Artificial intelligence based full-stack application systematic tuning method

Abstract

The application relates to an artificial intelligence-based full-stack application systematic optimization method, which comprises the following steps: realizing full-link data collection through an intelligent probe; data preprocessing; inputting the preprocessed data into a professional model system, identifying memory leakage risks, correlatively analyzing hardware indexes and application performance indexes, and locating root cause problems; generating a dynamic optimization strategy, generating optimization parameters, and injecting the target system; establishing a standardized feedback data collection mechanism, collecting key performance indexes every predetermined time, and evaluating the optimization result by comparing the key performance index fluctuation before and after optimization. The application can realize intelligentization, dynamicization and cross-level collaborative optimization of system performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of application system performance optimization, and particularly relates to a full-stack application system optimization method based on artificial intelligence. BACKGROUND

[0002] In the field of system performance optimization, the existing technical system has multi-dimensional defects, which is difficult to meet the dynamic optimization demand in complex IT environment. These defects mainly reflect in three aspects of the lagging nature of the optimization method, the insufficient adaptability and the limitation of technical coverage, which are specifically shown as follows.

[0003] 1. The inherent pain points of traditional manual optimization

[0004] The traditional manual optimization mode relying on expert experience has significant efficiency bottlenecks. On the one hand, the response lag problem is prominent-when the system performance bottleneck appears, it needs to go through problem diagnosis, technical expert intervention, scheme formulation and other links, and the average response cycle often exceeds the business tolerable threshold; on the other hand, the high dependence on individual experience leads to unstable optimization quality, different technical experts may produce differential judgments on the same performance problem, and the scarcity of high-end technical talents further limits the large-scale application of this mode.

[0005] 2. Technical limitations of non-AI automatic optimization

[0006] The non-AI automatic optimization scheme based on preset rules has realized the automation of part of the process, but it has the fundamental defects of rule solidification and insufficient environmental adaptability. This kind of scheme usually executes optimization operation through static threshold or fixed script, which cannot cope with dynamic scenarios such as elastic scaling in cloud environment and micro-service architecture iteration.

[0007] 3. Covering short board of existing AI optimization scheme

[0008] The current AI-driven optimization technology still has two major deficiencies in practical application:

[0009] Full-stack collaborative optimization is missing: most schemes focus on a single level (such as the application layer or the database layer), and lack the ability to coordinate the scheduling of cross-layer resources.

[0010] Hardware characteristics are not well adapted: the perception ability of the underlying hardware architecture is weak, especially in the NUMA (Non-Uniform Memory Access) architecture environment, the existing scheme has not fully considered the matching relationship between memory node affinity and CPU scheduling, resulting in increased memory access delay in multi-processor systems, and the hardware performance potential cannot be fully released.

[0011] 4. Performance bottleneck of data collection and processing link

[0012] Efficient performance tuning relies on high-quality full-link data support, but the existing technology has obvious short boards in data collection and preprocessing links:

[0013] Limited collection dimensions: Traditional monitoring tools focus on basic indicators (such as CPU utilization and memory occupancy), and lack the ability to collect deep features such as application call chains, database lock waiting, and network data packet flow;

[0014] Insufficient real-time processing: In the face of TB-level / second performance data, existing preprocessing modules often cause data backlog due to insufficient computing resources, and cannot provide real-time training samples for AI models.

[0015] 5. Splitting problem of full-stack layers

[0016] In traditional optimization schemes, the tuning strategies of the application layer, the software and hardware combined layer, and the basic layer show hierarchical splitting characteristics, and each layer independently adjusts parameters (such as optimizing thread pool configuration in the application layer, and adjusting cache strategy in the basic layer), and lacks a "hardware-to-application full-stack coordination" architecture coordination mechanism. This splitting limits the overall performance improvement and cannot achieve global optimal configuration of cross-layer resources.

