Server resource dynamic prediction and scheduling system and method based on long short-term memory network

Abstract

The invention relates to the technical field of server resource management, and provides a server resource dynamic prediction and scheduling system and method based on a long short-term memory network in order to solve the problems of insufficient prediction capability, single scheduling strategy, low energy efficiency and the like in existing server resource management. The system adopts a layered architecture and comprises a data acquisition layer, a prediction analysis layer and a scheduling execution layer, wherein the prediction analysis layer adopts an innovative double-layer LSTM network structure, integrates a short-term prediction network and a long-term prediction network, and realizes accurate prediction of server resource use through a time embedding feature and an attention mechanism; the scheduling execution layer realizes a multi-objective optimization decision mechanism and a progressive scheduling strategy combining service performance, energy efficiency and operation and maintenance cost. According to the method, the resource prediction accuracy and the resource utilization rate are remarkably improved, the energy consumption is reduced, the system operation and maintenance efficiency is improved, and the method is suitable for intelligent resource management of a large-scale server cluster.

Description

Technical Field

[0001] The present invention relates to the technical field of server resource management, and particularly to a dynamic prediction and scheduling system and method for server resources based on deep learning, which is applicable to the intelligent resource management of large-scale server clusters. Background Art

[0002] With the rapid development of cloud computing and large-scale distributed systems, the efficient management of data center server resources has become an important technical challenge. Traditional server resource management methods mainly rely on fixed-threshold triggering mechanisms and simple rule engines. When facing complex and changing workloads, such methods often cannot adjust resource allocation in a timely and accurate manner, resulting in low server resource utilization or degraded service quality.

[0003] Currently, the industry mainly adopts the following technical solutions for server resource management: 1. Traditional methods based on threshold triggering; 2. Prediction methods based on time series analysis; 3. Scheduling strategies based on simple machine learning; 4. The status quo of commercial solutions.

[0004] Traditional methods based on threshold triggering set fixed thresholds for metrics such as CPU usage and memory occupancy. When the monitored metrics exceed the thresholds, corresponding resource scheduling operations are triggered. For example, when the CPU usage exceeds 80%, computing resources are increased, and when the memory usage is lower than 30%, resources are recycled. This method is simple to implement, but has the following problems: a. The threshold setting lacks flexibility and is difficult to adapt to dynamic business scenarios; b. It is impossible to predict the changing trend of resource requirements and can only respond passively; c. It may lead to oscillation phenomena in resource allocation, affecting system stability.

[0005] Prediction methods based on time series analysis use traditional time series analysis techniques, such as ARIMA, exponential smoothing and other algorithms to predict the usage of server resources. Although it has improved compared to the pure threshold-triggering scheme, there are still obvious deficiencies: a. The prediction model is too simple to capture complex resource usage patterns; b. It is unable to effectively handle the mutual correlation between multi-dimensional resource metrics; c. Its prediction ability for sudden load changes is limited.

[0006] In recent years, some studies have begun to attempt to apply machine learning to server resource management, such as using support vector machines (SVMs) or simple neural networks for resource usage prediction. Although such methods have achieved certain results in some scenarios, there are still the following limitations: a. The prediction model only focuses on a single time scale and is difficult to simultaneously consider short-term and long-term resource demand predictions; b. There is a lack of an evaluation mechanism for the credibility of prediction results; c. The scheduling strategy is relatively simple and fails to fully utilize prediction information for multi-objective optimization.

[0007] Current commercial solutions on the market, such as the auto-scaling service of a certain cloud service provider, mainly use a rule engine combined with a simple prediction model for resource scheduling. Although these solutions have been applied in the actual production environment, there are still the following problems: a. The prediction algorithm is relatively simple and the accuracy is limited; b. Lack of fine-grained resource scheduling ability; c. Insufficient consideration for energy efficiency optimization; d. Limited system scalability.

[0008] By analyzing the existing technical solutions, we can see that the following technical bottlenecks mainly exist in the current server resource management field: 1. Insufficient prediction ability: It is difficult for existing solutions to accurately grasp the dynamic change law of resource requirements; 2. Single scheduling strategy: Lack of a comprehensive scheduling solution considering multi-objective optimization; 3. Limited system scalability: It is difficult to meet the management requirements of large-scale server clusters; 4. Low energy efficiency: Insufficient consideration for optimizing the energy consumption of servers.

