Hyper-converged computing power scheduling method and system based on prediction model

Abstract

The present invention relates to the field of resource scheduling technology, specifically to a hyper-converged computing power scheduling method and system based on a prediction model, comprising the following steps: collecting and organizing real-time data in a hyper-converged environment, including GPU usage and network bandwidth information, removing noise therefrom using digital filtering technology, extracting key performance indicators, numerically standardizing the key performance indicators, and establishing standardized resource characteristic values. In the present invention, through the application of real-time data collection and digital filtering technology, the quality of data processing is improved, the accuracy of resource prediction is ensured, and the use of standardized key performance indicators and time series analysis can accurately predict short-term resource requirements, achieve accurate matching and optimal allocation of resources, and in addition, the neural network model constructed using feedback loops and nonlinear activation functions can dynamically adjust resource configuration, effectively respond to real-time changes in resource requirements, and improve the system's adaptability to workloads and overall data processing capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of resource scheduling, and in particular to a hyper-converged computing power scheduling method and system based on a prediction model. Background Art

[0002] Resource scheduling is a key technical field in computer science, focusing on how to effectively allocate and manage computing resources, including processor time, memory, storage space, and network bandwidth. Research and applications in this field ensure that system resources are used fairly and efficiently in a multi-tasking and multi-user environment. Resource scheduling technology is widely used in operating systems, distributed computing, cloud computing, and big data processing. Its purpose is to optimize resource utilization, reduce task execution time, and improve overall system performance and user satisfaction. Through complex algorithms and strategies, resource scheduling can predict resource requirements, avoid resource conflicts and bottlenecks, and support system scalability and elasticity.

[0003] Among them, the hyper-converged computing power scheduling method refers to a resource scheduling strategy used in hyper-converged infrastructure to optimize and manage the allocation of computing, storage and network resources. This method makes resource scheduling more centralized and automated by integrating software-defined storage, computing and networking into a single system management platform. The purpose of hyper-converged computing power scheduling is to improve the operational efficiency of data centers, reduce complexity and costs, and enhance the system's responsiveness and adaptability to different workloads. By predicting resource requirements through intelligent algorithms, hyper-converged computing power scheduling can dynamically adjust resource allocation to support business continuity and performance optimization.

[0004] Existing technologies are insufficient in real-time data analysis and dynamic resource scheduling, which often leads to a mismatch between resource allocation and actual demand, affecting system efficiency and performance. Due to the lack of rapid response capabilities to real-time data changes, traditional resource scheduling may not be able to adapt to workload fluctuations in a timely manner, resulting in idle or overloaded resources. This static resource management method limits the flexibility and scalability of the system, and sometimes may affect user experience and business continuity due to response delays. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings existing in the prior art and propose a hyper-converged computing power scheduling method and system based on a prediction model.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a hyper-converged computing power scheduling method based on a prediction model, comprising the following steps:

[0007] S1: Collect and organize real-time data in the hyper-converged environment, including GPU usage and network bandwidth information, use digital filtering technology to remove noise, extract key performance indicators, standardize the key performance indicators, and establish standardized resource feature values;

[0008] S2: Perform time series analysis on the standardized resource feature values, build a prediction model to analyze the usage trend of resources in the short term, calculate the demand growth rate of GPU and storage in the time period, and obtain the resource demand prediction value;

[0009] S3: input the resource demand forecast value, construct a primary neural network including a feedback loop and a nonlinear activation function, perform preliminary adjustment of weights and parameters, match the current resource usage pattern, and generate a preliminary neural network model;

[0010] S4: Optimizing the preliminary neural network model by adjusting the depth and width of the network layer, selecting the activation function, adjusting the weight parameter of the loss function, and implementing the small batch gradient descent method for cyclic training to match future changes in resource demand, thereby generating an optimized neural network model;

[0011] S5: Apply the optimized neural network model to implement real-time prediction of resource usage and network load, adjust resource allocation and data processing path according to the prediction results, and obtain an adjusted resource configuration strategy.

[0012] The standardized resource characteristic values ​​specifically include GPU utilization rate and network bandwidth information; the resource demand prediction values ​​include GPU demand growth rate and storage demand growth rate; the preliminary neural network model specifically refers to a neural network structure including a feedback loop and a nonlinear activation function; the optimized neural network model specifically includes the adjusted network layer depth and width, the selected activation function, and the adjusted loss function weight parameters; the adjusted resource allocation strategy specifically includes real-time resource usage prediction and network load prediction.

