Deep learning-based computing power performance dynamic allocation optimization method and system

Abstract

The application discloses a deep learning-based computing performance dynamic allocation optimization method and system, comprising: constructing a generative adversarial network combined with an improved Transformer model architecture, the generative model of the generative adversarial network generates a preliminary computing power allocation scheme in line with the actual situation with the help of a constraint loss function and a task dependency weight matrix; the discriminant model adopts a gradient-based update strategy to improve the discriminant ability, and the improved Transformer model is optimized through position coding dynamic adjustment and introduction of a regularization term, and accurately processes computing power data generation strategies. The units in the system cooperate, the data acquisition feedback unit acquires computing power data in real time, the driving model is periodically optimized, and the allocation instruction conversion unit implements the strategy combined with the fluctuation compensation factor. Finally, the dynamic and accurate allocation of computing power resources is realized, the application effectively overcomes the problems of low resource utilization and slow response of the traditional allocation method, and significantly improves the computing power performance and system operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer computing power allocation, and in particular to a method and system for dynamic allocation and optimization of computing power performance based on deep learning. Background Technology

[0002] With the accelerating pace of digital transformation, the demand for computing resources from applications such as artificial intelligence, big data analytics, and cloud computing is increasing daily, while the diversity, complexity, and real-time requirements of computing tasks are constantly rising. Traditional computing power allocation methods are gradually revealing numerous drawbacks in addressing these changes, failing to meet the demands of modern computing environments for efficient resource utilization and flexible allocation. Against this backdrop, deep learning, with its powerful feature learning and data processing capabilities, has opened up new paths for optimizing computing performance, becoming a key technological means to overcome the bottlenecks of traditional computing power allocation.

[0003] Existing computing power allocation technologies have significant shortcomings. Firstly, resource allocation lacks dynamic flexibility. Traditional methods typically employ static or semi-static allocation strategies, allocating resources based on pre-defined fixed resource requirements for the task. This approach cannot adapt to the dynamically changing computing power demands of a task during its execution, leading to resource waste or inadequacy. For example, in the early stages of a data processing task, the computing power requirement is low, but as the data volume increases, the demand surges, and the statically allocated resources cannot be adjusted in time, resulting in low task execution efficiency.

[0004] Secondly, it has poor adaptability to complex computing environments. Traditional solutions struggle to effectively handle multi-dimensional and dynamically changing computing parameters, failing to delve into the complex relationships between parameters such as CPU utilization, memory usage, and network bandwidth. When faced with the differentiated computing resource requirements of different types of computing tasks, it is difficult to achieve precise resource matching, resulting in an inability to reasonably and evenly allocate computing resources among various computing nodes. This severely restricts the improvement of overall computing performance and makes it difficult to meet the demands of complex and ever-changing modern computing scenarios. Summary of the Invention

[0005] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a method and system for dynamic allocation and optimization of computing power performance based on deep learning.

[0006] The technical solution adopted in this invention is a deep learning-based method for dynamically allocating and optimizing computing power performance, comprising the following steps:

[0007] S1: Construct a generative adversarial network architecture that includes a generative model and a discriminative model, and use multi-layer convolutional and fully connected layers for feature extraction and classification for source discrimination;

[0008] S2: The initial computing power allocation scheme generated by the generative adversarial network and the computing power state parameters at the corresponding time are input into the improved Transformer model to optimize the multi-head attention mechanism, add a dynamic adjustment mechanism for position encoding, and perform feature fusion and sequence modeling on the input data.

[0009] S3: In the encoding stage of the improved Transformer model, self-attention calculation is performed on the input computing power related data. By adjusting the attention weight matrix, the model focuses on the key parameter dimensions that have a significant impact on computing power allocation. At the same time, a feedforward neural network is used to perform nonlinear transformation on the fused features.

[0010] S4: In the decoding stage of the improved Transformer model, based on the feature vector output from the encoding stage and combined with the current computing power demand prediction index, a refined computing power allocation strategy is generated. This strategy represents the resource allocation ratio of each computing node in vector form.

[0011] S5: Based on the computing power allocation strategy vector output by the improved Transformer model, and convert it into actual computing power allocation instructions, numerical allocation quotas are set for the specific computing power resources of different computing nodes.

[0012] S6: During the operation of the computing power system, the actual computing power usage data of each computing node is collected in real time, and the actual computing power usage data is fed back to the discriminant model of the generative adversarial network to update the parameters of the discriminant model;

[0013] S7: Periodically input the updated discriminant model parameters and the latest computing power operation data back into the generative adversarial network's generative model and the improved Transformer model, triggering the model's retraining and optimization process, and continuously updating and adapting the dynamic allocation scheme of computing power performance.

[0014] Furthermore, the generative adversarial network's generative model incorporates a constraint loss function based on computational power parameters. The expression is ,in, The number of types of computing power parameters. For the first Weighting coefficients for various computing power parameters For the first A real computing power parameter value, The corresponding computing power parameter values ​​generated for the model.

[0015] Furthermore, the improved Transformer model refines the position encoding, resulting in a new position encoding function. Defined as ,in, For traditional positional coding functions at position Dimensions The value, To adjust the coefficient, The number of features related to computing power, For the position pos, the first The value of each computing power characteristic.

[0016] Furthermore, an intermediate fusion layer is introduced between the generative adversarial network and the improved Transformer model, which utilizes a fusion matrix. Feature vectors of the initial computational power allocation scheme for the output of the generative adversarial network The feature vector output from the encoding stage of the improved Transformer model The feature vectors are fused together. ,in, This represents the concatenation of two vectors, achieved by adjusting the fusion matrix. The parameters are used to integrate the characteristics of different stages and optimize the generation of the final computing power allocation strategy.