[0017] 6. Real-time and feedback mechanism lag problem

[0018] The existing scheme has a significant delay in generating tuning strategies, which cannot meet the real-time needs of high-dynamic load scenarios. Due to the lack of a "performance trend prediction" module to predict the system state, the optimization decision lags behind the load change (such as not being able to expand in advance during a sudden traffic peak), resulting in an increase of more than 20% in system response delay. The timeliness of the feedback loop is insufficient, which becomes a key bottleneck for performance optimization. SUMMARY

[0019] To solve the problems existing in the prior art, the present application provides a full-stack application system optimization method based on artificial intelligence, comprising the following steps:

[0020] Step S1: Collect full-link data through intelligent probes;

[0021] Step S2: Preprocess the data collected in step S1;

[0022] Step S3: Input the preprocessed data into a specialized model system, identify memory leak risks, and perform correlation analysis on hardware indicators and application performance indicators to locate root causes;

[0023] The professional model system includes a JVM memory prediction model based on long short-term memory and a performance attribution model fusing NUMA architecture features: the JVM memory prediction model based on long short-term memory is used for training time series prediction ability through historical memory fluctuation data, and identifying memory leakage risks in advance according to preprocessed data; the performance attribution model fusing NUMA architecture features is used for correlating analysis of hardware indicators and application performance indicators, and locating root causes; the hardware indicators include memory access efficiency, cache hit rate and CPU utilization;

[0024] Step S4: integrating the results of step S3, generating a dynamic tuning strategy, generating optimization parameters and injecting the target system;

[0025] Step S5: establishing a standardized feedback data collection mechanism, collecting key performance indicators every predetermined time, and evaluating the tuning results by comparing the key performance indicators before and after tuning.

[0026] The step S1 includes:

[0027] The heap memory usage, memory allocation rate and GC pause time are obtained inside the JVM process through bytecode injection technology;

[0028] A database probe is deployed on the database node to monitor query database performance parameters, and the database performance parameters include throughput and lock waiting time;

[0029] A hardware probe is deployed on the underlying hardware to obtain underlying resource data, and the underlying resource data includes CPU utilization, memory bandwidth, storage device input operation per second, storage device output operation per second, memory access efficiency, cache hit rate and processor performance.

[0030] The data preprocessing includes:

[0031] The quartile distance-based truncation strategy is adopted for abnormal values;

[0032] Time series interpolation is performed on missing values.

[0033] Consistency check is performed on the data to identify and eliminate invalid samples caused by probe failure;

[0034] The data is de-duplicated and feature engineered.

[0035] The data de-duplication processing includes: for high-frequency acquisition index data, a sliding window of a predetermined time is set, MD5 hash values of repeated data in the window are calculated, only the original value collected for the first time is retained, and the rest of the repeated data is marked as redundant and discarded;

[0036] The feature engineering processing includes: normalizing the data, mapping the original value to the interval [0, 1], eliminating the dimension difference, then constructing a sliding window of a predetermined time, extracting the maximum value and variance in the window as key features, and forming a two-dimensional feature vector.

[0037] The training method of the JVM memory prediction model of the long short-term memory comprises:

[0038] Step S31: generating sample data on the operating system through a stress testing tool;

[0039] Step S32: dividing the sample data into a training set, a validation set and a test set according to a predetermined proportion;

[0040] Step S33: performing feature engineering processing on the training set data to obtain a feature vector;

[0041] Step S34: inputting the feature vector obtained in step S33 into the JVM memory prediction model of the long short-term memory for training, constantly adjusting the model parameters, and minimizing the error between the model prediction result and the true value;

[0042] Step S35: verifying the trained JVM memory prediction model of the long short-term memory on the validation set, and adjusting the parameters of the JVM memory prediction model of the long short-term memory according to the verification result;

[0043] Step S36: testing the JVM memory prediction model of the long short-term memory on the test set, and evaluating the prediction accuracy and generalization ability of the JVM memory prediction model of the long short-term memory.