[0013] As a further solution of the present invention, the step of obtaining the standardized resource characteristic value is specifically as follows:

[0014] S111: Collect GPU usage and network bandwidth information from the hyper-converged environment and organize the data to generate a preliminary data set;

[0015] S112: applying digital filtering technology to the preliminary data set to obtain a cleaned performance indicator data set by eliminating statistical noise and outliers;

[0016] S113: numerically normalizing the cleaned performance indicator data set, converting GPU usage and network bandwidth to a unified ratio and dimension using a Z-score method, and obtaining a standardized performance indicator data set;

[0017] S114: Based on the standardized performance indicator data set, select key performance indicators and calculate indicator weights using the formula:

[0018] ;

[0019] Obtaining standardized resource characteristic values;

[0020] in, represents the normalized value of a single performance indicator, Indicates the business criticality of the performance indicator, Represents the total number of performance indicators, Represents the integrated and standardized resource characteristic value.

[0021] As a further solution of the present invention, the step of obtaining the resource demand forecast value is specifically as follows:

[0022] S211: using the standardized resource characteristic values, segmenting the resource characteristic values ​​based on continuous time periods, rearranging them in chronological order and marking them with timestamps, and establishing time series data;

[0023] S212: Based on the time series data, for each continuous time period, the resource characteristic value is analyzed for the change trend within the time period, and the growth rate of the resource characteristic value in multiple time periods is calculated using the formula:

[0024] ;

[0025] Get the growth rate of GPU and storage;

[0026] in, Indicates the resource growth rate, is the resource characteristic value within the target time period, is the weighting parameter of the time slice, To calculate the adjustment factor, is the offset adjustment parameter, is the noise correction term, Represents the number of time periods referenced in the calculation;

[0027] S213: Based on the growth rates of the GPU and storage, a resource demand growth ratio within the time period is obtained by performing a set operation with resource usage in adjacent time periods to comprehensively calculate;

[0028] S214: calling the resource demand growth rate, based on the resource usage of each time segment, cumulatively calculating the resource growth rate of all segments, and obtaining a resource demand prediction value.

[0029] As a further solution of the present invention, the steps of obtaining the preliminary neural network model are specifically as follows:

[0030] S311: calling the resource demand prediction value, inputting it into the primary neural network, initializing the feedback loop structure and setting the nonlinear activation function, performing random distribution and preliminary setting on multiple connection weights, and generating an initial weight matrix;

[0031] S312: Execute the operation of the feedback loop, adjust the weight parameters in the initial weight matrix based on the feedback output, and combine the dynamic output of the nonlinear activation function response, using the formula:

[0032] ;

[0033] Adjust the weight of each connection and update the current weight matrix to obtain the initially adjusted weights and parameters;

[0034] in, represents the weight value after the next iteration, is the current weight value, is the learning rate, is the output value of the feedback loop, is the input activation strength of the connection, To adjust the deviation factor, is the noise correction term, Represents the total number of elements involved in the sum operation;

[0035] S313: Utilize the preliminarily adjusted weights and parameters, combine them with the resource demand forecast value, perform multiple iterative calculations on multiple weights and parameters, match the current resource usage pattern, and generate a preliminary neural network model.

[0036] As a further solution of the present invention, the step of obtaining the optimized neural network model is specifically as follows:

[0037] S411: Based on the preliminary neural network model, increase or decrease the number of layers and the number of nodes in each layer of the network, and optimize the number of hidden layers and the number of nodes based on resource demand prediction to obtain an adjusted neural network architecture;

[0038] S412: Based on the adjusted neural network architecture, select matching activation functions for multiple hidden layers, and perform a combination test on each layer using a differentiated activation function, using the formula:

[0039] ;

[0040] Calculate the response outputs of multiple nodes to obtain an optimized combination of activation functions;

[0041] in, Represents the output of the neural network layer, which is the total output of the weighted sum after multiple node activation functions are processed. is the node activation function, is the weight coefficient of the activation function, which is used to adjust the strength of the node input signal. is the input signal, is the bias parameter, which is used to adjust the activation threshold of the activation function. Represents the total number of nodes, indicating the number of nodes in the current layer;

[0042] S413: Based on the optimized activation function combination, the weight parameters of the loss function are adjusted, small batch gradient descent training is implemented in combination with the current weight matrix, the neural network error is iteratively calculated, and an optimized neural network model is generated.

[0043] As a further solution of the present invention, the step of obtaining the adjusted resource allocation strategy is specifically:

[0044] S511: Based on the optimized neural network model, the real-time monitoring data is processed to obtain the current resource usage, and a prediction analysis is performed based on the real-time resource usage and the network load to obtain a real-time resource prediction result;

[0045] S512: Based on the real-time resource prediction result, calculate the resource demand change trend, adjust the current resource configuration, and use weighted parameters to smooth the prediction error, using the formula:

[0046] ;

[0047] Get the adjusted resource ratio;

[0048] in, Indicates the adjusted resource allocation, is the weighting coefficient, is the real-time resource usage, To predict the load, is the current configuration capacity, is the noise correction term, Represents the total number of resource types or nodes involved in the calculation;

[0049] S513: Based on the adjusted resource allocation and in combination with the real-time resource prediction result, the data processing paths and resource allocations among the multiple network nodes are reallocated to obtain an adjusted resource configuration strategy.