[0009] In response to the above problems, a new technical solution is urgently needed, which can make full use of the advantages of deep learning technology to achieve more intelligent and efficient server resource management. Based on this background, the present invention proposes an innovative deep learning-based server resource dynamic prediction and scheduling system. Summary of the Invention

[0010] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a server resource dynamic prediction and scheduling system and method based on a long short-term memory network. This system can accurately predict the server resource usage situation and achieve intelligent and automated resource scheduling, improve the server resource utilization rate, reduce energy consumption, and optimize the system performance.

[0011] To achieve the above purpose, the technical solution provided by the present invention is: A server resource dynamic prediction and scheduling system based on a long short-term memory network, including a data collection layer, a prediction analysis layer, and a scheduling execution layer.

[0012] The data collection layer adopts a distributed architecture and realizes omni-directional data collection by deploying lightweight collection agents (Agents) on each managed server. Each Agent includes a data collection unit, a data processing unit, and a communication management unit; among them, the data collection unit is responsible for collecting system resource metrics, application layer metrics, and hardware status metrics; the data processing unit is responsible for preprocessing the collected data, including data cleaning, data normalization, and time feature extraction; the communication management unit is responsible for data transmission and management.

[0013] The prediction and analysis layer adopts an innovative double-layer LSTM network structure, including a short-term prediction network and a long-term prediction network. The short-term prediction network is mainly responsible for predicting resource usage within 5 - 15 minutes. Its structure includes an input layer, a bidirectional LSTM layer, an attention layer, and an output layer. The long-term prediction network is mainly responsible for predicting resource usage within 1 - 4 hours. Its structure includes an input layer, a stacked LSTM layer, a temporal attention layer, and an output layer. The system also designs an adaptive fusion mechanism for prediction results to achieve the optimal combination of short-term and long-term prediction results through learnable fusion parameters.

[0014] The scheduling and execution layer implements a resource scheduling decision-making mechanism for multi-objective optimization and a progressive scheduling strategy. The multi-objective optimization decision-making mechanism comprehensively considers three dimensions: service performance, energy efficiency, and operation and maintenance costs, and achieves the balance of objectives through dynamic weight allocation. The progressive scheduling strategy includes a safety threshold control and a hierarchical scheduling mechanism. The former dynamically sets the safety threshold for resource adjustment according to the prediction credibility, and the latter divides resource scheduling into three levels: emergency scheduling, regular scheduling, and optimization scheduling. The system also designs a complete scheduling execution and feedback control process, including functions such as scheduling command generation, execution monitoring and feedback, and self-learning optimization.

[0015] According to one aspect of the present invention, there is provided a method for dynamically predicting and scheduling server resources based on the above system, including the following steps: 1. Collect multi-dimensional monitoring data; 2. Preprocess and extract features from the data; 3. Use a double-layer LSTM network for resource usage prediction; 4. Execute a resource scheduling strategy based on the prediction results.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Improvement in resource prediction ability: Through the innovative design of the double-layer LSTM network structure, the system significantly improves the prediction accuracy and realizes the transformation from a single resource metric to multi-dimensional collaborative prediction; 2. Optimization of system operation efficiency: Based on accurate prediction results and a multi-objective optimization scheduling strategy, the system significantly improves resource utilization and improves system performance; 3. Optimization of energy efficiency: Through intelligent resource scheduling, the system effectively reduces the overall energy consumption of the server, optimizes the heat dissipation effect, and reduces the cooling cost; 4. Improvement in operation and maintenance efficiency: Through an automated prediction and scheduling mechanism, the system greatly reduces the need for manual intervention, improves the automated processing ability, and enhances the fault handling ability; 5. Improvement in system scalability: The hierarchical architecture design of the system improves the system's ability to support the management of large-scale server clusters and has good functional scalability. Description of the Drawings

[0017] Figure 1 is the overall system architecture diagram of the present invention. Figure 2 is the schematic diagram of the data acquisition layer structure of the present invention. Figure 3It is the flowchart of the data processing unit of the present invention. Figure 4 It is the structure diagram of the double-layer LSTM network of the present invention. Figure 5 It is the function module diagram of the scheduling and execution layer of the present invention. Specific implementation manners

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining and illustrating the present invention, and are not used to limit the present invention.

