A method and system for soft and hardware co-optimization based on DeepSeek

Abstract

The application discloses a kind of based on DeepSeek's software and hardware collaborative optimization method and system, belong to computer software and hardware optimization technical field.The method is by collecting hardware and software related data in computer system and pre-processing, constructs DeepSeek comprehensive analysis model to learn the performance influence law of software and hardware interaction relationship, generates diversified test scene based on user demand and system characteristics and is automatically executed, according to model output formulates optimization strategy and then iteratively optimizes.The system includes data acquisition, pre-processing, DeepSeek analysis model, test scene generation and execution and optimization strategy formulation and implementation module, realize software and hardware collaborative optimization, break through independent optimization limit, improve test and optimization efficiency, reduce resource waste, enhance the adaptability to different computer systems, guarantee system running stability and reliability.

Description

Technical Field

[0001] This invention relates to the field of computer software and hardware optimization technology, and in particular to a software and hardware co-tuning method and system based on DeepSeek. Background Technology

[0002] As computer systems continue to evolve towards higher performance and greater complexity, the collaborative operation of hardware and software has become a core element in ensuring stable and efficient system operation. Currently, the industry is witnessing a continuous enrichment of hardware types and performance improvements, ranging from basic computing components to dedicated acceleration hardware, covering application needs across various scenarios. Software systems are evolving towards functional diversification and modular architecture, with increasingly tighter interactions between various applications, operating systems, and middleware. Simultaneously, hardware and software testing and optimization techniques are also continuously developing. Hardware testing is increasingly focusing on monitoring the collaborative status of multiple components, while software testing emphasizes performance control throughout the entire lifecycle. Related tools and methods are constantly iterating to adapt to increasingly complex system environments. Whether it's server clusters in data centers or personal terminal devices, the collaborative efficiency of hardware and software directly affects the overall user experience and operational efficiency, becoming a key focus within the industry.

[0003] However, current software and hardware optimization technologies still face many unresolved issues. Traditional optimization methods often employ independent optimization of software and hardware, failing to fully consider their interaction and mutual influence, resulting in limited optimization effectiveness and difficulty in achieving optimal overall system performance. At the data processing level, there is a lack of comprehensive and standardized data collection and preprocessing mechanisms; different data types have inconsistent formats and varying quality, affecting the accuracy of subsequent analysis. Regarding model construction and application, existing analytical models struggle to deeply explore the interaction patterns between software and hardware and their impact on system performance, failing to provide precise support for optimization. The generation of test scenarios lacks diversity and specificity, failing to cover complex operating environments and load conditions, thus limiting the reference value of test results. The formulation of optimization strategies relies heavily on manual experience, resulting in low automation and a lack of closed-loop iterative mechanisms, making continuous optimization based on test feedback difficult and adaptable to dynamically changing system requirements. These problems lead to underutilization of software and hardware resources, slow system performance improvement, and difficulty in guaranteeing operational stability and reliability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a software and hardware co-tuning method and system based on DeepSeek.

[0005] The objective of this invention is achieved through the following technical solution: A hardware / software co-tuning method based on DeepSeek is provided, which includes the following steps: S1. Collect hardware and software related data from the computer system and preprocess the collected raw data; S2. Construct a DeepSeek comprehensive analysis model. Input the preprocessed data into the DeepSeek comprehensive analysis model and learn the interaction relationship between hardware and software and their impact on system performance through the DeepSeek comprehensive analysis model. S3. Based on user needs and system characteristics, use the DeepSeek comprehensive analysis model to generate test scenarios and execute them automatically, collecting hardware and software data in real time during the test process; S4. Based on the output of the DeepSeek comprehensive analysis model, formulate hardware and software optimization strategies, automatically implement the optimization strategies, and then run the test scenario again. Compare the system performance data before and after optimization, and iterate until the system performance meets the requirements.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Set up multiple data acquisition points within the computer system to collect hardware performance data and software operation-related data. Hardware components include CPU, GPU, memory, and hard disk. Hardware performance data includes the temperature, usage rate, frequency, and read / write speed of these components. Software operation-related data includes application operation logs, system call counts, and resource usage. S1.2. Clean the collected raw data, remove noisy data and outliers, and convert different types of data into a unified format to prepare for subsequent DeepSeek comprehensive analysis model analysis.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. A multi-layer neural network structure is constructed to form the DeepSeek comprehensive analysis model. The structure includes an input layer, a feature encoding layer, a deep feature extraction layer, a hardware-software interaction relationship learning module, a system performance impact law learning module, a multi-task learning module, and an output layer. The input layer receives hardware data and software data. The feature encoding layer encodes the received data and then passes it to the deep feature extraction layer. The deep feature extraction layer extracts data features and then passes them to the hardware-software interaction relationship learning module. The hardware-software interaction relationship learning module learns the data interaction relationship and then passes it to the system performance impact law learning module. The system performance impact law learning module, in conjunction with the multi-task learning module, processes the data and then outputs the system performance prediction results and interpretability analysis results through the output layer. S2.2. Collect historical test data of different types of computer systems. The historical test data includes hardware configuration information, software operation status and corresponding system performance indicators. Divide the historical test data and the actual test data into training set and test set. Train the DeepSeek comprehensive analysis model through supervised learning and reinforcement learning. In supervised learning, the hardware and software configuration combinations and corresponding system performance indicators in the historical data are used as labels to adjust the model parameters. In reinforcement learning, the improvement of system performance is used as a reward signal to optimize the model decision. The model parameters are continuously adjusted to optimize the prediction accuracy of the DeepSeek comprehensive analysis model.

[0008] Furthermore, step S3 includes the following sub-steps: The S3.1.DeepSeek comprehensive analysis model generates diverse test scenarios based on the learned hardware-software interaction patterns and user needs. These test scenarios include different hardware load conditions, software application combinations, and system operating environment settings. S3.2. Execute the generated test scenarios in a simulated environment using automated testing tools. At the same time, use hardware monitoring tools and software automation tools to comprehensively collect hardware performance data and software operation data during the testing process. Feed the collected data back to the DeepSeek comprehensive analysis model for further optimization.