[0050] A hyper-converged computing power scheduling system based on a prediction model, wherein the hyper-converged computing power scheduling system based on a prediction model is used to execute the hyper-converged computing power scheduling method based on a prediction model, and the system comprises:

[0051] The data preprocessing module collects GPU usage and network bandwidth data in the hyper-converged environment, uses digital filtering technology to remove random noise in the data, screens key performance indicators of GPU usage and network bandwidth, and converts the indicators into standardized resource characteristic values ​​through linear transformation;

[0052] The trend prediction model module uses the standardized resource characteristic values ​​to construct a time series analysis model, evaluate the dynamic changes in recent resource usage, calculate and predict the demand growth rate of GPU and storage in the short term, and thus generate a resource demand forecast value;

[0053] The neural network construction module takes the resource demand prediction value as input, sets the feedback loop in the neural network and selects the nonlinear activation function, preliminarily adjusts the network parameters to match the current resource usage pattern, creates a preliminary neural network model, adjusts the depth, width, and loss function weight of the network layer, and uses a small batch gradient descent method to train and optimize the network to adapt to future changes in resource demand, and obtain an optimized neural network model;

[0054] The resource allocation strategy module deploys the optimized neural network model to perform real-time resource usage and network load prediction, dynamically adjusts resource allocation and data processing paths according to the prediction results, and forms and implements the adjusted resource allocation strategy.

[0055] Compared with the prior art, the advantages and positive effects of the present invention are:

[0056] In the present invention, the quality of data processing is improved and the accuracy of resource prediction is ensured by the application of real-time data collection and digital filtering technology. By using standardized key performance indicators and time series analysis, short-term resource requirements can be accurately predicted, and accurate matching and optimal allocation of resources can be achieved. In addition, the neural network model constructed by using feedback loops and nonlinear activation functions can dynamically adjust resource configuration, effectively respond to real-time changes in resource requirements, and improve the system's adaptability to workloads and overall data processing capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0058] Figure 2 A flowchart of the steps for obtaining the characteristic value of a standardized resource according to the present invention;

[0059] Figure 3 A flowchart of the steps for obtaining the resource demand forecast value of the present invention;

[0060] Figure 4 A flowchart of the steps for obtaining a preliminary neural network model of the present invention;

[0061] Figure 5 A flowchart of the steps for obtaining the optimized neural network model of the present invention;

[0062] Figure 6 This is a flow chart of the steps for obtaining the adjusted resource allocation strategy of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0064] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0065] Example 1: Please refer to Figure 1 The present invention provides a technical solution: a hyper-converged computing power scheduling method based on a prediction model, comprising the following steps:

[0066] S1: Collect and organize real-time data in the hyper-converged environment, including GPU usage and network bandwidth information, use digital filtering technology to remove noise, extract key performance indicators, standardize the key performance indicators, and establish standardized resource feature values;

[0067] S2: Perform time series analysis on the standardized resource feature values, build a prediction model to analyze the resource usage trend in the short term, calculate the demand growth rate of GPU and storage in the time period, and obtain the resource demand prediction value;

[0068] S3: Input the predicted resource demand value, build a primary neural network including feedback loop and nonlinear activation function, perform preliminary adjustment of weights and parameters, match the current resource usage pattern, and generate a preliminary neural network model;

[0069] S4: Optimize the preliminary neural network model by adjusting the depth and width of the network layer, selecting the activation function, adjusting the weight parameters of the loss function, and implementing the small batch gradient descent method for cyclic training to match future changes in resource demand and generate an optimized neural network model;

[0070] S5: Apply the optimized neural network model to implement real-time prediction of resource usage and network load, adjust resource allocation and data processing path according to the prediction results, and obtain the adjusted resource configuration strategy.

[0071] The standardized resource characteristic values ​​specifically include GPU utilization rate and network bandwidth information. The resource demand prediction values ​​include GPU demand growth rate and storage demand growth rate. The preliminary neural network model specifically refers to a neural network structure that includes a feedback loop and a nonlinear activation function. The optimized neural network model specifically includes the adjusted network layer depth and width, the selected activation function, and the adjusted loss function weight parameters. The adjusted resource allocation strategy specifically includes real-time resource usage prediction and network load prediction.

[0072] See also Figure 2 , the specific steps for obtaining the standardized resource characteristic values ​​are:

[0073] S111: Collect GPU usage and network bandwidth information from the hyper-converged environment and organize the data to generate a preliminary data set;

[0074] The usage of GPU and network bandwidth is recorded in real time, and the collected data is arranged and labeled in chronological order to ensure the integrity and traceability of the data. Through preliminary data screening, data points due to system maintenance or abnormal working conditions are excluded, and data sets that meet normal working conditions are retained. This step ensures the purity of the data set and the accuracy of subsequent processing to obtain a preliminary data set.