[0017] Furthermore, during the training of the discriminative model of the generative adversarial network, an update strategy based on the gradient of computing power parameters is adopted to update the loss value output by the discriminative model regarding the discrimination between real and generated computing power allocation data. Calculate different computing power parameters gradient Adjust the update step size of the corresponding parameters of the discriminant model according to the gradient magnitude. ,in, This is the step size coefficient.

[0018] Furthermore, the improved Transformer model introduces a regularization term based on computing resource balancing when generating the computing power allocation strategy vector. The expression is ,in, To calculate the number of nodes, For the number of computing resource types, For the first The computing node was allocated to the first The amount of computing resources is constrained by this regularization term, which makes the allocation of computing resources among computing nodes more balanced.

[0019] Furthermore, when generating the initial computing power allocation scheme, the generative adversarial network's generative model incorporates the dependency weight matrix of different computing tasks on computing power parameters. Output of the generative model pass The calculation yielded that, To calculate the number of task types, To generate the model for the first Task types in the input noise vector The initial allocation scheme components generated below.

[0020] Furthermore, when converting the computing power allocation strategy vector output by the improved Transformer model into actual allocation instructions, a computing power fluctuation compensation factor is introduced. For computing nodes A certain computing power resource Allocation quota Actual allocated amount ,in, This represents the historical standard deviation of the computing power resources of this computing node.

[0021] Furthermore, during the periodic retraining of the generative adversarial network and the improved Transformer model, a sample weighting strategy based on computational task complexity is adopted. For samples from different computational tasks, the weighting is adjusted according to the computational complexity index of the task. Assign weights ,in, and To adjust the parameters, e It is a natural constant.

[0022] A deep learning-based dynamic allocation and optimization system for computing power performance includes:

[0023] Generative Adversarial Network (GAN) building blocks are used to construct a GAN architecture that includes a generative model and a discriminative model. The generative model generates an initial computing power allocation scheme, and the discriminative model performs source discrimination.

[0024] The data input fusion unit is used to input the initial scheme generated by the generative adversarial network and the computing power state parameters into the improved Transformer model, and to perform feature fusion between the generative adversarial network and the improved Transformer model;

[0025] The Transformer encoding processing unit utilizes an improved Transformer model to perform self-attention calculation and feature enhancement processing on the input data during the encoding stage.

[0026] The Transformer decoding generation unit generates a refined computing power allocation strategy during the decoding stage of the improved Transformer model, based on the results of the encoding stage and the estimated computing power demand indicators.

[0027] The allocation instruction conversion unit converts the policy vectors generated by the Transformer into allocation instructions for specific computing resources;

[0028] The real-time data acquisition and feedback unit collects computing power usage data of computing nodes in real time and feeds it back to the discriminative model of the generative adversarial network.

[0029] The model update and optimization unit periodically retrains and optimizes the generative adversarial network and the improved Transformer model based on feedback data and the latest information. The units are connected in sequence to work together to dynamically allocate and optimize computing power performance.

[0030] Beneficial Effects: This invention proposes a method and system for dynamic allocation and optimization of computing power performance based on deep learning. It deeply integrates generative adversarial networks (GANs) with an improved Transformer model, achieving efficient dynamic resource allocation around various computing power parameters. At the methodological level, the GAN, through a game-like adversarial process between a generative model and a discriminative model, can quickly generate preliminary computing power allocation schemes and continuously optimize them based on real-world data, ensuring the schemes closely match actual application scenarios. The improved Transformer model innovatively optimizes multi-head attention mechanisms and positional encoding, accurately capturing the time-series characteristics of computing power data and deeply mining the correlations between various computing power parameters, enabling accurate prediction of computing power needs and generation of refined allocation strategies. Simultaneously, a periodic model update mechanism combined with real-time computing power data feedback allows the system to adapt to dynamic changes in the computing power environment in a timely manner. At the system level, all functional units collaborate closely, forming a complete closed loop from data acquisition and model processing to instruction execution and feedback optimization. This method and system effectively solve the problems of low resource utilization and lagging allocation strategies in traditional computing power allocation, significantly improve the efficiency of computing power resource utilization, reduce the execution latency of computing tasks, enhance the system's ability to cope with complex and ever-changing computing power demands, enable computing power resources to be allocated in a balanced and efficient manner among different computing tasks and nodes, provide reliable resource guarantees for various application scenarios with stringent computing power requirements, and promote the development of computing power infrastructure towards intelligence and efficiency. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall steps of the present invention;

[0032] Figure 2 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0033] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, the deep learning-based method for dynamic allocation and optimization of computing power performance includes the following steps:

[0035] S1: Construct a generative adversarial network architecture that includes a generative model and a discriminative model. The generative model takes a random noise vector as the initial input and generates a preliminary computing power allocation scheme through multiple nonlinear transformation layers. The discriminative model uses multiple convolutional and fully connected layers to perform source discrimination feature extraction and classification processing on the input real computing power allocation data and the scheme output by the generative model.

[0036] Specifically, in step S1, constructing the generative adversarial network (GAN) architecture is the initial step of the entire method. The GAN consists of a generative model and a discriminative model. The generative model takes a random noise vector as input, which can be considered a type of irregular data signal. Its dimensions, distribution, and other parameters can be adjusted according to actual needs, for example, it can be set to 128 dimensions or 256 dimensions. Through multiple nonlinear transformation layers, such as common fully connected layers or convolutional layers, the generative model progressively processes the noise vector, transforming it into a preliminary computing power allocation scheme. These schemes cover the allocation of computing resources such as the number of CPU cores, memory capacity, and network bandwidth for different computing nodes. For example, allocating 8 CPU cores, 16GB of memory, and 100Mbps of network bandwidth to a certain node.