[0009] Furthermore, step S4 includes the following sub-steps: The S4.1.DeepSeek comprehensive analysis model analyzes the feedback test data. If it identifies a performance bottleneck in a software module under the current hardware configuration, it formulates a software code optimization strategy. If it finds problems in the collaboration between hardware components, it formulates a hardware parameter adjustment strategy or recommends hardware component upgrades. S4.2. Perform optimization operations according to the established tuning strategy. After completion, execute the corresponding test scenario again, compare and analyze the system performance data before and after tuning. If the performance does not meet the expected requirements, formulate a new tuning strategy based on the new test data and repeat the optimization process.

[0010] Furthermore, in step S1.2, normalization technology is used to convert the format of numerical data, and preprocessing is performed on text log data. The preprocessing operations include word segmentation, removal of words without practical meaning, and text standardization. Then, word embedding technology is used to convert the text into a fixed-length vector representation. Word embedding technology includes continuous bag-of-words model and skip word model. The training process of word embedding technology includes building a vocabulary, initializing word vectors, and iteratively training and adjusting word vectors until the model converges.

[0011] Furthermore, in step S2.1, the feature encoding layer uses numerical normalization or category encoding to process hardware data and word embedding technology to process software data; the deep feature extraction layer includes a basic feature extraction component, a temporal data processing component, and an interactive learning mechanism. The basic feature extraction component is a fully connected layer, the temporal data processing component can be equipped with a convolutional neural network or a long short-term memory network, and the interactive learning mechanism is a multi-head attention mechanism; the hardware-software interaction relationship learning module uses a graph neural network and a causal inference network. The graph neural network treats hardware and software as nodes and the interaction relationship as an edge to learn the global relationship. The causal inference network analyzes the causal impact of hardware and software configuration on system performance; the multi-task learning module includes a system performance prediction task, an anomaly detection task, and a hardware-software configuration recommendation task.

[0012] Furthermore, in step S3.1, the DeepSeek comprehensive analysis model samples from the hardware configuration space and software parameter space through algorithms. The hardware configuration space includes CPU frequency and memory capacity, while the software parameter space includes the number of threads and caching strategy. The generated test scenarios are output in the form of configuration files, which specify the hardware parameter settings and software startup parameters. The sampling process combines the learned hardware and software interaction rules to ensure scenario diversity.

[0013] Furthermore, in step S4.2, the software code optimization strategy includes algorithm improvement and resource allocation adjustment, which are implemented and integrated into the continuous integration and continuous deployment process, automatically triggering code modification, compilation and deployment; the hardware parameter adjustment strategy is implemented through the hardware management interface to achieve remote automated configuration. After each tuning operation is completed, the new test data and tuning results are recorded in the historical dataset for further optimization of the training data of the DeepSeek comprehensive analysis model.

[0014] In some embodiments, a software and hardware collaborative optimization system based on DeepSeek is provided, which includes a data acquisition module, a data preprocessing module, a DeepSeek analysis model module, a test scenario generation and execution module, and an optimization strategy formulation and implementation module. The data acquisition module is used to set up multiple data acquisition points to collect hardware performance data and software operation-related data; the data preprocessing module is used to clean, normalize, and convert the format of the collected raw data; the DeepSeek analysis model module is used to build a multi-layer neural network structure, and form a DeepSeek comprehensive analysis model through supervised learning and reinforcement learning to learn the interaction relationship between hardware and software and the performance impact law; the test scenario generation and execution module is used to generate diverse test scenarios using the DeepSeek comprehensive analysis model and execute them automatically, collecting test data in real time and providing feedback; the optimization strategy formulation and implementation module is used to formulate optimization strategies based on the output of the DeepSeek comprehensive analysis model, automatically implement and iteratively optimize until the system performance meets the requirements.

[0015] The beneficial effects of this invention are: (1) Through the core process of “data acquisition and preprocessing - model building and learning - test scenario generation - optimization strategy iteration”, we can achieve software and hardware collaborative optimization, break the limitations of independent optimization, and make software and hardware resources more fully adapted; (2) Automated testing scenario generation, optimization strategy implementation and data feedback reduce manual intervention costs, avoid resource waste caused by blind optimization, and improve testing and optimization efficiency; (3) By continuously learning the model and supplementing and optimizing the dataset, the adaptability of the method to different computer systems is enhanced, and the stability and reliability of the system operation are improved. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a software and hardware co-tuning method based on DeepSeek. Figure 2 The following is a flowchart illustrating the specific steps of a software and hardware co-tuning method based on DeepSeek, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0018] Example 1 See Figure 1 This embodiment provides a hardware and software co-tuning method based on DeepSeek, which includes the following steps: S1. Collect hardware and software related data from the computer system and preprocess the collected raw data; S2. Construct a DeepSeek comprehensive analysis model. Input the preprocessed data into the DeepSeek comprehensive analysis model and learn the interaction relationship between hardware and software and their impact on system performance through the DeepSeek comprehensive analysis model. S3. Based on user needs and system characteristics, use the DeepSeek comprehensive analysis model to generate test scenarios and execute them automatically, collecting hardware and software data in real time during the test process; S4. Based on the output of the DeepSeek comprehensive analysis model, formulate hardware and software optimization strategies, automatically implement the optimization strategies, and then run the test scenario again. Compare the system performance data before and after optimization, and iterate until the system performance meets the requirements.

[0019] Furthermore, step S1 includes the following sub-steps: S1.1. Set up multiple data acquisition points within the computer system to collect hardware performance data and software operation-related data. Hardware components include CPU, GPU, memory, and hard disk. Hardware performance data includes the temperature, usage rate, frequency, and read / write speed of these components. Software operation-related data includes application operation logs, system call counts, and resource usage. S1.2. Clean the collected raw data, remove noisy data and outliers, and convert different types of data into a unified format to prepare for subsequent DeepSeek comprehensive analysis model analysis.