[0075] S112: applying digital filtering technology to the preliminary data set to obtain a cleaned performance indicator data set by eliminating statistical noise and outliers;

[0076] Digital filtering technology is applied to the preliminary data set, and a low-pass filter is used to remove high-frequency noise. At the same time, median filtering is applied to remove occasional extreme values. These technologies ensure that the most representative performance indicators are retained in the data. This processing step is particularly critical for subsequent performance evaluation because it directly affects the accuracy and reliability of the performance evaluation, and a cleaned performance indicator data set is obtained.

[0077] S113: performing numerical standardization on the cleaned performance indicator data set, converting the GPU usage and network bandwidth to a unified ratio and dimension using a Z-score method, and obtaining a standardized performance indicator data set;

[0078] The cleaned performance indicator data set is numerically standardized, and the Z-score method is used to calculate the standard deviation of each data point from the mean. In this step, each performance indicator is converted into a dimensionless form to facilitate comparison and analysis between differentiated types of data. The standardized data is easier to use in multivariate analysis, ensuring the fairness and consistency of comparisons between multiple performance indicators, and obtaining a standardized performance indicator data set.

[0079] S114: Based on the standardized performance indicator data set, select key performance indicators and calculate indicator weights using the formula:

[0080] ;

[0081] Obtaining standardized resource characteristic values;

[0082] in, represents the normalized value of a single performance indicator, Indicates the business criticality of the performance indicator, Represents the total number of performance indicators, Represents the integrated and standardized resource characteristic value.

[0083] formula:

[0084] ;

[0085] The benefit of the formula is that by weighting the performance indicators, the differentiated indicators gain multiple influences according to their business criticality, thereby more accurately evaluating the overall performance of the system.

[0086] Detailed explanation of the formula and the process of formula calculation and derivation:

[0087] There are three performance indicators, and their standardized values ​​are , , , and set the business weights of the indicators to , , ; Calculate the absolute value and weight denominator as:

[0088] ;

[0089] The weighted sum of the weight numerators is:

[0090] ;

[0091] Weighted overall performance score:

[0092] for ;

[0093] The results show that the weighted value of the overall performance index after referring to the business criticality is 0.37, which shows the performance level of the system under the current evaluation. The correlation between this numerical result and the step result is that it directly indicates the calculation basis and value of the standardized resource characteristic value.

[0094] See also Figure 3 ,The specific steps for obtaining the resource demand forecast value are:

[0095] S211: using the standardized resource characteristic values, segmenting the resource characteristic values ​​based on continuous time periods, rearranging them in chronological order and marking them with timestamps, and establishing time series data;

[0096] Through system monitoring, the collected raw time series data show the volatility of GPU usage and network bandwidth in multiple time periods. These data points are marked with timestamps, and the calculation of standardized resource feature values ​​is based on these data. First, the data is segmented to ensure that each segment includes continuous time series information. This data processing method ensures the accuracy and reliability of time series analysis. Through this method, the dynamic changes in resource usage can be effectively monitored, providing input data for subsequent analysis. The result of this step is the formation of a complete time series data set, which will be directly used for the next step of trend analysis and the establishment of a prediction model.

[0097] S212: Based on the time series data, for each continuous time period, the resource characteristic value is analyzed for the change trend within the time period, and the growth rate of the resource characteristic value in multiple time periods is calculated using the formula:

[0098] ;

[0099] Get the growth rate of GPU and storage;

[0100] in, Indicates the resource growth rate, is the resource characteristic value within the target time period, is the weighting parameter of the time slice, To calculate the adjustment factor, is the offset adjustment parameter, is the noise correction term, Represents the number of time periods referenced in the calculation;

[0101] formula:

[0102] ;

[0103] The benefit of the formula is that by introducing adjustment factors and offset parameters, the flexibility and adaptability of the model are increased, and changes in resource demand under differentiated conditions can be more accurately predicted.

[0104] Detailed explanation of the formula and the process of formula calculation and derivation:

[0105] Set the value to: ,in is the average GPU usage or storage usage during the target time period. is the weighted parameter of the time slice, which differentiates the criticality of the time slice. It is a fixed value set to adjust the baseline offset. and They are adjustment factors and noise correction items, respectively, used to deal with random fluctuations and outliers in the data. Calculation process:

[0106] ;

[0107] ;

[0108] ;

[0109] ;

[0110] The results show that with reference to time weight and noise correction, the predicted resource growth rate is 134.49, which means that resource demand has a significant increasing trend and matching resource allocation and adjustment are required.

[0111] S213: Based on the growth rate of GPU and storage, a resource demand growth ratio within the time period is obtained by performing a collective operation with resource usage in adjacent time periods.