[0037] The discriminative model plays a crucial role. It receives real computing power allocation data from actual operating computing systems, recording the actual resource allocation of each computing node under different tasks and at different times, including various detailed computing power parameters. Simultaneously, the discriminative model also receives preliminary schemes output by the generative model. Utilizing multiple convolutional and fully connected layers, the discriminative model performs feature extraction and classification processing on the input data. Convolutional layers extract local features from the data, such as local patterns of resource allocation between computing nodes, while fully connected layers synthesize these local features to determine the source of the data—that is, whether the data comes from allocation records of a real system or a simulated scheme generated by the generative model.

[0038] The significance of this step lies in the fact that, through the adversarial game between the generative model and the discriminative model, the generative model can continuously optimize its generated computing power allocation scheme. In this process, the generative model gradually learns the distribution characteristics and patterns of real computing power allocation data, making the generated scheme more in line with the needs of actual scenarios. This lays the foundation for generating more accurate computing power allocation strategies in the future, thereby improving overall computing power performance and avoiding unreasonable allocation and waste of resources.

[0039] S2: The initial computing power allocation scheme generated by the generative adversarial network and the computing power status parameters at the corresponding time, such as CPU utilization, memory usage, and network bandwidth usage, are input into the improved Transformer model. The improved Transformer model optimizes the multi-head attention mechanism and adds a dynamic adjustment mechanism for position encoding on the basis of the traditional architecture, so that it can better adapt to the time series characteristics of computing power data and perform feature fusion and sequence modeling on the input data.

[0040] Specifically, step S2 inputs the initial computing power allocation scheme generated by the generative adversarial network, along with the computing power status parameters at the corresponding moment, into the improved Transformer model. The initial computing power allocation scheme is a resource allocation plan obtained by training the generative model, while the computing power status parameters reflect the actual operating status of the current computing power system, including key indicators such as CPU utilization (e.g., the current node's CPU is at different usage levels such as 30% or 50%), memory occupancy (the proportion of memory space used to the total capacity), and network bandwidth usage (the amount of data transmitted over the network per unit time). These parameters can comprehensively present the resource usage status and load of the computing power system at a certain moment.

[0041] The improved Transformer model optimizes the traditional architecture, focusing on adjustments to the multi-head attention mechanism and the addition of a dynamic positional encoding adjustment mechanism. The multi-head attention mechanism allows the model to focus on and process input data from multiple different angles and dimensions. For example, it can simultaneously monitor the correlation between CPU and memory resources on different computing nodes, as well as their changing trends at different points in time. The dynamic positional encoding adjustment mechanism is specifically designed for the time-series characteristics of computing power data. Because computing power requirements and resource usage change continuously over time, this mechanism allows the model to better capture the temporal characteristics of the data, such as the periodic changes in computing power requirements for certain tasks within different time periods.

[0042] The initial plan and computing power status parameters are input into the improved Transformer model. The aim is to perform deep feature fusion and sequence modeling on this data. Through model processing, potential correlations and patterns in the data can be uncovered. For example, the intrinsic relationship between certain task types and specific computing power resource requirements can be discovered, thus providing a basis for generating more accurate and reasonable computing power allocation strategies, further optimizing computing power performance, and improving resource utilization efficiency.

[0043] S3: In the encoding stage of the improved Transformer model, self-attention calculation is performed on the input computing power related data. By adjusting the attention weight matrix, the model focuses on the key parameter dimensions that have a significant impact on computing power allocation. At the same time, a feedforward neural network is used to perform nonlinear transformation on the fused features to enhance the feature expression capability.

[0044] Specifically, in the improved Transformer model encoding stage of step S3, self-attention computation and feedforward neural network processing are the core operations. During self-attention computation, the model analyzes the input computing power-related data and adjusts the attention weight matrix to focus on key parameter dimensions that have a significant impact on computing power allocation. For example, in some computing scenarios, the number of CPU cores and memory capacity may be key factors determining task execution efficiency. The model will give these parameters higher attention weights, making them play a more important role in the computation process, while relatively reducing the attention paid to some secondary parameters.

[0045] Simultaneously, the feedforward neural network performs a non-linear transformation on the fused features. The feedforward neural network contains multiple hidden layers, each processing the input data through an activation function. This non-linear transformation enhances the expressive power of the features, transforming raw computing power data into more abstract and representative feature vectors. For example, it transforms raw data such as CPU utilization and memory usage into features containing complex semantic information, which can more accurately reflect the operating status and task requirements of the computing system.

[0046] The significance of this step lies in the fact that, through the synergy of self-attention computation and feedforward neural networks, the input computing power data is deeply processed and refined, highlighting key information and enhancing the expressive power of features. This enables the model to better understand the inherent logic and patterns of the computing power data, providing high-quality feature representations for generating refined computing power allocation strategies in the subsequent decoding stage. This helps improve the accuracy and rationality of computing power allocation, thereby enhancing computing performance and ensuring the efficient execution of computing tasks.

[0047] S4: In the decoding stage of the improved Transformer model, based on the feature vector output from the encoding stage, combined with the current computing power demand prediction indicators, such as task priority weight and expected task execution time, a refined computing power allocation strategy is generated. This strategy represents the resource allocation ratio of each computing node in vector form.