[0020] Furthermore, step S2 includes the following sub-steps: S2.1. A multi-layer neural network structure is constructed to form the DeepSeek comprehensive analysis model. The structure includes an input layer, a feature encoding layer, a deep feature extraction layer, a hardware-software interaction relationship learning module, a system performance impact law learning module, a multi-task learning module, and an output layer. The input layer receives hardware data and software data. The feature encoding layer encodes the received data and then passes it to the deep feature extraction layer. The deep feature extraction layer extracts data features and then passes them to the hardware-software interaction relationship learning module. The hardware-software interaction relationship learning module learns the data interaction relationship and then passes it to the system performance impact law learning module. The system performance impact law learning module, in conjunction with the multi-task learning module, processes the data and then outputs the system performance prediction results and interpretability analysis results through the output layer. S2.2. Collect historical test data of different types of computer systems. The historical test data includes hardware configuration information, software operation status and corresponding system performance indicators. Divide the historical test data and the actual test data into training set and test set. Train the DeepSeek comprehensive analysis model through supervised learning and reinforcement learning. In supervised learning, the hardware and software configuration combinations and corresponding system performance indicators in the historical data are used as labels to adjust the model parameters. In reinforcement learning, the improvement of system performance is used as a reward signal to optimize the model decision. The model parameters are continuously adjusted to optimize the prediction accuracy of the DeepSeek comprehensive analysis model.

[0021] Furthermore, step S3 includes the following sub-steps: The S3.1.DeepSeek comprehensive analysis model generates diverse test scenarios based on the learned hardware-software interaction patterns and user needs. These test scenarios include different hardware load conditions, software application combinations, and system operating environment settings. S3.2. Execute the generated test scenarios in a simulated environment using automated testing tools. At the same time, use hardware monitoring tools and software automation tools to comprehensively collect hardware performance data and software operation data during the testing process. Feed the collected data back to the DeepSeek comprehensive analysis model for further optimization.

[0022] Furthermore, step S4 includes the following sub-steps: The S4.1.DeepSeek comprehensive analysis model analyzes the feedback test data. If it identifies a performance bottleneck in a software module under the current hardware configuration, it formulates a software code optimization strategy. If it finds problems in the collaboration between hardware components, it formulates a hardware parameter adjustment strategy or recommends hardware component upgrades. S4.2. Perform optimization operations according to the established tuning strategy. After completion, execute the corresponding test scenario again, compare and analyze the system performance data before and after tuning. If the performance does not meet the expected requirements, formulate a new tuning strategy based on the new test data and repeat the optimization process.

[0023] Furthermore, in step S1.2, normalization technology is used to convert the format of numerical data, and preprocessing is performed on text log data. The preprocessing operations include word segmentation, removal of words without practical meaning, and text standardization. Then, word embedding technology is used to convert the text into a fixed-length vector representation. Word embedding technology includes continuous bag-of-words model and skip word model. The training process of word embedding technology includes building a vocabulary, initializing word vectors, and iteratively training and adjusting word vectors until the model converges.

[0024] Furthermore, in step S2.1, the feature encoding layer uses numerical normalization or category encoding to process hardware data and word embedding technology to process software data; the deep feature extraction layer includes a basic feature extraction component, a temporal data processing component, and an interactive learning mechanism. The basic feature extraction component is a fully connected layer, the temporal data processing component can be equipped with a convolutional neural network or a long short-term memory network, and the interactive learning mechanism is a multi-head attention mechanism; the hardware-software interaction relationship learning module uses a graph neural network and a causal inference network. The graph neural network treats hardware and software as nodes and the interaction relationship as an edge to learn the global relationship. The causal inference network analyzes the causal impact of hardware and software configuration on system performance; the multi-task learning module includes a system performance prediction task, an anomaly detection task, and a hardware-software configuration recommendation task.

[0025] Furthermore, in step S3.1, the DeepSeek comprehensive analysis model samples from the hardware configuration space and software parameter space through algorithms. The hardware configuration space includes CPU frequency and memory capacity, while the software parameter space includes the number of threads and caching strategy. The generated test scenarios are output in the form of configuration files, which specify the hardware parameter settings and software startup parameters. The sampling process combines the learned hardware and software interaction rules to ensure scenario diversity.

[0026] Furthermore, in step S4.2, the software code optimization strategy includes algorithm improvement and resource allocation adjustment, which are implemented and integrated into the continuous integration and continuous deployment process, automatically triggering code modification, compilation and deployment; the hardware parameter adjustment strategy is implemented through the hardware management interface to achieve remote automated configuration. After each tuning operation is completed, the new test data and tuning results are recorded in the historical dataset for further optimization of the training data of the DeepSeek comprehensive analysis model.

[0027] In some embodiments, a software and hardware collaborative optimization system based on DeepSeek is provided, which includes a data acquisition module, a data preprocessing module, a DeepSeek analysis model module, a test scenario generation and execution module, and an optimization strategy formulation and implementation module. The data acquisition module is used to set up multiple data acquisition points to collect hardware performance data and software operation-related data; the data preprocessing module is used to clean, normalize, and convert the format of the collected raw data; the DeepSeek analysis model module is used to build a multi-layer neural network structure, and form a DeepSeek comprehensive analysis model through supervised learning and reinforcement learning to learn the interaction relationship between hardware and software and the performance impact law; the test scenario generation and execution module is used to generate diverse test scenarios using the DeepSeek comprehensive analysis model and execute them automatically, collecting test data in real time and providing feedback; the optimization strategy formulation and implementation module is used to formulate optimization strategies based on the output of the DeepSeek comprehensive analysis model, automatically implement and iteratively optimize until the system performance meets the requirements.