[0112] When determining the resource demand growth rate in each time period, the growth rate of GPU and storage is used. Through set operations, the resource usage in consecutive time periods is compared with that in adjacent time periods to calculate the resource demand growth rate. In this way, the dynamic changes in resource usage in multiple time periods can be analyzed and future demand trends can be predicted. The calculation process includes the entire process from actual resource usage to demand prediction, ensuring the accuracy and practicality of the prediction results, thereby making resource management more efficient and accurate.

[0113] S214: calling the resource demand growth rate, based on the resource usage of each time segment, cumulatively calculating the resource growth rate of all segments, and obtaining a resource demand prediction value.

[0114] Based on the cumulative calculation results of the resource demand growth rate, this step involves summarizing the resource usage of each time segment in the short term, and obtaining the overall resource demand forecast value by accumulating the resource growth rate of all segments. This method not only improves the coverage of the forecast, but also improves the accuracy of the forecast, so that resource allocation can be adjusted more scientifically according to actual needs.

[0115] See also Figure 4 , the specific steps for obtaining the preliminary neural network model are:

[0116] S311: calling the resource demand prediction value, inputting it into the primary neural network, initializing the feedback loop structure and setting the nonlinear activation function, performing random distribution and preliminary setting on multiple connection weights, and generating an initial weight matrix;

[0117] Calling resource demand forecast values ​​and inputting them into the primary neural network as initial input depends critically on the accuracy of resource demand forecast values, which are obtained through time series analysis based on previous data and current resource usage patterns. The input process includes initializing the neural network structure, including setting the activation function and randomly assigned initial weights, and determining the feedback loop structure. These steps ensure that the network can start self-adjusting according to the input data. Each neuron of the neural network is initialized and ready to accept input data during the training process. Through this series of settings, the neural network model will be able to perform basic training and learning.

[0118] S312: Execute the operation of the feedback loop, adjust the weight parameters in the initial weight matrix based on the feedback output, and combine the dynamic output of the nonlinear activation function response, using the formula:

[0119] ;

[0120] Adjust the weight of each connection and update the current weight matrix to obtain the initially adjusted weights and parameters;

[0121] in, represents the weight value after the next iteration, is the current weight value, is the learning rate, is the output value of the feedback loop, is the input activation strength of the connection, To adjust the deviation factor, is the noise correction term, Represents the total number of elements involved in the sum operation;

[0122] formula:

[0123] ;

[0124] The benefit of the formula is that by introducing the adjustment bias factor and noise correction term, the weights can be adjusted dynamically in actual training, which not only enhances the matching of the model but also improves the accuracy of prediction. , learning rate , the feedback loop output is , the input activation strength of the connection is , adjustment bias factor , noise correction term , calculated according to the formula:

[0125]

[0126] ;

[0127] ;

[0128] ;

[0129] The results show that the weights adjusted through feedback can better adapt to the current training data, improving the response sensitivity and accuracy of the model.

[0130] S313: Utilize the preliminarily adjusted weights and parameters, combine them with the resource demand forecast value, perform multiple iterative calculations on multiple weights and parameters, match the current resource usage pattern, and generate a preliminary neural network model.

[0131] Using the initially adjusted weights and parameters, which are obtained through multiple iterations of neural network training, the network cyclically receives resource demand forecast values ​​during the training process, performs pattern matching and adjustment based on these values, and each iteration adjusts the weights based on the current resource usage pattern and predicted output, including multiple forward propagations and backward propagations, adjusting the weights so that the output is as close to the expected result as possible. The goal of each iteration is to reduce the prediction error, improve the network's generalization ability and ability to match new data, and each update of the weights is based on the current error rate and the previous weight value. In this way, the neural network gradually approaches the optimal parameter configuration, and the generated preliminary neural network model can effectively reflect and match the actual resource usage pattern, so that the model has better prediction accuracy and stability in practical applications.

[0132] See also Figure 5 , the specific steps for obtaining the optimized neural network model are:

[0133] S411: Based on the preliminary neural network model, increase or decrease the number of network layers and the number of nodes per layer, optimize the number of hidden layers and the number of nodes based on resource demand prediction, and obtain the adjusted neural network architecture;

[0134] When adjusting the architecture of a neural network model, you first need to analyze the impact of the depth and width of the network layer on the model performance. By increasing the number of layers, you can improve the model's learning ability, while increasing the number of nodes helps capture feature information. By adjusting the depth and width of the network, you can determine the optimal structure through experiments. Then, based on resource demand forecast data, you can configure the number of hidden layers and the number of nodes in each layer. The determination of each number of nodes needs to be driven by actual data. Experiment with multiple groups of settings to verify the performance improvement effect of the model and gradually determine the neural network architecture.