[0048] Specifically, step S4 is the decoding stage of the improved Transformer model. This stage generates a refined computing power allocation strategy based on the feature vector output from the encoding stage and the current computing power demand prediction metrics. The feature vector output from the encoding stage is the result of deep processing of the input computing power data, containing complex correlation information and patterns in the data. The computing power demand prediction metrics, such as task priority weights (different tasks are assigned different weights according to their importance, with high-priority tasks requiring priority computing power resources) and estimated task execution time (estimated based on historical data or task characteristics), provide additional decision-making basis for the decoding stage.

[0049] During the decoding process, based on the feature vectors from the encoding stage and the estimated computing power requirements, the model generates a computing power allocation strategy represented as a vector through a series of calculations and processing. Each element in this vector corresponds to a certain proportion of computing resources allocated to different computing nodes. For example, the first element in the vector represents the CPU core allocation proportion for node 1, the second element represents the memory allocation proportion for node 1, and so on. In this way, the model can formulate detailed and accurate allocation schemes for different computing nodes and types of computing resources.

[0050] The refined computing power allocation strategy generated in this step is of great significance. It enables precise allocation of computing resources based on the actual state of the current computing system and task requirements. This avoids over-allocation or under-allocation of resources, ensuring that each computing node receives the most suitable computing resources for the current task, thereby improving overall computing performance, accelerating task execution efficiency, and enhancing system stability and reliability.

[0051] S5: Based on the computing power allocation strategy vector output by the improved Transformer model, it is transformed into actual computing power allocation instructions, and numerical allocation quotas are set for specific computing power resources such as the number of CPU cores, memory capacity, and network bandwidth of different computing nodes.

[0052] Specifically, step S5 involves converting the computing power allocation strategy vector output by the improved Transformer model into actual computing power allocation instructions. The computing power allocation strategy vector is an abstract mathematical representation that describes the proportional relationship of computing power resources that should be allocated to each computing node in vector form. The actual computing power allocation instructions need to be specifically implemented in terms of various computing power resources of the computing nodes, such as the number of CPU cores, memory capacity, and network bandwidth.

[0053] During the transformation process, the system maps the proportion values ​​in the policy vector to specific numerical values ​​according to preset rules and algorithms. For example, if the CPU allocation ratio of a node in the policy vector is 0.3, and the entire system has 100 CPU cores, then the actual number of CPU cores allocated to that node is 30. The allocation of resources such as memory and network bandwidth follows a similar calculation process. At the same time, the system also considers factors such as the hardware limitations and compatibility of computing nodes to ensure that the allocated resources are within the node's capacity and that resource conflicts and other problems do not occur.

[0054] This step transforms the abstract allocation strategy into executable operations, enabling computing resources to be accurately allocated to each computing node according to the planned scheme. It is a crucial link connecting theoretical strategies and practical applications. Only by accurately translating strategies into instructions can we truly achieve effective management and allocation of computing resources, thereby improving computing performance and ensuring that computing tasks can be executed smoothly and efficiently in the actual operating environment.

[0055] S6: During the operation of the computing power system, real-time data on the actual computing power usage of each computing node is collected, including but not limited to CPU clock frequency fluctuations, memory read and write speeds, and network data packet transmission volume. This real-time data is fed back to the discriminative model of the generative adversarial network to update the parameters of the discriminative model and improve its ability to distinguish between the generated scheme and real data.

[0056] Specifically, in step S6, the core operation is to collect real-time data on the actual computing power usage of each computing node. This data covers a wealth of information, including but not limited to CPU clock frequency fluctuations (reflecting changes in the CPU's operating frequency during operation, such as fluctuations from 2.0GHz to 3.0GHz), memory read / write rates (the speed at which data is read from and written to memory per unit time), and network packet transmission volume (the number of data packets transmitted in the network per unit time). This data can reflect the resource usage status and load of the computing nodes in real time and accurately.

[0057] The collected real-time data is fed back into the discriminative model of the generative adversarial network (GAN). The discriminative model uses this new data to update its parameters. This update process is based on machine learning optimization algorithms, which adjust the values ​​of various parameters in the model by calculating the data error and gradient. For example, based on the difference between the new data and the model's predictions, the weight parameters of the convolutional and fully connected layers are adjusted, thereby improving the discriminative model's ability to distinguish between generated schemes and real data.

[0058] The significance of this step lies in enabling the generative adversarial network (GAN) to promptly perceive the actual operational status of the computing power system through real-time data collection and feedback. After updating its parameters, the discriminant model can more accurately distinguish between real data and the schemes generated by the generative model, thereby prompting the generative model to generate computing power allocation schemes that better meet actual needs. This feedback mechanism allows the entire system to continuously adapt to the dynamic changes in the computing power environment, improving the accuracy and effectiveness of computing power allocation, thereby continuously optimizing computing power performance and ensuring that the system can operate efficiently under different workloads.

[0059] S7: Periodically input the updated discriminant model parameters and the latest computing power operation data back into the generative adversarial network's generative model and the improved Transformer model to trigger the model's retraining and optimization process, so as to achieve continuous updating and adaptation of the dynamic allocation scheme of computing power performance.

[0060] Specifically, step S7 is a periodic model update and optimization process. The updated discriminant model parameters and the latest computing power operation data are then input back into the generative model and the improved Transformer model of the generative adversarial network. The updated discriminant model parameters reflect its improved ability to distinguish between real and generated data after receiving new data, while the latest computing power operation data contains the latest status information of the computing power system, such as changes in the current task and updates to the resource usage of each node.

