.

Images