[0019] As Figure 1 shown, the present invention provides a server resource dynamic prediction and scheduling system based on a long short-term memory network. The system adopts a hierarchical architecture design, mainly including three core levels: a data collection layer, a prediction and analysis layer, and a scheduling and execution layer. This hierarchical design can achieve modular management of system functions and improve the scalability and maintainability of the system.

[0020] As Figure 2 shown, the data collection layer adopts a distributed architecture, and realizes all-round data collection by deploying lightweight collection agents (Agents) on each managed server. Each Agent includes three core components: a data collection unit, a data processing unit, and a communication management unit.

[0021] The data collection unit is responsible for real-time collection of multi-dimensional metrics, specifically including: (1) System resource metric collection: CPU usage rate, collecting the overall usage rate and the specific usage of each core; memory usage, including physical memory usage, virtual memory usage, page swap rate; disk I / O status, monitoring read and write speeds, I / O wait time, disk queue length; network traffic, recording inbound traffic, outbound traffic, network latency, packet loss rate; system load, collecting the average load values for 1 minute, 5 minutes, and 15 minutes. (2) Application layer metric collection: Request processing ability, including concurrent request number, request queue length; service response time, recording the average response time and response time distribution; transaction processing situation, counting transaction success rate, transaction processing volume; application error rate, monitoring system error rate and business error rate. (3) Hardware status metric collection: Temperature monitoring, real-time collecting CPU temperature and motherboard temperature; power consumption monitoring, recording the overall power consumption and CPU power consumption; hardware health status, monitoring fan speed and power supply status.

[0022] As Figure 3As shown, the data processing unit implements a multi - level data processing pipeline, mainly including the following processing steps: (1) Data pre - processing: Data cleaning, removing outliers and handling missing values; Data standardization, unifying metrics with different dimensions to the same scale; Time feature extraction, extracting time - periodic features, such as periodicity at the hour, day, week levels. (2) Data compression: Adopting an adaptive compression strategy, dynamically selecting appropriate compression algorithms according to data characteristics: For numerical data, using differential coding combined with lossless compression; For state - type data, using run - length encoding; For text - type data, using dictionary compression.

[0023] The prediction and analysis layer is the core innovative part of the present invention, realizing multi - dimensional resource usage prediction based on deep learning. As Figure 4 shown, the present invention designs an innovative double - layer LSTM network structure, including: (1) Short - term prediction network: Input layer, receiving standardized multi - dimensional monitoring metrics; LSTM layer, adopting a bidirectional LSTM structure to extract short - term time - series features; Attention layer, realizing dynamic weight allocation for features at different time steps; Output layer, predicting the resource usage in the next 5 - 15 minutes. (2) Long - term prediction network: The long - term prediction network is mainly responsible for predicting the resource usage in 1 - 4 hours, and its structure is designed as follows: Input layer, in addition to receiving standardized multi - dimensional monitoring metrics, also introducing time - embedding features; Deep LSTM, adopting a three - layer stacked LSTM structure to enhance the ability to capture long - term dependencies; Time attention layer, realizing adaptive weighting of features at different time scales; Output layer, predicting the resource usage trend in the next 1 - 4 hours.

[0024] In the present invention, the time - embedding vector is calculated using the following method: TE(pos, 2i) = sin(pos / 10000 ^ (2i / d)), TE(pos, 2i + 1) = cos(pos / 10000 ^ (2i / d)). Where pos represents the time position, i represents the dimension index, and d represents the embedding dimension. This encoding method can effectively capture the periodic features of the time series.