[0061] When this data is input into the generative model and the improved Transformer model, it triggers a retraining and optimization process. During training, the model adjusts its weights and structure based on the new data and parameters to better adapt to the new computing environment and task requirements. For example, the generative model adjusts its generation rules based on new discriminative feedback to make them closer to reality; the improved Transformer model optimizes its attention mechanism and positional encoding, improving its ability to process computing data and the accuracy of its allocation strategies.

[0062] This step is crucial for the continuous updating and adaptation of the dynamic computing power allocation scheme. As the computing system operates, task requirements and resource status constantly change. Through periodic model updates and optimizations, the system can keep up with these changes in a timely manner, continuously improving and refining the computing power allocation scheme. This ensures that computing resources are always allocated in the optimal way during long-term operation, continuously improving computing performance, meeting the ever-changing computing task demands for computing resources, and guaranteeing the efficient and stable operation of the system.

[0063] Preferably, a constraint loss function based on computational power parameters is introduced into the generative model of the generative adversarial network. The expression is ,in, The number of types of computing power parameters. For the first Weighting coefficients for various computing power parameters For the first A real computing power parameter value, The corresponding computing power parameter values ​​generated for the generation model are constrained by this loss function to make the computing power allocation scheme generated by the generation model more closely match the actual computing power parameter range.

[0064] Specifically, a constraint loss function based on computational power parameters is introduced into the generative model of generative adversarial networks (GANs) to constrain the computational power allocation scheme generated by the generative model. This loss function optimizes the generative model by calculating the difference between the actual computational power parameter values ​​and the corresponding values ​​output by the generative model, and combining this with the weight coefficients of each computational power parameter. For example, in a system containing multiple computational power parameters such as the number of CPU cores, memory capacity, and network bandwidth, different weight coefficients are set according to the importance of each parameter to the computational power allocation. If the number of CPU cores has a greater impact on task execution, it can be assigned a higher weight. Through this loss function, the computational power allocation scheme generated by the generative model can more closely match the actual range of computational power parameters, avoiding the generation of unrealistic allocation schemes, improving the feasibility and effectiveness of the generated schemes, thereby optimizing computational power performance and reducing resource waste and low task execution efficiency caused by unreasonable allocation.

[0065] In the preferred and improved Transformer model, the position encoding is improved, and a new position encoding function is used. Defined as ,in, For traditional positional coding functions at position Dimensions The value, To adjust the coefficient, The number of features related to computing power, For the position pos, the first The value of each computing power feature is used by this function to enable the location encoding to be dynamically adjusted according to the computing power feature, thereby improving the model's ability to process computing power sequence data.

[0066] Specifically, the positional encoding of the improved Transformer model is enhanced. The new positional encoding function combines traditional positional encoding with computing power-related features, enabling dynamic adjustment of the encoding based on these features. In real-world computing systems, the utilization of computing resources by different nodes changes over time. For example, the CPU utilization and memory read / write speeds of certain nodes may differ within specific time periods. The new positional encoding function incorporates these computing power features into its calculations. By adjusting the combined calculations of coefficients and computing power features, it changes the fixed pattern of traditional positional encoding, allowing the model to better capture the dynamic changes in computing power data over time. This helps the improved Transformer model process computing power sequence data more accurately, improves the accuracy of computing power demand prediction and allocation strategy generation, and ultimately enhances overall computing power performance, making computing power resource allocation more aligned with actual task requirements.

[0067] Preferably, an intermediate fusion layer is introduced between the generative adversarial network and the improved Transformer model, which utilizes a fusion matrix. Feature vectors of the initial computational power allocation scheme for the output of the generative adversarial network The feature vector output from the encoding stage of the improved Transformer model The feature vectors are fused together. ,in, This represents the concatenation of two vectors, achieved by adjusting the fusion matrix. The parameters enable the effective fusion of characteristics at different stages, thereby optimizing the generation of the final computing power allocation strategy.

[0068] Specifically, an intermediate fusion layer is introduced between the Generative Adversarial Network (GAN) and the improved Transformer model, using a fusion matrix to fuse the feature vectors output by the two models. The feature vector of the initial computing power allocation scheme output by the GAN contains allocation pattern information obtained through adversarial learning, while the feature vector output by the encoding stage of the improved Transformer model fuses computing power state parameters and deep features of the initial scheme. By adjusting its own parameters, the fusion matrix can weight and combine the two feature vectors, highlighting key features and weakening secondary features. For example, when dealing with CPU resource-sensitive computing tasks, the fusion matrix can enhance the weight of CPU-related features, making the fused feature vector more reflective of the task's CPU resource requirements. In this way, effective fusion of features from different stages is achieved, providing more comprehensive and accurate feature information for the improved Transformer model to generate a better computing power allocation strategy, thereby improving the rationality of computing power allocation and computing performance.

[0069] Preferably, during the training of the discriminative model of the generative adversarial network, an update strategy based on the gradient of the computing power parameters is adopted to update the loss value output by the discriminative model regarding the discrimination between real and generated computing power allocation data. Calculate different computing power parameters gradient Adjust the update step size of the corresponding parameters of the discriminant model according to the gradient magnitude. ,in, The step size coefficient is used to make the discriminative model pay more attention to the parameter dimensions that are sensitive to computing power allocation, thereby accelerating model convergence.