[0135] S412: Based on the adjusted neural network architecture, select matching activation functions for multiple hidden layers, and perform combined tests using differentiated activation functions for each layer, using the formula:

[0136] ;

[0137] Calculate the response outputs of multiple nodes to obtain an optimized combination of activation functions;

[0138] in, Represents the output of the neural network layer, which is the total output of the weighted sum after multiple node activation functions are processed. is the node activation function, is the weight coefficient of the activation function, which is used to adjust the strength of the node input signal. is the input signal, is the bias parameter, which is used to adjust the activation threshold of the activation function. Represents the total number of nodes, indicating the number of nodes in the current layer;

[0139] formula:

[0140] ;

[0141] The benefit of the formula is that by dynamically adjusting the weight coefficient of the activation function , the output of each layer can be optimized according to the characteristics of the multiple layers, thereby improving the learning and matching of the overall network.

[0142] Detailed explanation of the formula and the process of formula calculation and derivation:

[0143] Set the network to have three layers, and select the same activation function for each layer. Assigning differentiated weights , these weights are adjusted during the network training process. The input of the first layer 0.5 respectively 0.3 0.9, weight Initial setting is 1.2 0.8 1.0. Through actual operation and adjustment, the weight will be updated to 1.15 according to the gradient descent 0.85 1.05. The output of the first layer is calculated as follows:

[0144] ;

[0145] in, is the bias term, which is a small constant such as 0.05. If the Sigmoid function is used as the activation function , the output of each item will be compressed between 0 and 1.

[0146] The results show that by adjusting the weights and biases of the differentiation layer, it is possible to control the output of each layer of the neural network and optimize the model's ability to process data.

[0147] S413: Based on the optimized activation function combination, the weight parameters of the loss function are adjusted, small batch gradient descent training is implemented in combination with the current weight matrix, the neural network error is iteratively calculated, and the optimized neural network model is generated.

[0148] Adjusting the weight parameters of the loss function is a key step in optimizing the neural network model. By adjusting the weight parameters of the loss function, the model can better match the target type of data and increase the weight of the minority class when processing unbalanced data. This requires adjusting the weight parameters based on the distribution and characteristics of the data. The implementation of the mini-batch gradient descent method requires calculating the loss of each batch and updating the model parameters. This process involves calculating the gradient of the loss function and updating the weights based on it. Through multiple iterations, the loss function value is gradually reduced, so that the model parameters gradually approach the optimal solution.

[0149] See also Figure 6 , the specific steps for obtaining the adjusted resource allocation strategy are:

[0150] S511: Based on the optimized neural network model, the real-time monitoring data is processed to obtain the current resource usage, and a prediction analysis is performed based on the real-time resource usage and network load to obtain a real-time resource prediction result;

[0151] Based on the existing architecture of the optimized neural network model and the input of real-time monitoring data, the real-time monitoring data is first cleaned and preprocessed to remove outliers and noise to ensure data quality. Then, by adjusting the parameter configuration of the network layer, including adjusting the number of hidden layers and the number of neurons in each layer, these parameters are adjusted in real time according to network traffic and resource usage to match the dynamically changing network load, and implement resource usage and network load prediction. Through continuous data training and model optimization, the accuracy and response speed of the prediction are improved to obtain real-time resource prediction results.

[0152] S512: Based on the real-time resource prediction results, calculate the resource demand change trend, adjust the current resource configuration, and use weighted parameters to smooth the prediction error. The formula is:

[0153] ;

[0154] Get the adjusted resource ratio;

[0155] in, Indicates the adjusted resource allocation, is the weighting coefficient, is the real-time resource usage, To predict the load, is the current configuration capacity, is the noise correction term, Represents the total number of resource types or nodes involved in the calculation;

[0156] ;

[0157] The benefit of the formula is that it can dynamically adjust resource allocation, match cyclically changing network conditions, and optimize system performance by integrating real-time resource usage and predicted load, taking a weighted average, and then referring to the current configuration capacity and noise correction.

[0158] Detailed explanation of the formula and the process of formula calculation and derivation:

[0159] There are three types of resources (GPU, storage, bandwidth) with weights of 0.5, 0.3, and 0.2 respectively. The current usage is 70%, 50%, and 60%. The predicted growth is 10%, 15%, and 5%. The current configuration capacity is 100, 200, and 300. The noise correction items are 0.1, 0.1, and 0.1.

[0160] 1. Calculate the weighted usage and predicted sum of multiple resources:

[0161] ;

[0162] 2. Calculate the adjusted sum of the configured capacity:

[0163] ;

[0164] 3. Resource allocation:

[0165] ;

[0166] The results show that the adjusted resource ratio is about 2.92 resources that can be supported per unit configuration. This value guides the dynamic adjustment of resources, responds to future changes in resource demand, and ensures that system resources are neither overloaded nor idle.

[0167] S513: Based on the adjusted resource allocation ratio and in combination with the real-time resource prediction result, the data processing paths and resource allocations among the multiple network nodes are reallocated to obtain an adjusted resource configuration strategy.