[0025] The present invention designs different attention mechanisms in the short - term and long - term prediction networks respectively. For time t, the attention weight is calculated as follows: α(t, s) = softmax(e(t, s)), e(t, s) = v^T * tanh(W_1 * h(t)+W_2 * h(s)+b). Where: α(t, s) represents the attention weight of time t to time s; h(t) and h(s) represent the hidden states at the current time and historical time respectively; W_1, W_2 are learnable weight matrices; v and b are learnable parameter vectors.

[0026] To comprehensively utilize the results of short-term and long-term predictions, the present invention designs an adaptive fusion mechanism. The calculation of the fusion weight β is as follows: β = sigmoid(W_f * [h_s; h_l] + b_f), y_final = β * y_short + (1 - β) * y_long. Where: h_s and h_l respectively represent the hidden states of the short-term and long-term networks; y_short and y_long respectively represent the short-term and long-term prediction results; W_f and b_f are learnable fusion parameters.

[0027] The present invention adopts an innovative training strategy to improve the model performance: (1) Phase-based training: In the first phase, the short-term and long-term prediction networks are trained separately, L_phase1 = L_short + L_long; in the second phase, the fusion network is trained while fixing the parameters of the short-term and long-term networks, L_phase2 = L_total. (2) Dynamic sampling strategy: To address the problem of sample imbalance, the present invention designs a dynamic sampling mechanism based on the fluctuation of resource utilization rate: p(x) = softmax(λ * σ(x)). Where: p(x) represents the probability that sample x is selected; σ(x) represents the standard deviation of the resource utilization rate within the time window corresponding to the sample; λ is the sampling temperature parameter. Through the above design, the double-layer LSTM network of the present invention can effectively integrate the short-term and long-term prediction capabilities and achieve accurate prediction of server resource usage.

[0028] The scheduling and execution layer is the key execution unit of the present invention, responsible for formulating and executing resource scheduling policies according to the prediction results. As Figure 5 shown, the present invention designs a resource scheduling decision-making mechanism considering multiple objectives. This mechanism comprehensively considers three dimensions of service performance, energy efficiency, and operation and maintenance costs, and achieves the balance of objectives through dynamic weight allocation. The specific innovation points include: (1) Performance objective: The present invention quantifies service performance into three key indicators of response time, throughput, and resource utilization rate. Based on the predicted resource usage values provided by the prediction model and combined with historical data analysis, the system can set reasonable performance target thresholds for each service. When it is detected that the performance indicators may violate the thresholds, the system will trigger corresponding resource adjustment strategies. (2) Energy efficiency objective: An energy efficiency evaluation model is innovatively introduced. By analyzing the power consumption characteristic curve of the server and combining the workload prediction, the optimal resource allocation scheme is calculated. This model considers the energy efficiency performance of the server at different load levels and realizes the optimization of energy consumption. (3) Operation and maintenance cost objective: An operation and maintenance cost evaluation framework is designed, which converts the impact of resource scheduling on system stability into quantifiable cost indicators. This framework considers factors such as resource migration overhead and system fluctuation risk to ensure the economy of scheduling decisions.

[0029] To avoid drastic impacts on the system caused by resource adjustment, the present invention innovatively proposes a progressive scheduling strategy: (1) Safety threshold control: Dynamically set the safety threshold for resource adjustment according to the credibility of the prediction results. The higher the credibility, the greater the allowed adjustment range; when the credibility is low, the system will adopt a more conservative adjustment strategy. (2) Hierarchical scheduling mechanism: Design a three-level scheduling mechanism: Emergency scheduling, which is used to handle sudden performance anomalies, can respond quickly but has limited adjustment range; Regular scheduling, which conducts planned resource adjustment based on the prediction results and adopts a gentle adjustment strategy; Optimization scheduling, which conducts resource optimization during the system's stable period and mainly focuses on improving energy efficiency.