[0028] Example 2 This embodiment provides a specific implementation process for a software and hardware collaborative optimization method based on DeepSeek. This process revolves around the comprehensive collection and preprocessing of software and hardware data, the construction and training of the DeepSeek comprehensive analysis model, the generation and execution of diverse test scenarios, and the formulation and iterative optimization of targeted optimization strategies. Through the close connection and data exchange of each step, the goal of software and hardware collaborative optimization is achieved. Figure 2 As shown, the specific steps are as follows: S1. Collect and preprocess relevant hardware and software data: S1.1. Set up multiple data acquisition points within the computer system to collect hardware performance data and software operation-related data. Hardware components include CPU, GPU, memory, and hard drive. Hardware performance data includes the temperature, usage rate, frequency, and read / write speed of these components. Software operation-related data includes application runtime logs, system call counts, and resource usage. Multiple data acquisition points are set up at key hardware nodes and software execution paths in the computer system to ensure comprehensive capture of relevant data on hardware operating status and software execution. For hardware, the data collected covers core components such as CPU, GPU, memory, and hard drive. For each component, performance data reflecting its operating status, such as temperature, usage rate, frequency, and read / write speed, are collected. This data directly reflects the hardware's load and efficiency during operation. For software, the data collected includes application runtime logs, system call counts, and resource usage. Runtime logs record the entire process of software startup, operation, and termination. System call counts reflect the frequency of software interaction with the operating system, and resource usage reflects the software's consumption of hardware resources. During the acquisition process, hardware data is captured in real time by monitoring units adapted to each hardware component, while software data is recorded through data acquisition logic integrated into the software system, ensuring that the collected data accurately reflects the actual operating status of the hardware and software.

[0029] In some embodiments, hardware components may also include a motherboard, graphics card, and network card, and the corresponding hardware performance data may include the power supply stability parameters of the motherboard, the rendering frame rate of the graphics card, and the data transmission rate of the network card; software operation-related data may also include the software response time, the number of errors thrown, and the number of threads occupied by the process, etc. By expanding the types and scope of collected data, the comprehensiveness and effectiveness of the data can be further improved.

[0030] S1.2. Clean the collected raw data, remove noisy data and outliers, and convert different types of data into a unified format to prepare for subsequent analysis using the DeepSeek comprehensive analysis model: The raw data collected may contain noise and outliers caused by equipment malfunctions, transmission interference, or momentary hardware / software anomalies. These data can affect the accuracy of subsequent analysis, so data cleaning is necessary first. Through pre-defined filtering rules and anomaly identification logic, the raw data is verified one by one, removing noise and outliers that significantly deviate from the normal range to ensure the authenticity and reliability of the retained data.

[0031] After data cleaning, since hardware performance data is mostly numerical and software operation data includes both numerical and text data (such as operation logs), the different types of data have different formats and value ranges, making them unsuitable for direct input into the DeepSeek comprehensive analysis model. Therefore, it is necessary to convert different types of data into a unified format.

[0032] For numerical data, normalization technology is used for format conversion. Through specific calculation methods, numerical data with different value ranges are mapped to a unified numerical interval, eliminating the impact of differences in data units and making various types of numerical data comparable and compatible.

[0033] For text-based log data, preprocessing operations are performed first. These operations include word segmentation, removal of meaningless words, and text standardization. Word segmentation breaks down continuous log text into independent word units, facilitating subsequent feature extraction. Removal of meaningless words eliminates words in the log that do not carry effective information, reducing data redundancy. Text standardization unifies text with different formats and expressions into a standardized form, such as unifying capitalization and correcting typos. After preprocessing, word embedding technology is used to transform the text into a fixed-length vector representation. Word embedding technology includes the continuous bag-of-words model and the skip-word model. The training process of word embedding technology includes building a vocabulary, initializing word vectors, and iteratively training and adjusting word vectors until the model converges. In the process of building the vocabulary, all different words appearing in all log texts after preprocessing are counted to form a complete vocabulary set. In the process of initializing word vectors, a randomly generated dense vector is assigned to each word in the vocabulary, and the dimension of the vector is pre-set according to actual needs. In the process of iteratively training and adjusting word vectors, appropriate learning rate and iteration number parameters are set. The continuous bag-of-words model predicts the target word with context words as input, and the skip-word model predicts the context words with the target word as input. By continuously adjusting the word vectors, the difference between the model's prediction results and the actual results gradually decreases until the model converges, and finally a fixed-length vector that can accurately represent the semantic information of the text is obtained.

[0034] In some embodiments, for the format conversion of numerical data, in addition to normalization techniques, standardization techniques can also be used. By calculating the mean and standard deviation of the data, the data can be converted into a format that conforms to a specific distribution pattern. For the preprocessing of text log data, a word weight calculation step can also be added. By statistically analyzing the frequency and importance of words in the log text, corresponding weights can be assigned to different words to improve the effectiveness of subsequent word vector representation.

[0035] S2. Build models and learn the patterns of software-hardware interaction and performance impact: S2.1. A multi-layer neural network structure is constructed to form the DeepSeek comprehensive analysis model. This structure includes an input layer, a feature encoding layer, a deep feature extraction layer, a hardware-software interaction learning module, a system performance impact law learning module, a multi-task learning module, and an output layer. The input layer receives hardware and software data. The feature encoding layer encodes the received data and then passes it to the deep feature extraction layer. The deep feature extraction layer extracts data features and then passes them to the hardware-software interaction learning module. The hardware-software interaction learning module learns the data interaction relationships and then passes them to the system performance impact law learning module. The system performance impact law learning module, in conjunction with the multi-task learning module, processes the data and outputs the system performance prediction results and interpretability analysis results at the output layer. A multi-layered neural network structure was built to form the DeepSeek comprehensive analysis model. This model has a clear hierarchical structure, with each part working collaboratively to ensure the deep extraction of effective information from the data. The input layer, serving as the model's data entry point, specifically receives pre-processed hardware and software data, integrating these two types of data into an input format that the model can process, providing fundamental data support for subsequent processing at each layer.

[0036] After receiving hardware and software data from the input layer, the feature encoding layer performs encoding processing on the two types of data separately: For hardware data, numerical normalization or categorical encoding is applied. Numerical normalization further optimizes the numerical distribution of the hardware data to better suit the model's training requirements, while categorical encoding converts the categorical information in the hardware data into numerical form. For software data, word embedding technology is used to further optimize the preprocessed text vectors in the software data, enhancing the integration of textual and numerical information. After encoding, the encoded hardware and software data are fed into the deep feature extraction layer.