[0044] The training method of the performance attribution model fusing the NUMA architecture features comprises:

[0045] Step S3a: obtaining hardware indicators and application performance indicators, and performing feature engineering processing on the hardware indicators and the application performance indicators to obtain a feature vector;

[0046] Step S3b: inputting the feature vector obtained in step S3a into the performance attribution model fusing the NUMA architecture features for training, and constantly iterating and optimizing the parameters of the model, so that the performance attribution model fusing the NUMA architecture features can accurately find the correlation between the hardware indicators and the application performance indicators;

[0047] Step S3c: using a cross-validation method to evaluate and optimize the performance attribution model fusing the NUMA architecture features, and improving the accuracy and stability.

[0048] The step S4 is dynamically optimized through a dynamic policy library of a hierarchical classification storage architecture, the hierarchical classification storage architecture divides a set of optimization rules according to system abstraction levels, and the set of optimization rules includes application layer policies and basic layer policies.

[0049] The application layer policies focus on business logic optimization, and include cache invalidation time adjustment and connection pool size configuration.

[0050] The basic layer policies are for hardware and operating system level optimization, and cover NUMA memory affinity configuration and CPU core binding.

[0051] The application is suitable for a full-stack IT architecture constructed based on localized hardware (such as Feiteng CPU and Kunpeng CPU), localized operating systems (such as Kylin advanced server operating system), localized middleware and localized databases (such as Dameng database), and can realize intelligent, dynamic and cross-level collaborative optimization of system performance. DETAILED DESCRIPTION

[0052] In order to have a further understanding of the technical solutions and beneficial effects of the application, the technical solutions of the application and the beneficial effects generated thereby will be described in detail below.

[0053] The overall architecture of the application aims to optimize the full-stack localized application system, and constructs a collaborative mechanism including four core modules, and realizes full-link optimization from data collection to policy execution through intelligent closed loop. Each module independently undertakes a special function, and forms an organic whole through a clear data flow, and specifically includes the following steps S1 to S5.

[0054] Step S1: realizing full-link data collection through intelligent probes.

[0055] The intelligent probes of the application undertake the full-link data collection duty, and through technical means such as bytecode injection and kernel state tracking, real-time capture system key indicators, covering each level from web application server, database, network, JVM, tool chain to kernel. Specifically, it includes application layer JVM (Java Virtual Machine) memory state (such as heap memory usage, memory allocation rate, GC pause time), middleware layer connection pool utilization rate, and basic layer NUMA memory access efficiency (such as local memory hit rate, cross-node delay) and other core data. The collection frequency supports dynamic adjustment, and can be improved to microsecond level precision in high load scenarios, ensuring data timeliness and integrity.

[0056] In order to realize full-stack performance data collection, a layered embedded deployment strategy is adopted:

[0057] 1. Application memory layer: Through Java Agent probe technology, the application memory performance is deeply monitored inside JVM. The probe uses bytecode injection capability to implant collection logic in the key execution path of the application, thereby accurately obtaining three core indicators: heap memory usage, memory allocation rate, and GC pause time.

[0058] 2. Database layer: Deploy database probes on database nodes (such as DM8) to monitor transaction throughput and lock waiting time, and other database performance parameters.

[0059] 3. At the bottom layer of hardware (such as CPU, memory, disk controller): Through the combination of hardware performance monitoring unit and intelligent software probe, the CPU utilization, memory bandwidth, IOPS (input / output operations per second of storage device), memory access efficiency, cache hit rate, and processor performance are obtained.

[0060] 4. Application performance layer: Through the application layer extensible Java Agent probe technology, the request throughput, response delay, error rate, and other core indicators directly reflecting user experience are obtained. Step S2: Data preprocessing. Data preprocessing is a key link to ensure the quality of AI model training. Through multi-dimensional cleaning mechanism, the original collected data is processed. For abnormal values, the IQR (interquartile range) based truncation strategy is adopted, the missing values are interpolated and completed by time series, and the dimension difference is eliminated by feature normalization (such as Min-Max scaling). In addition, the module has built-in data consistency verification logic, which can automatically identify and exclude invalid samples caused by probe failure, and finally output standardized training data set and real-time feature vector.

[0061] Data preprocessing is aimed at noise filtering and feature extraction of original collected data, and the main innovative content includes two aspects: de-duplication algorithm and feature engineering.