[0030] The present invention realizes a complete scheduling execution and feedback control process: (1) Scheduling command generation: Automatically generate a standardized scheduling command based on the decision result, including: Resource allocation instruction, specifying the type of resource to be adjusted and the target value; Execution timing, detailed scheduling steps and time arrangements; Rollback strategy, defining the coping plan in case of scheduling failure. (2) Execution monitoring and feedback: Design a real-time monitoring and feedback mechanism, including: Execution progress tracking, monitoring the execution status of the scheduling command in real time; Effect evaluation, analyzing the actual impact of the scheduling operation on the system performance; Exception handling, detecting and handling exceptions during the scheduling process. (3) Self-learning optimization: Innovatively introduce a self-learning optimization mechanism for the scheduling strategy: Record the execution effect of each scheduling operation; Analyze the key factors affecting the scheduling effect; Continuously optimize the parameter configuration of the scheduling strategy. Through the above design, the scheduling execution layer of the present invention realizes the intelligent and automated scheduling of resources, ensuring both the safety of the scheduling operation and the optimization of resource utilization.

[0031] The system deployment of the present invention needs to meet the following conditions: (1) Hardware environment requirements: Computing nodes: The CPU is not lower than the Intel Xeon E5-2600v4 series, the memory of each node is not less than 128GB, and the network is interconnected by a 10G network; Storage system: A high-performance distributed storage system with a storage capacity of not less than 50TB and an IOPS of not less than 100K. (2) Software environment requirements: The operating system is Linux kernel version 4.18 or higher, the container platform is Kubernetes 1.18 or higher, and the database system is a time-series database (such as InfluxDB, TimescaleDB, etc.).

[0032] The server resource dynamic prediction and scheduling system based on long short-term memory network proposed by the present invention can effectively improve the intelligent and automated level of server resource management, significantly improve the system performance, reduce energy consumption, improve the operation and maintenance efficiency, and provide strong technical support for the efficient management of large-scale server clusters.

[0033] Those skilled in the art should understand that the present invention is not limited to the exact configuration of the above embodiments, and various modifications and extensions can be made. For example: 1. The prediction model can be extended to other deep learning models, such as Transformer, GRU, etc., to adapt to different application scenarios; 2. The scheduling strategy can be customized according to specific business requirements, adding more decision dimensions and optimization goals; 3. The system can be deeply integrated with existing container orchestration platforms (such as Kubernetes) to enhance the resource management ability in the container environment; 4. A federated learning mechanism can be introduced to achieve collaborative training and optimization of models between multiple data centers. These modifications and extensions are all within the protection scope of the present invention.

[0034] Through the innovative double-layer LSTM network structure and the scheduling strategy of multi-objective optimization, the present invention effectively solves the problems of insufficient prediction ability, single scheduling strategy, limited system scalability, and low energy efficiency in server resource management, providing a comprehensive solution for the intelligent resource management of large-scale server clusters.

Claims

1. A server resource dynamic prediction and scheduling system based on a long short-term memory network, characterized in that Including: The data acquisition layer adopts a distributed architecture and realizes multi-dimensional data acquisition by deploying lightweight acquisition agents (Agents) on the managed servers. The Agent includes a data acquisition unit, a data processing unit, and a communication management unit; The prediction and analysis layer adopts a double-layer LSTM network structure, including a short-term prediction network and a long-term prediction network. The short-term prediction network is used for resource usage prediction within 5 - 15 minutes, the long-term prediction network is used for resource usage prediction within 1 - 4 hours, and a mechanism for realizing adaptive fusion of prediction results; The scheduling and execution layer realizes a resource scheduling decision mechanism for multi-objective optimization and a progressive scheduling strategy. The multi-objective optimization decision mechanism comprehensively considers three dimensions: service performance, energy efficiency, and operation and maintenance costs. The progressive scheduling strategy includes a safety threshold control and a hierarchical scheduling mechanism.

2. The system according to claim 1, wherein The data acquisition unit is responsible for acquiring the following multi-dimensional metrics: System resource metrics, including CPU usage rate, memory usage, disk I / O status, network traffic, and system load; Application layer metrics, including request processing ability, service response time, transaction processing situation, and application error rate; Hardware status metrics, including temperature monitoring, power consumption monitoring, and hardware health status.

3. The system according to claim 1, characterized in that The data processing unit realizes a multi-level data processing pipeline, including: Data preprocessing, realizing data cleaning, data standardization, and time feature extraction; Data compression, dynamically selecting compression algorithms according to data characteristics, including using differential coding combined with lossless compression for numerical data, run-length encoding for status data, and dictionary compression for text data.