[0037] The deep feature extraction layer comprises a basic feature extraction component, a temporal data processing component, and an interactive learning mechanism. The basic feature extraction component is a fully connected layer, which extracts basic features from the encoded data through comprehensive connections between neurons, capturing global feature information. The temporal data processing component can be equipped with either a convolutional neural network or a long short-term memory network. When the data contains temporal features, the convolutional neural network can effectively extract local temporal features, while the long short-term memory network can capture long-term dependencies. The appropriate temporal data processing component is selected based on the actual characteristics of the data. The interactive learning mechanism is a multi-head attention mechanism, which focuses on the interactive information of key features in the data through parallel computation of multiple attention heads, improving the targeting and effectiveness of feature extraction. After completing feature extraction, the deep feature extraction layer feeds the extracted deep features into the hardware-software interactive relationship learning module.

[0038] The hardware-software interaction learning module employs graph neural networks and causal inference networks. The graph neural network treats hardware and software as nodes in a graph, and the interactions between them as edges. It learns the global interactions between hardware and software through graph convolution operations, uncovering the correlation patterns between different hardware components and different software modules. The causal inference network constructs a causal graph model to analyze the direct and indirect impacts of hardware configuration and software settings on system performance, clarifying the causal relationships between various factors and system performance. After completing the interaction learning, this module feeds the learning results into the system performance impact law learning module.

[0039] The system performance impact learning module receives the output from the hardware-software interaction learning module and then collaborates with the multi-task learning module. This multi-task learning module includes three tasks: system performance prediction, anomaly detection, and hardware-software configuration recommendation. The system performance prediction task focuses on learning the mapping relationship between hardware and software configuration combinations and system performance indicators; the anomaly detection task identifies abnormal patterns during hardware and software operation; and the hardware-software configuration recommendation task explores optimal hardware and software configuration combinations. Through this multi-task collaborative learning, the system performance impact learning module comprehensively grasps the impact patterns of hardware and software on system performance. Finally, the output layer outputs system performance prediction results and interpretability analysis results. The system performance prediction results represent the expected system performance under different configuration combinations, while the interpretability analysis results clarify the impact weights and action paths of each hardware and software factor on system performance.

[0040] In some embodiments, the deep feature extraction layer can also be configured with a gated recurrent unit to replace the long short-term memory network. The gated recurrent unit, through the design of reset gates and update gates, can also effectively capture the temporal dependencies in the data and has certain advantages in terms of computational complexity. The causal inference network in the hardware-software interaction relationship learning module can adopt different causal graph construction algorithms. The optimal construction method can be selected according to the characteristics of the data and the actual application scenario to improve the accuracy of causal analysis.

[0041] S2.2. Collect historical test data of different types of computer systems. This historical test data includes hardware configuration information, software operation status, and corresponding system performance indicators. Divide the historical test data and actual test data into training and testing sets. Train the DeepSeek comprehensive analysis model using supervised learning and reinforcement learning. In supervised learning, use the hardware and software configuration combinations and corresponding system performance indicators from the historical data as labels to adjust model parameters. In reinforcement learning, use system performance improvement as a reward signal to optimize model decisions. Continuously adjust model parameters to improve the prediction accuracy of the DeepSeek comprehensive analysis model. To ensure the DeepSeek comprehensive analysis model has good generalization ability and predictive accuracy, it is necessary to collect historical test data from different types of computer systems. These computer systems cover different hardware configuration combinations, software installation environments, and operating scenarios. The historical test data specifically includes hardware configuration information, software operation status, and corresponding system performance indicators: the hardware configuration information records in detail the model, parameter settings, etc. of the hardware components in each computer system; the software operation status includes information such as the software version, operating parameters, and task type; and the corresponding system performance indicators are the actual performance data of the system under the hardware configuration and software operation conditions, such as throughput, response speed, and resource utilization.

[0042] The collected historical test data is integrated with the actual test data collected and preprocessed in step S1, and then divided into training and test sets according to a preset ratio. The training set is used for model parameter learning, and the test set is used for model performance verification and evaluation. During the division process, it is ensured that the data distribution in the training and test sets is consistent to avoid bias in model training due to differences in data distribution.

[0043] A combination of supervised learning and reinforcement learning was used to train the DeepSeek holistic analysis model. In the supervised learning process, hardware and software configuration combinations and corresponding system performance indicators from historical data were used as labels. Preprocessed hardware and software data were input into the DeepSeek holistic analysis model, which outputs system performance prediction results. By calculating the difference between the prediction results and the labels, a loss function was constructed. Then, the gradient descent algorithm was used to adjust the parameters of each level of the model, gradually reducing the value of the loss function and improving the model's ability to fit known data.

[0044] In reinforcement learning, the decision-making process of the DeepSeek synergistic analysis model is regarded as an interaction between the agent and the environment. The environment is the hardware and software operating scenario of the computer system, and the agent is the DeepSeek synergistic analysis model. The model outputs hardware and software configuration adjustment suggestions based on the current hardware and software data, and uses the improvement of system performance as a reward signal. If the adjusted system performance is better than before the adjustment, a positive reward is given; otherwise, a negative reward is given. The model continuously optimizes its own decision-making logic and adjusts its parameters based on the reward signals, so that the model can learn the optimal decision strategy that can bring about the improvement of system performance.

[0045] During model training, the model's performance on the test set is continuously monitored. The model's prediction accuracy is verified using test set data. If the model's error on the test set exceeds a preset threshold, the training process is returned to its previous state. The weights of the supervised learning loss function, the reward coefficient of reinforcement learning, and the model's hyperparameters (such as learning rate, number of iterations, number of network layers, number of neurons, etc.) are adjusted, and training is performed again. Through multiple iterations of training, the model parameters are continuously adjusted, and the prediction accuracy of the DeepSeek comprehensive analysis model is continuously optimized until the model achieves stable and ideal performance on both the training and test sets.

[0046] In some embodiments, in addition to supervised learning and reinforcement learning, semi-supervised learning can be introduced to assist model training. This involves expanding the training samples with a large amount of unlabeled data, mining the potential information in the unlabeled data through semi-supervised learning algorithms, assisting in adjusting model parameters, and further improving the model's generalization ability and prediction accuracy. During the data partitioning process, cross-validation can be used to partition the training set and the test set. By performing multiple different partitions and training sessions, the impact of data partitioning on model performance evaluation can be reduced.