[0062] De-duplication algorithm: sliding window + hash table comparison mechanism is adopted to realize real-time de-duplication of high-frequency collection indicators (such as CPU utilization, network throughput). The specific implementation is as follows: set a 1-second sliding window, calculate the MD5 hash value of the repeated data in the window, and only keep the original value collected for the first time, and mark the remaining repeated data as redundant and discard. This mechanism can reduce the repetition rate of CPU utilization data by about 20%, effectively reducing the storage and computing resource consumption.

[0063] Feature engineering: according to the input requirements of AI model, the preprocessed data is standardized and feature extracted. Taking memory occupation data as an example, first, the Min-Max normalization formula is used:

[0064] ,

[0065] The original value is mapped to the interval [0, 1] to eliminate dimensional differences; then a 5-minute sliding time window is constructed, and the maximum value (reflecting the peak load) and the variance (reflecting the load fluctuation) in the window are extracted as key features, forming a feature vector with a dimension of 2, providing input for subsequent AI model performance prediction and bottleneck positioning.

[0066] Step S3: AI training and problem analysis

[0067] AI training and problem analysis are based on the construction of a specialized model system based on domain characteristics, which includes two sub-models: one is a JVM (Java Virtual Machine) memory prediction model based on LSTM (Long Short-Term Memory), which trains the time series prediction ability through historical memory fluctuation data to identify memory leak risks in advance; the second is an XGBoost (Large-Scale Parallel Tool) performance attribution model that integrates NUMA architecture features, which analyzes the correlation between memory efficiency, cache hit rate, and application performance indicators (such as response delay) to locate root causes. The training process uses an incremental learning strategy, updating model parameters every 24 hours to adapt to system dynamics.

[0068] 1. JVM memory prediction model based on LSTM (Long Short-Term Memory):

[0069] Feature engineering: Input features include hardware parameters of domestic servers (such as Feiteng S2500 64-core processors, memory bandwidth 25.6GB / s), JVM memory status (such as heap memory usage, memory allocation rate, GC pause time), and application performance indicators (such as the request throughput of TongWeb). These indicator parameters can reflect the running state of the system from different dimensions, hardware parameters reflect the influence of underlying hardware on JVM memory usage, JVM memory status indicators directly reflect the running situation of memory, and application performance indicators reflect the demand situation of the application for memory. Combining these indicator parameters can more accurately predict the trend of JVM memory usage.

[0070] Model architecture: LSTM neural network structure is used, which has long-term memory capability and can better handle long-term dependencies in time series data. For the instruction set characteristics of ARM architecture CPU, the network structure parameters are adjusted, such as setting appropriate hidden layer node number, learning rate, etc. Through 5-fold cross-validation, the hyperparameters are optimized to improve the prediction accuracy and generalization ability of the model.

[0071] Training data source: On the Kirin Advanced Server V10 operating system, 100,000 sample data are generated through stress testing tools (such as the domestic stress testing tool "Kylin PET 7.1"), covering different load scenarios (CPU-intensive, IO-intensive).

[0072] Training steps:

[0073] (1) Divide 100,000 sample data into certain proportions (such as 70% training set, 15% validation set, and 15% test set);

[0074] (2) Perform feature engineering processing on the training set data (the method of feature engineering processing is the same as described in the data preprocessing section above), and extract the key feature vector;

[0075] (3) Input the processed feature vector into the LSTM model for training, and continuously adjust the model parameters to minimize the error between the predicted results of the model and the true values;

[0076] (4) Validate the trained model on the validation set, adjust the model parameters such as learning rate and hidden layer node number according to the validation results, to further improve the performance of the model;

[0077] (5) Finally, test the model on the test set to evaluate the prediction accuracy and generalization ability of the model, combine the time series prediction ability of historical memory fluctuation data with the prediction ability of the LSTM model, use dynamic threshold and multi-feature correlation to identify memory leaks in advance, and assist in root cause analysis to locate problems

[0078] 2. XGBoost performance attribution model with NUMA architecture features:

[0079] Feature engineering: Input features include memory access efficiency, cache hit rate, CPU utilization, and application performance indicators (such as response delay). These feature parameters can reflect the performance of the system from both hardware and application levels, and by analyzing the correlation between them, the root cause of the system performance problem can be accurately found.