4. The system according to claim 1, wherein The structure of the long-term prediction network includes: The input layer, in addition to receiving standardized multi-dimensional monitoring metrics, also introduces time embedding features. The calculation formula for the time embedding vector TE is: TE(pos,2i) = sin(pos / 10000 ^ (2i / d)); TE(pos,2i + 1) = cos(pos / 10000 ^ (2i / d)); where pos represents the time position, i represents the dimension index, and d represents the embedding dimension; Deep LSTM, adopting a three-layer stacked LSTM structure to enhance the ability to capture long-term dependencies; The time attention layer realizes adaptive weighting of features at different time scales; The output layer is used to predict the resource usage trend in the next 1 - 4 hours.

5. The system according to claim 1, wherein The prediction result adaptive fusion mechanism is realized in the following way: The calculation formula for the fusion weight β is: β = sigmoid(W_f * [h_s; h_l] + b_f); The calculation formula for the final prediction result y_final is: y_final = β * y_short + (1 - β) * y_long; where h_s and h_l respectively represent the hidden states of the short-term and long-term networks, y_short and y_long respectively represent the short-term and long-term prediction results, and W_f and b_f are learnable fusion parameters.

6. The system according to claim 1, characterized in that, In the progressive scheduling strategy: The safety threshold control dynamically sets the safety threshold for resource adjustment according to the credibility of the prediction results. The higher the credibility, the greater the allowable adjustment range. When the credibility is low, a more conservative adjustment strategy is adopted. The hierarchical scheduling mechanism includes three levels: emergency scheduling, regular scheduling, and optimization scheduling. Among them, emergency scheduling is used to handle sudden performance anomalies, regular scheduling performs planned resource adjustments based on prediction results, and optimization scheduling optimizes resources during the system stable period.

7. A method for dynamically predicting and scheduling server resources based on the system according to any one of claims 1-8, characterized in that, It includes the following steps: Collect multi-dimensional monitoring data through the Agent deployed on the server, including system resource metrics, application layer metrics, and hardware status metrics. Preprocess the collected data, including data cleaning, data standardization, and time feature extraction. Use a double-layer LSTM network to predict resource usage for the preprocessed data, including short-term prediction, long-term prediction, and prediction result fusion. Execute the resource scheduling strategy based on the prediction results, including multi-objective optimization decision-making and progressive resource adjustment.

8. The method according to claim 9, characterized in that, The specific implementation of the prediction step includes: Predict the resource usage in the next 5 - 15 minutes through the short-term prediction network, which adopts a bidirectional LSTM structure and an attention mechanism. Predict the resource usage trend in the next 1 - 4 hours through the long-term prediction network, which introduces time embedding features and adopts a stacked LSTM structure. Adopt an adaptive fusion mechanism to combine the short-term and long-term prediction results, and the fusion weights are dynamically adjusted through learnable parameters.

9. The method according to claim 9, wherein The training strategy of the LSTM network includes: Phased training. In the first phase, the short-term and long-term prediction networks are trained separately, and in the second phase, the fusion network is trained. Dynamic sampling strategy, adopting a dynamic sampling mechanism based on the fluctuation of resource utilization rate: p(x) = softmax(λ * σ(x)), where p(x) represents the probability that sample x is selected, σ(x) represents the standard deviation of the resource utilization rate within the time window corresponding to the sample, and λ is the sampling temperature parameter.

10. The method according to claim 9, characterized in that, The execution steps of the scheduling strategy include: Make scheduling decisions based on the prediction results and the multi-objective optimization mechanism, comprehensively weighing service performance, energy efficiency, and operation and maintenance costs. Adopt a progressive strategy to execute resource adjustment, dynamically set the adjustment range according to the prediction credibility, and process it according to the hierarchical scheduling mechanism. Implement the generation and execution of scheduling commands, including formulating resource allocation instructions, execution timings, and rollback strategies. Evaluate the scheduling effect through real-time monitoring and feedback mechanisms, and continuously optimize the scheduling strategy parameters.

Images