[0047] S3. Generate and execute test scenarios to collect test data: S3.1. The DeepSeek comprehensive analysis model generates diverse test scenarios based on learned hardware-software interaction patterns and user needs. These test scenarios include different hardware load conditions, software application combinations, and system operating environment settings. The DeepSeek comprehensive analysis model samples from the hardware configuration space and software parameter space through algorithms. The hardware configuration space includes CPU frequency and memory capacity, while the software parameter space includes the number of threads and caching strategies. The generated test scenarios are output in the form of configuration files, which specify the hardware parameter settings and software startup parameters. The sampling process, combined with the learned hardware-software interaction patterns, ensures scenario diversity. After training, the DeepSeek comprehensive analysis model has fully learned the interaction patterns between hardware and software, as well as their impact on system performance. Combining user-defined system performance requirements (such as improving system speed and reducing resource consumption) with the characteristics of the computer system itself (such as the maximum load capacity of hardware components and the compatibility range of software), the DeepSeek comprehensive analysis model initiates the test scenario generation process.

[0048] Test scenario generation is based on sampling operations in the hardware configuration space and software parameter space. The hardware configuration space includes adjustable ranges for core hardware parameters such as CPU frequency and memory capacity, while the software parameter space includes configurable options for key software operation parameters such as thread count and caching strategy. The DeepSeek comprehensive analysis model samples from the hardware configuration space and software parameter space using a specific algorithm. The sampling process strictly combines the learned hardware-software interaction rules to ensure the rationality and effectiveness of the sampled hardware and software parameter combinations. It also covers different hardware load scenarios, software application combinations, and system operating environment settings, achieving test scenario diversification.

[0049] The diverse test scenarios specifically include different hardware load conditions (such as low load, medium load, high load, etc.), software application combinations (such as single software running, multiple software running in parallel, etc.), and system operating environment settings (such as different operating system configurations, network environment settings, etc.). The generated test scenarios are output in the form of configuration files, which clearly specify the corresponding hardware parameter settings (such as the specific CPU frequency, specific memory capacity, etc.) and software startup parameters (such as the specific number of threads, specific caching strategy type, etc.) for each test scenario, providing a clear and explicit basis for the subsequent automatic execution of test scenarios.

[0050] In some embodiments, the hardware configuration space may also include parameters such as GPU frequency and hard disk read / write cache size, and the software parameter space may also include parameters such as software process priority and data transmission protocol. By expanding the parameter types in the sampling space, the diversity of test scenarios can be further enriched. The sampling algorithm may adopt Monte Carlo tree search algorithm or genetic algorithm, etc., and the appropriate algorithm can be selected according to actual needs to improve sampling efficiency and the rationality of scenario generation.

[0051] S3.2. Utilize automated testing tools to execute the generated test scenarios in a simulated environment. Simultaneously, employ hardware monitoring tools and software automation tools to comprehensively collect hardware performance data and software operation data during the testing process. Feed the collected data back to the DeepSeek comprehensive analysis model for further optimization. First, a simulation environment consistent with the actual computer system's operating environment is built. This simulation environment can accurately reproduce the hardware configuration logic, software operation mechanism, and data interaction method of the actual system, ensuring that the execution results of the test scenarios are valuable for reference. Then, using automated testing tools, the configuration file generated in step S3.1 is read, and each test scenario is automatically executed sequentially in the simulation environment according to the parameter settings in the configuration file.

[0052] During the test scenario execution, hardware monitoring tools are used to collect hardware performance data in real time. These tools establish data interaction channels with the simulated hardware components in the simulation environment, continuously capturing performance data such as temperature, usage, frequency, and read / write speeds of hardware components like the CPU, GPU, memory, and hard drive. Simultaneously, software automation tools are used to collect software runtime data. These tools are embedded in the software's execution flow, recording application logs, system call counts, resource usage, and other data. The data collection process ensures real-time data quality and completeness, fully reflecting the dynamic operating status of the hardware and software during the test scenario execution.

[0053] After the test scenario is completed, the collected hardware performance data and software operation data are pre-processed according to the data preprocessing standards in step S1.2 to remove obviously invalid data. The processed data is then fed back to the DeepSeek comprehensive analysis model. This feedback data will serve as an important basis for further model optimization, helping the model to correct its own interaction pattern cognition and performance prediction logic, and improve the model's accuracy and adaptability.

[0054] In some embodiments, the simulation environment can be built using virtualization technology. Multiple independent virtual test environments can be created using virtualization software to enable the parallel execution of multiple test scenarios and improve test efficiency. Hardware monitoring tools and software automation operation tools can be deployed in a distributed manner, respectively, on different monitoring nodes, to collaboratively complete data collection tasks and ensure the comprehensiveness and stability of data collection.

[0055] S4. Develop and implement optimization strategies and iteratively optimize them: The S4.1.DeepSeek comprehensive analysis model analyzes the feedback test data. If it identifies a performance bottleneck in a software module under the current hardware configuration, it formulates software code optimization strategies. If it finds problems in the collaboration between hardware components, it formulates hardware parameter adjustment strategies or hardware component upgrade suggestions. After receiving the test data from step S3.2, the DeepSeek comprehensive analysis model performs in-depth analysis of the test data, combining it with its learned hardware-software interaction patterns and system performance impact patterns. During the analysis, the model focuses on the matching degree between software operating efficiency and hardware configuration, the collaborative working status between hardware components, and the performance indicators of the system.

[0056] If the model identifies a performance bottleneck in a software module under the current hardware configuration—meaning the module's operation is inefficient and this inefficiency is not due to insufficient hardware resources but rather to unreasonable code logic or runtime parameter settings—then a software code optimization strategy is formulated. This strategy targets the bottlenecked software module, clearly defining the code logic that needs improvement and the runtime parameters that need adjustment. The aim is to improve the software's performance and enhance its compatibility with the hardware configuration.