[0080] Model architecture: XGBoost algorithm is used, which is a high-efficiency gradient boosting decision tree algorithm with the ability to handle high-dimensional data and complex nonlinear relationships. According to the characteristics of NUMA architecture, the model parameters are adjusted, such as setting appropriate tree depth, learning rate, etc., to improve the attribution ability of the model to the performance problems of the system under NUMA architecture.

[0081] Training data source: Similarly, on the Kirin high-end server V10 operating system, simulate different NUMA architecture scenarios and system load conditions to generate sample data containing various performance problems.

[0082] Training steps:

[0083] (1) Perform feature engineering processing on the generated sample data (the method of feature engineering processing is the same as described in the data preprocessing section above), and extract the key feature vector.

[0084] (2) The processed key feature vector is input into the XGBoost model for training. By continuously iterating and optimizing the parameters of the model, the model can accurately find the correlation between hardware indicators and application performance indicators.

[0085] (3) The model is evaluated and optimized using methods such as cross-validation to improve the accuracy and stability of the model. The hardware indicators such as memory access efficiency and cache hit rate are analyzed in relation to the application performance indicators to locate the root cause of the problem. For example, when the system has a problem of increasing response delay, the model can analyze the relationship between hardware indicators and application performance indicators to determine whether it is caused by memory access delay under NUMA architecture or other reasons.

[0086] Step S4: Intelligent optimization.

[0087] Intelligent optimization as the decision center, integrates AI (artificial intelligence) analysis results to generate dynamic optimization strategies. For JVM (Java Virtual Machine) memory problems, it can automatically adjust the ratio of young generation / old generation, survivor area size and other parameters; in the face of NUMA memory bottleneck, it optimizes the core binding logic through thread allocation algorithm of dual-channel CPU (such as topology-aware load balancing strategy). After the strategy is generated, the module has a built-in gray execution engine, which supports phased strategy deployment and effect verification, ensuring the safety and controllability of optimization actions.

[0088] 1. Dynamic strategy library: adopts hierarchical classification storage architecture, divides the optimization rule set according to the system abstraction level: application layer strategy focuses on business logic optimization, including cache invalidation time adjustment, connection pool size configuration, etc.; basic layer strategy optimizes hardware and operating system level, covering NUMA memory affinity configuration, CPU core binding and other underlying resource scheduling rules. This hierarchical design ensures that the optimization strategy can match the application characteristics and deeply tap the hardware potential.

[0089] 2. Performance trend prediction: based on ARIMA-LSTM hybrid model to build multi-time scale prediction ability, through the fusion of the linear trend capture of statistical model and the nonlinear pattern fitting advantage of deep learning model, to realize the accurate prediction of future 1 hour system load. When the predicted CPU utilization exceeds the 80% threshold, the optimization process is automatically triggered to avoid performance bottlenecks.

[0090] Step S5: execution and feedback: the tuning strategy calls the parameter configuration interface of the full-stack tuning module through the REST API interface, and injects the generated optimization parameters (such as cache invalidation time, NUMA node mapping relationship, etc.) into the target system. To ensure that the tuning effect can be quantitatively evaluated, the system establishes a standardized feedback data collection mechanism: key performance indicators (CPU utilization, memory bandwidth, response delay, etc.) are collected every 30 seconds after tuning, and a performance baseline is formed for 5 minutes, and the optimization effect is evaluated by comparing the index fluctuations before and after tuning.

[0091] Therefore, the overall data flow and closed loop logic of the present application are as follows: the system realizes intelligent closed-loop optimization through the following data flow: the raw data collected by the intelligent probe is first cleaned and converted by the data preprocessing module to form standardized training data, which is input into the AI training and problem analysis module; after the AI module completes model training and problem diagnosis, the optimization suggestion is transmitted to the intelligent tuning module to generate specific strategies; after the strategy is executed by the full-stack tuning module, the system performance changes are captured again by the intelligent probe to form a complete closed loop of "collection-analysis-decision-execution-feedback". This closed loop mechanism ensures that the optimization strategy has self-iteration capability: when the system environment or load characteristics change, the intelligent probe can sense in real time and trigger a new round of model training, so that the tuning strategy continuously adapts to dynamic scenarios, and finally realizes long-term stable optimization of full-stack performance.