[0070] Specifically, during the training of the discriminative model in the generative adversarial network (GAN), an update strategy based on the gradient of computational power parameters is adopted. When the discriminative model distinguishes between real and generated computational power allocation data, it generates a loss value. By calculating the gradient of this loss value with respect to each computational power parameter, it can be determined which parameters have a significant impact on the discrimination result. For example, when judging whether a computational power allocation scheme is reasonable, a large gradient in the memory occupancy parameter indicates that memory occupancy has a significant impact on the discrimination result. The update step size of the corresponding parameters in the discriminative model is adjusted according to the gradient magnitude. For parameters with larger gradients, a larger update step size is given to allow them to adapt to data changes more quickly; for parameters with smaller gradients, a smaller update step size is used to avoid over-adjustment. This strategy makes the discriminative model focus more on parameters sensitive to computational power allocation, accelerates the model's convergence speed, improves the discriminative model's ability to distinguish between generated schemes and real data, and ultimately prompts the generative model to generate computational power allocation schemes that better meet actual needs, thus optimizing computational performance.

[0071] The preferred and improved Transformer model introduces a regularization term based on computing resource balancing when generating the computing power allocation strategy vector. The expression is ,in, To calculate the number of nodes, For the number of computing resource types, For the first The computing node was allocated to the first The amount of computing resources is constrained by this regularization term, which makes the allocation of computing resources among computing nodes more balanced.

[0072] Specifically, when generating the computing power allocation strategy vector in the improved Transformer model, a regularization term based on computing resource balancing is introduced. In a system containing multiple computing nodes and various types of computing resources, the regularization term constrains the allocation strategy by calculating the difference between the different amounts of computing resources allocated to each computing node and the average allocation. For example, if there are three computing nodes and three types of computing resources—CPU, memory, and network bandwidth—the regularization term calculates the sum of squared deviations of the allocation ratio of each node on each resource from the average allocation ratio. By minimizing this regularization term, the allocation strategy generated by the improved Transformer model can make the allocation of computing resources among the computing nodes more balanced, avoiding situations where some nodes have excessively concentrated resources while others lack resources. This helps improve the overall utilization efficiency of computing resources, reduces task execution bottlenecks caused by uneven resource allocation, improves computing performance, and ensures the stable operation of the system.

[0073] Preferably, when generating the initial computing power allocation scheme, the generative model of the generative adversarial network incorporates the dependency weight matrix of different computing tasks on computing power parameters. Output of the generative model pass The calculation yielded that, To calculate the number of task types, To generate the model for the first Task types in the input noise vector The initial allocation scheme components generated below are used to make the generated computing power allocation scheme more closely match the computing power requirements of different tasks.

[0074] Specifically, the generative adversarial network (GAN) model considers the dependency weight matrix of different computing tasks on computing power parameters when generating initial computing power allocation schemes. Different types of computing tasks, such as scientific computing, data processing, and graphics rendering, have different requirements for computing power parameters such as the number of CPU cores, memory capacity, and network bandwidth. The dependency weight matrix records the degree of dependence of each task type on different computing power parameters. For example, scientific computing tasks may have a higher dependency on the number of CPU cores and memory capacity, while graphics rendering tasks may rely more on GPU performance and memory bandwidth. When generating initial allocation schemes, the generative model obtains the corresponding weights from the dependency weight matrix according to the task type and performs weighted processing on the generation of different computing power parameters, making the generated initial allocation scheme more closely match the computing power requirements of different tasks. This helps to improve the relevance and effectiveness of the generated schemes, enabling computing resources to more accurately match task requirements, improve computing performance, and accelerate task execution.

[0075] Preferably, when converting the computing power allocation strategy vector output by the improved Transformer model into actual allocation instructions, a computing power fluctuation compensation factor is introduced. For computing nodes A certain computing power resource Allocation quota Actual allocated amount ,in, The standard deviation of the historical fluctuation of this type of computing power resource for this computing node is used as a compensation factor to pre-compensate for possible computing power fluctuations and improve the stability of the allocation scheme.

[0076] Specifically, a computing power fluctuation compensation factor is introduced when converting the computing power allocation strategy vector output by the improved Transformer model into actual allocation instructions. During the actual operation of a computing system, the utilization of computing resources on computing nodes can fluctuate due to factors such as changes in task load and hardware performance. For example, the CPU clock frequency may automatically decrease due to rising temperature, leading to a decline in computing power. The computing power fluctuation compensation factor adjusts the allocation quota based on the historical standard deviation of a computing node's computing resource. For instance, if the historical standard deviation of a computing node's CPU resource fluctuation is large, it indicates that its CPU performance is unstable. When allocating CPU cores, a compensation factor greater than 1 is multiplied to appropriately increase the CPU allocation quota for that node, in order to cope with potential decreases in computing power. Through this compensation mechanism, potential computing power fluctuations can be pre-compensated, improving the stability of the allocation scheme, avoiding task execution interruptions or efficiency reductions caused by computing power fluctuations, thereby improving computing power performance and ensuring the reliable operation of the system.

[0077] Preferably, when periodically retraining the generative adversarial network and the improved Transformer model, a sample weighting strategy based on computational task complexity is adopted. For samples with different computational tasks, the weighting strategy is adjusted according to the computational complexity index of the task. Assign weights ,in, and To adjust the parameters, during model training, weighted loss calculations are performed on samples with different weights, making the model more focused on allocation optimization in complex computing task scenarios.