[0168] Based on real-time prediction results and current network load data, advanced resource management strategies are adopted to evaluate the existing resource configuration one by one, analyze the peak and idle periods of resource usage, and implement dynamic adjustment of resources according to differentiated application and service requirements. By optimizing the data processing path, reducing data transmission delays and blockages, and improving data processing efficiency, the adjusted resource allocation strategy is obtained. The strategy refers to the resource demand forecast in the future time period and the current resource usage to ensure maximum resource utilization, avoid resource waste, and improve the overall performance of the system and the user's service experience.

[0169] A hyper-converged computing power scheduling system based on a prediction model, the hyper-converged computing power scheduling system based on a prediction model is used to execute the hyper-converged computing power scheduling method based on a prediction model, and the system includes:

[0170] The data preprocessing module collects GPU usage and network bandwidth data in the hyper-converged environment, uses digital filtering technology to remove random noise in the data, screens key performance indicators of GPU usage and network bandwidth, and converts the indicators into standardized resource characteristic values ​​through linear transformation;

[0171] The trend prediction model module uses standardized resource feature values ​​to build a time series analysis model to evaluate the dynamic changes in recent resource usage, calculate and predict the demand growth rate of GPU and storage in the short term, and thus generate resource demand forecast values;

[0172] The neural network building module takes the resource demand forecast value as input, sets the feedback loop in the neural network and selects the nonlinear activation function, preliminarily adjusts the network parameters to match the current resource usage pattern, creates a preliminary neural network model, adjusts the depth, width, and loss function weight of the network layer, and uses the mini-batch gradient descent method to train and optimize the network to adapt to future changes in resource demand and obtain the optimized neural network model;

[0173] The resource allocation strategy module deploys the optimized neural network model to perform real-time resource usage and network load prediction, dynamically adjusts resource allocation and data processing paths according to the prediction results, and forms and implements the adjusted resource allocation strategy.

[0174] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A hyper-converged computing power scheduling method based on a prediction model, characterized in that: The following steps are involved: Collect and organize real-time data in the hyper-converged environment, including GPU usage and network bandwidth information, use digital filtering technology to remove noise, extract key performance indicators, standardize the key performance indicators, and establish standardized resource feature values; Perform time series analysis on the standardized resource feature values, build a prediction model to analyze the usage trend of resources in the short term, calculate the demand growth rate of GPU and storage in the time period, and obtain the resource demand prediction value; The steps for obtaining the resource demand forecast value are specifically as follows: Using the standardized resource characteristic values, the resource characteristic values ​​are segmented based on continuous time periods, rearranged in chronological order and marked with timestamps to establish time series data; Based on the time series data, for each continuous time period, the resource characteristic value is analyzed for the change trend within the time period, and the growth rate of the resource characteristic value in multiple time periods is calculated using the formula: ； Get the growth rate of GPU and storage; in, Indicates the resource growth rate, is the resource characteristic value within the target time period, is the weighting parameter of the time slice, To calculate the adjustment factor, is the offset adjustment parameter, is the noise correction term, Represents the number of time periods referenced in the calculation; Based on the growth rate of the GPU and storage, a resource demand growth ratio within the time period is obtained by performing a collective operation with resource usage in adjacent time periods; The resource demand growth rate is called, and based on the resource usage of each time segment, the resource growth rate of all segments is cumulatively calculated to obtain a resource demand forecast value; Input the resource demand forecast value, construct a primary neural network including a feedback loop and a nonlinear activation function, perform preliminary adjustments to weights and parameters, match the current resource usage pattern, and generate a preliminary neural network model; The steps for obtaining the preliminary neural network model are specifically as follows: Calling the resource demand prediction value, inputting it into the primary neural network, initializing the feedback loop structure and setting the nonlinear activation function, performing random distribution and preliminary setting on multiple connection weights, and generating an initial weight matrix; The operation of the feedback loop is performed, and the weight parameters in the initial weight matrix are adjusted based on the feedback output, combined with the dynamic output of the nonlinear activation function response, using the formula: ； Adjust the weight of each connection and update the current weight matrix to obtain the initially adjusted weights and parameters; in, represents the weight value after the next iteration, is the current weight value, is the learning rate, is the output value of the feedback loop, is the input activation strength of the connection, To adjust the deviation factor, is the noise correction term, Represents the total number of elements involved in the sum operation; Utilizing the preliminarily adjusted weights and parameters, combining them with resource demand forecast values, performing multiple iterative calculations on multiple weights and parameters, matching current resource usage patterns, and generating a preliminary neural network model; Optimizing the preliminary neural network model by adjusting the depth and width of the network layer, selecting the activation function, adjusting the weight parameters of the loss function, and implementing the small batch gradient descent method for cyclic training to match future changes in resource demand, thereby generating an optimized neural network model; The steps for obtaining the optimized neural network model are specifically as follows: Based on the preliminary neural network model, increase or decrease the number of layers and the number of nodes in each layer of the network, optimize the number of hidden layers and the number of nodes based on resource demand prediction, and obtain an adjusted neural network architecture; Based on the adjusted neural network architecture, matching activation functions are selected for multiple hidden layers, and a combination test is performed on each layer using a differentiated activation function, using the formula: ； Calculate the response outputs of multiple nodes to obtain an optimized combination of activation functions; in, Represents the output of the neural network layer, which is the total output of the weighted sum after multiple node activation functions are processed. is the node activation function, is the weight coefficient of the activation function, which is used to adjust the strength of the node input signal. is the input signal, is the bias parameter, which is used to adjust the activation threshold of the activation function. Represents the total number of nodes, indicating the number of nodes in the current layer; Based on the optimized activation function combination, the weight parameters of the loss function are adjusted, small batch gradient descent training is performed in combination with the current weight matrix, the neural network error is iteratively calculated, and an optimized neural network model is generated; The optimized neural network model is applied to implement real-time prediction of resource usage and network load, and resource allocation and data processing path are adjusted according to the prediction results to obtain an adjusted resource configuration strategy.