[0092] The technical effects of the present application are as follows:

[0093] 1. Vertical and horizontal coordination mechanism constitutes the basic framework of the system: this architecture breaks through the limitations of traditional single-layer tuning, and forms a two-way linkage through the left vertical optimization level covering all layers of the full stack (including hardware, operating system, middleware and application layer) and the right horizontal processing flow (AI-driven performance analysis and tuning flow). The vertical level is responsible for data collection and local optimization of each layer, and the horizontal flow realizes cross-level data integration and global decision-making through AI algorithms. The two work together to dynamically schedule cross-level resources, effectively solving the resource mismatch problem caused by "layered islands" in traditional optimization, and improving the utilization rate of full-stack resources.

[0094] 2. AI deep integration realizes the core engine of intelligent tuning: the present application deeply embeds machine learning models (such as performance prediction model, root cause positioning model) into the performance analysis flow, replacing the traditional tuning method relying on static rules. The performance prediction model can identify potential bottlenecks in advance based on historical data and real-time indicators, and the root cause positioning model can quickly locate the source of the problem through multi-dimensional data correlation analysis, and the two form a dynamic optimization decision-making mechanism. This AI-driven mode can adapt to the characteristics of variable application load and heterogeneous environment, and realize the paradigm shift from "passive response" to "active prevention".

[0095] 3. The modular and scalable design provides continuous evolution capability for the system: each functional module (vertical level module, intelligent probe, data preprocessing module, AI model module, etc.) is connected through standardized interface to realize loose coupling, supporting two extension modes: one is independent upgrade of the module, for example, the AI prediction model can be replaced according to actual needs, such as upgrading the LSTM (Long Short Term Memory) model to the Transformer (Deep Model Learning Architecture); the other is horizontal expansion of functions, such as adding a monitoring module for domestic databases, or integrating a hardware performance collection unit for a specific chip architecture. This design ensures that the system can adapt to the rapid iteration of the domestic technology ecosystem and reduce the development cost of adapting new software and hardware.

[0096] In summary, the full-stack application system optimization method based on artificial intelligence provided by the present application can realize real-time capture and feature engineering of full-stack data through the distributed collection architecture and the stream computing engine of the "intelligent probe" and the "data preprocessing" module, realizes intelligent, dynamic and cross-level collaborative optimization of system performance, effectively improves system response speed, throughput and stability, and provides an innovative and efficient solution for system performance optimization, fundamentally solves the problem of insufficient data support capability in the prior art, solves the adaptability and collaboration bottleneck of the traditional scheme, and is applicable to the full-stack IT architecture constructed based on domestic hardware (such as Feiteng CPU and Kunpeng CPU), domestic operating system (such as Kylin advanced server operating system), domestic middleware and domestic database (such as Dameng database).

[0097] Although the present application has been described by the above preferred embodiments, it is not intended to limit the protection scope of the present application, and any person skilled in the art can make various changes and modifications to the above embodiments without departing from the spirit and scope of the present application, which are still within the protection scope of the present application. Therefore, the protection scope of the present application is defined by the claims.