[0057] If the model identifies problems in the collaboration between hardware components—such as data transmission delays or resource allocation conflicts during operation that affect overall system performance—then a hardware parameter adjustment strategy or hardware component upgrade recommendation will be formulated based on the severity and specific manifestation of the problem. The hardware parameter adjustment strategy addresses collaboration issues that can be resolved through parameter configuration optimization, specifying the hardware parameters that need adjustment (such as bus transfer rate and memory timings). The hardware component upgrade recommendation addresses collaboration issues caused by insufficient performance of the hardware components themselves, proposing the types of hardware components that need to be replaced or the direction of upgrades.

[0058] S4.2. Execute optimization operations according to the established tuning strategy. After completion, execute the corresponding test scenario again, compare and analyze the system performance data before and after tuning. If the performance does not meet the expected requirements, formulate a new tuning strategy based on the new test data and repeat the optimization process. The software code optimization strategy includes algorithm improvement and resource allocation adjustment, and its implementation is integrated into the continuous integration and continuous deployment process, automatically triggering code modification, compilation, and deployment. The hardware parameter adjustment strategy is implemented through the hardware management interface to achieve remote automated configuration. After each tuning operation, the new test data and tuning results are recorded in the historical dataset for further optimization of the training data of the DeepSeek comprehensive analysis model. Following the optimization strategy outlined in step S4.1, optimization operations are performed categorized by type. For software code optimization, this strategy includes algorithm improvement and resource allocation adjustment: algorithm improvement replaces inefficient algorithmic logic within software modules with more efficient algorithms or optimizes and refactors existing algorithms; resource allocation adjustment reallocates the proportion of hardware resources used during software operation based on the operational needs of each software module. The implementation of the software code optimization strategy is integrated into the continuous integration and continuous deployment process. Once the optimization strategy is determined, the code modification, compilation, and deployment processes are automatically triggered without manual intervention, ensuring efficient execution of optimization operations.

[0059] For hardware parameter adjustment strategies, remote automated configuration is achieved through a hardware management interface. This interface establishes communication connections with each hardware component and remotely sends parameter configuration commands based on the parameter adjustment requirements specified in the optimization strategy, automatically completing the hardware parameter adjustments and avoiding the inefficiencies and configuration errors associated with manual operation. For hardware component upgrade recommendations, the corresponding hardware components are replaced according to the recommendations, completing the hardware system upgrade.

[0060] After the optimization operation is completed, the corresponding test scenario in step S3.1 is executed again, using the same data collection method and evaluation criteria as before. Hardware performance data and software operation data are collected in real time during the test to obtain the optimized system performance data. The optimized system performance data is then compared and analyzed with the unoptimized system performance data, focusing on the changes in system performance indicators to determine whether the optimization operation achieved the expected results.

[0061] If the comparative analysis results show that the system performance meets the user's needs, the optimization process ends; if the performance does not meet the expected requirements, based on the new test data collected after this optimization, return to step S4.1, and the DeepSeek comprehensive analysis model will re-analyze the data, identify unresolved or newly emerging problems, formulate targeted optimization strategies, and execute optimization operations and test scenarios again to repeat the above optimization process.

[0062] After each tuning operation, all new test data (including tuned hardware performance data, software operation data, and system performance data) and tuning results (including tuning strategy details, optimization operation details, performance changes, etc.) generated during the tuning process are uniformly recorded in the historical dataset. This newly added historical data will serve as supplementary training data for the DeepSeek comprehensive analysis model, used to further optimize the model's training effect, improve the model's adaptability and prediction accuracy in complex hardware and software collaborative scenarios, and form a closed-loop optimization mechanism of "data collection - model training - scenario testing - tuning iteration - data supplementation".

[0063] In some embodiments, software code optimization strategies may also include operations such as code refactoring and redundant code removal, which further improve software operating efficiency by optimizing the structure of software code and reducing the amount of code; the hardware management interface can support multiple communication protocols, and select the appropriate communication protocol according to the type and brand of hardware components to ensure the compatibility and stability of remote automated configuration.

[0064] The DeepSeek-based hardware-software co-optimization method provided in this embodiment offers a high-quality data foundation for model analysis by comprehensively collecting and standardizing relevant hardware and software data. It constructs a well-structured and fully trained DeepSeek comprehensive analysis model to accurately learn the interaction relationships and performance impact patterns between hardware and software, providing reliable technical support for test scenario generation and optimization strategy formulation. By generating diverse test scenarios and executing them automatically, it comprehensively verifies the system performance under different hardware and software configurations. Through targeted optimization strategies and iterative optimization, it continuously improves system performance. This method achieves hardware-software co-optimization, breaking the limitations of traditional independent hardware and software optimization. It allows hardware resources to better adapt to software operation requirements, and software performance to better leverage hardware capabilities, effectively improving the overall stability and reliability of the system. Simultaneously, the highly automated process reduces manual intervention, improves testing and optimization efficiency, and avoids resource waste. Furthermore, by continuously supplementing historical datasets and optimizing the DeepSeek comprehensive analysis model, the method exhibits good adaptability and scalability, enabling it to adapt to different types of computer systems and diverse performance optimization needs. Among them, the realization of software and hardware co-optimization is the core technical effect of this method. Through the close connection and data exchange of each step, it truly improves performance from the overall system level. This effect is directly achieved through the core process of "collecting and preprocessing relevant software and hardware data - building models and learning the laws of software and hardware interaction and performance impact - generating and executing test scenarios to collect test data - formulating and implementing tuning strategies and iteratively optimizing".

[0065] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A hardware / software co-tuning method based on DeepSeek, characterized in that, Includes the following steps: S1. Collect hardware and software related data from the computer system and preprocess the collected raw data; S2. Construct a DeepSeek comprehensive analysis model. Input the preprocessed data into the DeepSeek comprehensive analysis model and learn the interaction relationship between hardware and software and their impact on system performance through the DeepSeek comprehensive analysis model. S3. Based on user needs and system characteristics, use the DeepSeek comprehensive analysis model to generate test scenarios and execute them automatically, collecting hardware and software data in real time during the test process; S4. Based on the output of the DeepSeek comprehensive analysis model, formulate hardware and software optimization strategies, automatically implement the optimization strategies, and then run the test scenario again. Compare the system performance data before and after optimization, and iterate until the system performance meets the requirements.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Set up multiple data acquisition points within the computer system to collect hardware performance data and software operation-related data. Hardware components include CPU, GPU, memory, and hard disk. Hardware performance data includes the temperature, usage rate, frequency, and read / write speed of these components. Software operation-related data includes application operation logs, system call counts, and resource usage. S1.