[0078] Specifically, when periodically retraining the Generative Adversarial Network (GAN) and the improved Transformer model, a sample weighting strategy based on the computational complexity of the task is adopted. Different computational tasks have different computational complexities; for example, complex AI training tasks have high computational complexity, while simple data query tasks have low computational complexity. Samples are assigned weights based on the computational complexity of the task. Samples from tasks with high computational complexity are given higher weights, allowing them to contribute more to the loss calculation during model training; samples from tasks with low computational complexity are given lower weights. For example, during training, a sample from a complex task might have a weight of 0.8, while a sample from a simple task might have a weight of 0.2. Through this weighted loss calculation method, the model focuses more on allocation optimization in complex computational task scenarios during training, improving its ability to generate reasonable computational allocation schemes when handling complex tasks. This helps improve overall computational performance, enabling the system to better handle various complex computational tasks, improving resource utilization efficiency and task execution effectiveness.

[0079] like Figure 2 As shown, the deep learning-based dynamic allocation and optimization system for computing power performance includes:

[0080] Generative Adversarial Network (GAN) building blocks are used to construct a GAN architecture that includes a generative model and a discriminative model. The generative model generates an initial computing power allocation scheme, and the discriminative model performs source discrimination.

[0081] The data input fusion unit is used to input the initial scheme generated by the generative adversarial network and the computing power state parameters into the improved Transformer model, and to perform feature fusion between the generative adversarial network and the improved Transformer model;

[0082] The Transformer encoding processing unit utilizes an improved Transformer model to perform self-attention calculation and feature enhancement processing on the input data during the encoding stage.

[0083] The Transformer decoding generation unit generates a refined computing power allocation strategy during the decoding stage of the improved Transformer model, based on the results of the encoding stage and the estimated computing power demand indicators.

[0084] The allocation instruction conversion unit converts the policy vectors generated by the Transformer into allocation instructions for specific computing resources;

[0085] The real-time data acquisition and feedback unit collects computing power usage data of computing nodes in real time and feeds it back to the discriminative model of the generative adversarial network.

[0086] The model update and optimization unit periodically retrains and optimizes the generative adversarial network and the improved Transformer model based on feedback data and the latest information. The units are connected in sequence to work together to dynamically allocate and optimize computing power performance.

[0087] The Generative Adversarial Network (GAN) construction unit first constructs the GAN, generates a preliminary computing power allocation scheme, and then uses a discriminant model for judgment. The data input fusion unit inputs the preliminary scheme and computing power state parameters into the improved Transformer model, and simultaneously uses an intermediate fusion layer to fuse the features of the GAN and the Transformer model. The Transformer encoding processing unit performs encoding processing on the input data, and the Transformer decoding generation unit generates a refined computing power allocation strategy based on the encoding results and computing power demand prediction indicators. The allocation instruction conversion unit converts the strategy vector into actual allocation instructions and applies them to the computing power system. The real-time data acquisition and feedback unit collects computing power usage data from computing nodes and feeds it back to the discriminant model of the GAN. The model update and optimization unit periodically retrains and optimizes the GAN and the improved Transformer model based on the feedback data and the latest information. These units work closely together and are sequentially connected to form a complete closed loop from scheme generation, data processing, instruction execution to model optimization, realizing dynamic allocation and optimization of computing power performance.

[0088] To address the low resource utilization caused by traditional static computing power allocation, this method and system leverage the collaborative operation of a generative adversarial network (GAN) and an improved Transformer model. The GAN's generative model introduces a constraint loss function based on computing power parameters to generate allocation schemes that closely match the actual range of computing power parameters, avoiding unrealistic resource allocation. Simultaneously, it considers the dependency weight matrix of different computing tasks on computing power parameters, making the generated schemes more tailored to task characteristics. The improved Transformer model, through dynamic adjustments to positional encoding and the introduction of regularization terms for computing power resource balancing, precisely processes computing power sequence data, achieving refined resource allocation. These techniques break away from the traditional fixed allocation model, enabling flexible allocation of computing power resources based on real-time task requirements, effectively avoiding resource idleness and waste, and significantly improving resource utilization.

[0089] To address the issues of slow response and difficulty in adapting to complex environmental changes in traditional computing power allocation, this method and system construct a real-time feedback and dynamic optimization mechanism. The real-time data acquisition and feedback unit continuously collects actual computing power usage data from each computing node, such as CPU clock frequency fluctuations and memory read / write rates, and feeds this data back to the discriminative model of the generative adversarial network (GAN), prompting it to update its parameters. Simultaneously, the model update and optimization unit periodically retrains the GAN and the improved Transformer model based on this data, adjusting model parameters to adapt to new computing power environments. When translating the allocation strategy into actual instructions, a computing power fluctuation compensation factor is introduced to pre-compensate for potential computing power fluctuations. This end-to-end dynamic adjustment capability enables the system to quickly perceive task changes and computing power fluctuations, promptly optimize allocation strategies, significantly improve adaptability to complex and ever-changing computing scenarios, effectively reduce task execution latency, and enhance overall system performance.

[0090] Furthermore, this method and system innovate in terms of inter-model interaction and policy generation. By introducing an intermediate fusion layer between the generative adversarial network and the improved Transformer model, effective fusion of features from different stages is achieved, providing comprehensive information for generating better allocation strategies. In the training of the discriminative model, an update strategy based on the gradient of computational power parameters is adopted to accelerate model convergence and improve the ability to discriminate generated schemes. During model training, a sample weighting strategy based on the complexity of computational tasks is employed to focus on optimizing allocation schemes under complex tasks. These innovative measures further enhance the system's intelligent decision-making capabilities, ensuring efficient and accurate allocation of computing resources in complex tasks and dynamic environments, and driving the advancement of computing infrastructure towards intelligence and efficiency.