Claims

1. A systematic optimization method for a full-stack application based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Achieve end-to-end data collection using smart probes; Step S2: Perform data preprocessing on the data collected in step S1; Step S3: Input the preprocessed data into the professional model system, identify memory leak risks, perform correlation analysis on hardware indicators and application performance indicators, and locate the root cause of the problem; The specialized model system includes a JVM memory prediction model based on long short-term memory and a performance attribution model that integrates NUMA architecture features. The JVM memory prediction model based on long short-term memory is used to train time-series prediction capabilities using historical memory fluctuation data and to identify memory leak risks in advance based on preprocessed data. The performance attribution model that integrates NUMA architecture features is used to perform correlation analysis between hardware metrics and application performance metrics to locate root causes of problems. Hardware metrics include memory access efficiency, cache hit rate, and CPU utilization. Step S4: Integrate the results of step S3, perform dynamic optimization through the dynamic strategy library of the hierarchical classification storage architecture, generate dynamic optimization strategies, generate optimization parameters and inject them into the target system; the hierarchical classification storage architecture divides the optimization rule set according to the system abstraction level, including application layer strategies and basic layer strategies. Application-layer strategies focus on business logic optimization, including adjusting cache expiration time and configuring connection pool size. The base layer strategy is optimized for hardware and operating system levels, covering NUMA memory affinity configuration and CPU core binding. Step S5: Establish a standardized feedback data collection mechanism, collect key performance indicators at predetermined intervals, and evaluate the optimization results by comparing the fluctuations of key performance indicators before and after optimization.

2. The systematic optimization method for full-stack application systems based on artificial intelligence as described in claim 1, characterized in that, Step S1 includes: By using bytecode injection technology, heap memory usage, memory allocation rate, and GC pause time can be obtained within the JVM process. Deploy database probes on database nodes to monitor and query database performance parameters, including transaction throughput and lock wait time. Hardware probes are deployed at the underlying hardware level to obtain underlying resource data, including CPU utilization, memory bandwidth, number of input operations per second (IOPS) of storage devices, number of output operations per second (IOPS) of storage devices, memory access efficiency, cache hit rate, and processor performance.

3. The systematic optimization method for full-stack applications based on artificial intelligence as described in claim 1, characterized in that: The data preprocessing includes: A truncation strategy based on interquartile range is adopted for outliers; Time series interpolation is used to complete missing values; Perform consistency checks on the data to identify and remove invalid samples caused by probe failures; The data is deduplicated and subjected to feature engineering.

4. The systematic optimization method for full-stack applications based on artificial intelligence as described in claim 3, characterized in that: Data deduplication includes: for high-frequency collected index data, setting a sliding window with a predetermined time, calculating the MD5 hash value of duplicate data within the window, retaining only the original value of the first collection, and marking the remaining duplicate data as redundant and discarding it; Feature engineering processing includes: normalizing the data, mapping the original values ​​to the [0,1] interval to eliminate dimensional differences, then constructing a sliding window for a predetermined time, extracting the maximum value and variance within the window as key features, and forming a two-dimensional feature vector.

5. The systematic optimization method for full-stack application systems based on artificial intelligence as described in claim 1, characterized in that, Training methods for JVM memory prediction models based on Long Short-Term Memory include: Step S31: On the operating system, generate sample data using a load testing tool; Step S32: Divide the sample data into training set, validation set and test set according to a predetermined ratio; Step S33: Perform feature engineering on the training set data to obtain feature vectors; Step S34: Input the feature vector obtained in step S33 into the JVM memory prediction model of long short-term memory for training, continuously adjust the model parameters, and minimize the error between the model prediction result and the true value. Step S35: Validate the trained JVM memory prediction model for long short-term memory on the validation set, and adjust the parameters of the JVM memory prediction model for long short-term memory based on the validation results. Step S36: Test the JVM memory prediction model for Long Short-Term Memory on the test set to evaluate the prediction accuracy and generalization ability of the JVM memory prediction model for Long Short-Term Memory.

6. The systematic optimization method for full-stack application systems based on artificial intelligence as described in claim 1, characterized in that, Training methods for performance attribution models that incorporate NUMA architecture features include: Step S3a: Obtain hardware metrics and application performance metrics, and perform feature engineering on the hardware metrics and application performance metrics to obtain feature vectors; Step S3b: Input the feature vector obtained in step S3a into the performance attribution model that integrates NUMA architecture features for training. By continuously iterating and optimizing the model parameters, the performance attribution model that integrates NUMA architecture features can accurately find the correlation between hardware indicators and application performance indicators. Step S3c: Use cross-validation to evaluate and optimize the performance attribution model that incorporates NUMA architecture features to improve accuracy and stability.