2. Clean the collected raw data, remove noisy data and outliers, and convert different types of data into a unified format to prepare for subsequent DeepSeek comprehensive analysis model analysis.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. A multi-layer neural network structure is constructed to form the DeepSeek comprehensive analysis model. The structure includes an input layer, a feature encoding layer, a deep feature extraction layer, a hardware-software interaction relationship learning module, a system performance impact law learning module, a multi-task learning module, and an output layer. The input layer receives hardware data and software data. The feature encoding layer encodes the received data and then passes it to the deep feature extraction layer. The deep feature extraction layer extracts data features and then passes them to the hardware-software interaction relationship learning module. The hardware-software interaction relationship learning module learns the data interaction relationship and then passes it to the system performance impact law learning module. The system performance impact law learning module, in conjunction with the multi-task learning module, processes the data and then outputs the system performance prediction results and interpretability analysis results through the output layer. S2.

2. Collect historical test data of different types of computer systems. The historical test data includes hardware configuration information, software operation status and corresponding system performance indicators. Divide the historical test data and the actual test data into training set and test set. Train the DeepSeek comprehensive analysis model through supervised learning and reinforcement learning. In supervised learning, the hardware and software configuration combinations and corresponding system performance indicators in the historical data are used as labels to adjust the model parameters. In reinforcement learning, the improvement of system performance is used as a reward signal to optimize the model decision. The model parameters are continuously adjusted to optimize the prediction accuracy of the DeepSeek comprehensive analysis model.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: The S3.1.DeepSeek comprehensive analysis model generates diverse test scenarios based on the learned hardware-software interaction patterns and user needs. These test scenarios include different hardware load conditions, software application combinations, and system operating environment settings. S3.

2. Execute the generated test scenarios in a simulated environment using automated testing tools. At the same time, use hardware monitoring tools and software automation tools to comprehensively collect hardware performance data and software operation data during the testing process. Feed the collected data back to the DeepSeek comprehensive analysis model for further optimization.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: The S4.1.DeepSeek comprehensive analysis model analyzes the feedback test data. If it identifies a performance bottleneck in a certain software module under the current hardware configuration, it formulates a software code optimization strategy. If problems are found in the coordination between hardware components, a hardware parameter adjustment strategy or hardware component upgrade recommendation will be formulated. S4.

2. Perform optimization operations according to the established tuning strategy. After completion, execute the corresponding test scenario again, compare and analyze the system performance data before and after tuning. If the performance does not meet the expected requirements, formulate a new tuning strategy based on the new test data and repeat the optimization process.

6. The method according to claim 2, characterized in that, In step S1.2, normalization technology is used to convert the format of numerical data, and preprocessing is performed on text log data. The preprocessing operations include word segmentation, removal of words with no practical meaning and text standardization. Then, word embedding technology is used to convert the text into a fixed-length vector representation. Word embedding technology includes continuous bag-of-words model and skip word model. The training process of word embedding technology includes building a vocabulary, initializing word vectors and iteratively training and adjusting word vectors until the model converges.

7. The method according to claim 3, characterized in that, In step S2.1, the feature encoding layer uses numerical normalization or category encoding to process hardware data and word embedding technology to process software data; the deep feature extraction layer includes a basic feature extraction component, a temporal data processing component, and an interactive learning mechanism. The basic feature extraction component is a fully connected layer, the temporal data processing component can be equipped with a convolutional neural network or a long short-term memory network, and the interactive learning mechanism is a multi-head attention mechanism; the hardware-software interaction relationship learning module uses a graph neural network and a causal inference network. The graph neural network treats hardware and software as nodes and the interaction relationship as an edge to learn the global relationship. The causal inference network analyzes the causal impact of hardware and software configuration on system performance; the multi-task learning module includes a system performance prediction task, an anomaly detection task, and a hardware-software configuration recommendation task.

8. The method according to claim 4, characterized in that, In step S3.1, the DeepSeek comprehensive analysis model samples from the hardware configuration space and software parameter space through algorithms. The hardware configuration space includes CPU frequency and memory capacity, while the software parameter space includes the number of threads and caching strategy. The generated test scenarios are output in the form of configuration files, which specify the hardware parameter settings and software startup parameters. The sampling process combines the learned hardware and software interaction rules to ensure scenario diversity.

9. The method according to claim 5, characterized in that, In step S4.2, the software code optimization strategy includes algorithm improvement and resource allocation adjustment, which are implemented and integrated into the continuous integration and continuous deployment process, automatically triggering code modification, compilation and deployment; the hardware parameter adjustment strategy is implemented through the hardware management interface to achieve remote automated configuration. After each tuning operation is completed, the new test data and tuning results are recorded in the historical dataset for further optimization of the training data of the DeepSeek comprehensive analysis model.

10. A hardware / software co-tuning system based on DeepSeek, used to execute the method described in any one of claims 1-9, characterized in that, It includes a data acquisition module, a data preprocessing module, a DeepSeek analysis model module, a test scenario generation and execution module, and an optimization strategy formulation and implementation module; The data acquisition module is used to set up multiple data acquisition points to collect hardware performance data and software operation-related data; The data preprocessing module is used to clean, normalize, and convert the format of the collected raw data; The DeepSeek analysis model module is used to build a multi-layer neural network structure. Through supervised learning and reinforcement learning, it forms a comprehensive DeepSeek analysis model, learning the interaction between hardware and software and the performance impact rules. The test scenario generation and execution module is used to generate diverse test scenarios using the comprehensive DeepSeek analysis model and execute them automatically, collecting test data in real time and providing feedback. The optimization strategy formulation and implementation module is used to formulate optimization strategies based on the output of the comprehensive DeepSeek analysis model, automatically implement and iteratively optimize until the system performance meets the requirements.

Images