2. The method for scheduling hyper-converged computing power based on a prediction model according to claim 1, characterized in that: The standardized resource characteristic values ​​specifically include GPU utilization rate and network bandwidth information; the resource demand prediction values ​​include GPU demand growth rate and storage demand growth rate; the preliminary neural network model specifically refers to a neural network structure including a feedback loop and a nonlinear activation function; the optimized neural network model specifically includes the adjusted network layer depth and width, the selected activation function, and the adjusted loss function weight parameters; the adjusted resource allocation strategy specifically includes real-time resource usage prediction and network load prediction.

3. The method for scheduling hyper-converged computing power based on a prediction model according to claim 2, characterized in that: The steps for obtaining the standardized resource characteristic value are specifically as follows: Collect GPU usage and network bandwidth information from the hyperconverged environment and organize the data to generate a preliminary dataset; Applying digital filtering technology to the preliminary data set to obtain a cleaned performance indicator data set by eliminating statistical noise and outliers; The cleaned performance indicator data set is numerically normalized, and the GPU usage and network bandwidth are converted to a uniform ratio and dimension using a Z-score method to obtain a standardized performance indicator data set; Based on the standardized performance indicator data set, key performance indicators are selected and indicator weights are calculated using the formula: ； Obtaining standardized resource characteristic values; in, represents the normalized value of a single performance indicator, Indicates the business criticality of the performance indicator, Represents the total number of performance indicators, Represents the integrated and standardized resource characteristic value.

4. The method for scheduling hyper-converged computing power based on a prediction model according to claim 1, characterized in that: The steps for obtaining the adjusted resource allocation strategy are specifically as follows: Based on the optimized neural network model, the real-time monitoring data is processed to obtain the current resource usage, and a prediction analysis is performed based on the real-time resource usage and network load to obtain a real-time resource prediction result; Based on the real-time resource prediction results, the resource demand change trend is calculated, and the current resource configuration is adjusted. The prediction error is smoothed using weighted parameters, using the formula: ； Get the adjusted resource ratio; in, Indicates the adjusted resource allocation, is the weighting coefficient, is the real-time resource usage, To predict the load, is the current configuration capacity, is the noise correction term, Represents the total number of resource types or nodes involved in the calculation; Based on the adjusted resource allocation ratio and in combination with the real-time resource prediction result, the data processing paths and resource allocations among multiple network nodes are reallocated to obtain an adjusted resource configuration strategy.

5. A hyper-converged computing power scheduling system based on a prediction model, characterized in that: According to a method for scheduling hyper-converged computing power based on a prediction model according to any one of claims 1 to 4, the system comprises: The data preprocessing module collects GPU usage and network bandwidth data in the hyper-converged environment, uses digital filtering technology to remove random noise in the data, screens key performance indicators of GPU usage and network bandwidth, and converts the indicators into standardized resource characteristic values ​​through linear transformation; The trend prediction model module uses the standardized resource characteristic values ​​to construct a time series analysis model, evaluate the dynamic changes in recent resource usage, calculate and predict the demand growth rate of GPU and storage in the short term, and thus generate a resource demand forecast value; The neural network construction module takes the resource demand prediction value as input, sets the feedback loop in the neural network and selects the nonlinear activation function, preliminarily adjusts the network parameters to match the current resource usage pattern, creates a preliminary neural network model, adjusts the depth, width, and loss function weight of the network layer, and uses a small batch gradient descent method to train and optimize the network to adapt to future changes in resource demand, and obtain an optimized neural network model; The resource allocation strategy module deploys the optimized neural network model to perform real-time resource usage and network load prediction, dynamically adjusts resource allocation and data processing paths according to the prediction results, and forms and implements the adjusted resource allocation strategy.

Images