[0091] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic allocation and optimization of computing power performance based on deep learning, characterized in that, Includes the following steps: S1: Construct a generative adversarial network architecture that includes a generative model and a discriminative model, and use multi-layer convolutional and fully connected layers for feature extraction and classification for source discrimination; S2: The initial computing power allocation scheme generated by the generative adversarial network and the computing power state parameters at the corresponding time are input into the improved Transformer model to optimize the multi-head attention mechanism, add a dynamic adjustment mechanism for position encoding, and perform feature fusion and sequence modeling on the input data. S3: In the encoding stage of the improved Transformer model, self-attention calculation is performed on the input computing power related data. By adjusting the attention weight matrix, the model focuses on the key parameter dimensions that have a significant impact on computing power allocation. At the same time, a feedforward neural network is used to perform nonlinear transformation on the fused features. S4: In the decoding stage of the improved Transformer model, based on the feature vector output from the encoding stage and combined with the current computing power demand prediction index, a refined computing power allocation strategy is generated. This strategy represents the resource allocation ratio of each computing node in vector form. S5: Based on the computing power allocation strategy vector output by the improved Transformer model, and convert it into actual computing power allocation instructions, numerical allocation quotas are set for the specific computing power resources of different computing nodes. S6: During the operation of the computing power system, the actual computing power usage data of each computing node is collected in real time, and the actual computing power usage data is fed back to the discriminant model of the generative adversarial network to update the parameters of the discriminant model; S7: Periodically input the updated discriminant model parameters and the latest computing power operation data back into the generative adversarial network's generative model and the improved Transformer model, triggering the model's retraining and optimization process, and continuously updating and adapting the dynamic allocation scheme of computing power performance.

2. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, The generative adversarial network introduces a constraint loss function based on computational power parameters in its generative model. The expression is ,in, The number of types of computing power parameters. For the first Weighting coefficients for various computing power parameters For the first A real computing power parameter value, The corresponding computing power parameter values ​​generated for the model.

3. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, The improved Transformer model refines the position encoding, introducing a new position encoding function. Defined as ,in, For traditional positional coding functions at position Dimensions The value, To adjust the coefficient, The number of features related to computing power, For the position pos, the first The value of each computing power characteristic.

4. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, An intermediate fusion layer is introduced between the generative adversarial network and the improved Transformer model, which utilizes a fusion matrix. Feature vectors of the initial computational power allocation scheme for the output of the generative adversarial network The feature vector output from the encoding stage of the improved Transformer model The feature vectors are fused together. ,in, This represents the concatenation of two vectors, achieved by adjusting the fusion matrix. The parameters are used to integrate the characteristics of different stages and optimize the generation of the final computing power allocation strategy.

5. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, During the training of the discriminative model of the generative adversarial network, an update strategy based on the gradient of computing power parameters is adopted to update the loss value output by the discriminative model regarding the discrimination between real and generated computing power allocation data. Calculate different computing power parameters gradient Adjust the update step size of the corresponding parameters of the discriminant model according to the gradient magnitude. ,in, This is the step size coefficient.

6. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, The improved Transformer model introduces a regularization term based on computing resource balancing when generating the computing power allocation strategy vector. The expression is ,in, To calculate the number of nodes, For the number of computing resource types, For the first The computing node was allocated to the first The amount of computing resources is constrained by this regularization term, which makes the allocation of computing resources among computing nodes more balanced.

7. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, When generating the initial computing power allocation scheme, the generative adversarial network's generative model incorporates the dependency weight matrix of different computing tasks on computing power parameters. Output of the generative model pass The calculation yielded that, To calculate the number of task types, To generate the model for the first Task types in the input noise vector The initial allocation scheme components generated below.

8. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, When converting the computing power allocation strategy vector output by the improved Transformer model into actual allocation instructions, a computing power fluctuation compensation factor is introduced. For computing nodes A certain computing power resource Allocation quota Actual allocated amount ,in, This represents the historical standard deviation of the computing power resources of this computing node.

9. The method for dynamic allocation and optimization of computing power performance based on deep learning according to claim 1, characterized in that, When periodically retraining the Generative Adversarial Network and the improved Transformer model, a sample weighting strategy based on computational task complexity is adopted. For samples with different computational tasks, the weighting is adjusted according to the computational complexity index of the task. Assign weights ,in, and To adjust the parameters, e It is a natural constant.

10. A deep learning-based dynamic allocation and optimization system for computing power performance, characterized in that, include: Generative Adversarial Network (GAN) building blocks are used to construct a GAN architecture that includes a generative model and a discriminative model. The generative model generates an initial computing power allocation scheme, and the discriminative model performs source discrimination. The data input fusion unit is used to input the preliminary scheme generated by the generative adversarial network and the computing power state parameters into the improved Transformer model, and to perform feature fusion between the generative adversarial network and the improved Transformer model; The Transformer encoding processing unit utilizes an improved Transformer model to perform self-attention calculation and feature enhancement processing on the input data during the encoding stage. The Transformer decoding generation unit generates a refined computing power allocation strategy during the decoding stage of the improved Transformer model, based on the results of the encoding stage and the estimated computing power demand indicators. The allocation instruction conversion unit converts the policy vectors generated by the Transformer into allocation instructions for specific computing resources; The real-time data acquisition and feedback unit collects computing power usage data of computing nodes in real time and feeds it back to the discriminative model of the generative adversarial network. The model update and optimization unit periodically retrains and optimizes the generative adversarial network and the improved Transformer model based on feedback data and the latest information. The units are connected in sequence to work together to dynamically allocate and optimize computing power performance.

Images