Method for evaluating real-time performance of computing power network based on analytic hierarchy process

By establishing a multi-level comprehensive evaluation index system and Stacking heterogeneous integrated learning model, the problems of insufficient real-time, multi-level inaccuracy and insufficient heterogeneity processing in computing power network performance evaluation are solved, and efficient and accurate performance is achieved The evaluation has improved the utilization efficiency of computing power resources and the allocation optimization of computing tasks.

CN118897952APending Publication Date: 2024-11-05JINAN UNIVERSITY
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202410971736.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-05

AI Technical Summary

Technical Problem

The existing computing power network performance evaluation methods have problems such as insufficient real-time, multi-level inaccuracy, insufficient heterogeneity processing and low model efficiency. They cannot effectively adapt to the dynamically changing computing environment and deal with differences in heterogeneous nodes, which affects the accuracy of evaluation. Sex and efficiency.

Method used

Using a real-time evaluation method based on hierarchical analysis, by establishing a multi-level comprehensive evaluation index system and Stacking heterogeneous integrated learning model, collecting and analyzing the operating data of computing power network nodes, performing top-down feature analysis and prediction, and realizing Efficient and accurate evaluation of the performance of multi-layer computing power network.

Benefits of technology

Realize real-time and accurate evaluation of computing power network performance, improve the real-time and multi-level accuracy of evaluation, can handle the heterogeneity of different hardware and architectures, improve the universality of evaluation and computing efficiency, and support computing Optimized configuration and scheduling of power resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118897952A_ABST
    Figure CN118897952A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of computing power network performance evaluation, in particular to a computing power network real-time performance evaluation method based on an analytic hierarchy process. The method comprises the following steps: firstly, establishing a multi-level comprehensive evaluation index system, starting from a plurality of first-level indexes of a server node from a processor, a memory, a network, a disk and the like, subdividing the first-level indexes into a second-level index and a third-level index, such as a processor utilization rate, a memory utilization rate and the like, and performing the same processing on other level nodes; then, collecting operation data of computing power network nodes, and carrying out cleaning, normalization processing and noise elimination to obtain a cleaned computing power feature vector; thirdly, training the model by adopting a Stacking heterogeneous ensemble learning technology, performing calculation power prediction, summarizing evaluation results layer by layer through a weighted average method, and obtaining overall evaluation on the performance of the whole calculation power network; according to the method, the real-time performance and the accuracy of performance evaluation of the multi-layer computing power network are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention mainly relates to the field of computing power network, and in particular to the technology of real-time evaluation of computing power network performance. Specifically, the present invention relates to a real-time evaluation method based on hierarchical analysis method, which is used to evaluate and predict the performance of nodes in the computing power network. By collecting and analyzing the operating data of each node in the computing power network, and using a multi-level comprehensive evaluation index system and a heterogeneous integrated learning model, the performance of computing power networks at different levels is evaluated in real time to improve the utilization efficiency of computing power resources and optimize the allocation of computing tasks. Background Art

[0002] With the rapid development of cloud computing, edge computing, and IoT technologies, computing resources are gradually expanding from centralized cloud to distributed edge and terminal devices. Such a multi-level computing architecture makes the management and scheduling of computing resources more complex and important. The current computing network consists of multiple heterogeneous nodes, including cloud data centers, large regional computing centers, small edge nodes, and terminal devices. Evaluating the performance of these different levels of nodes is crucial to ensuring the efficiency and stability of the overall computing network.

[0003] However, existing evaluation methods have the following major problems:

[0004] 1) Lack of real-time performance: Traditional evaluation methods are mostly based on static data and cannot adapt to the dynamically changing computing environment, resulting in delayed evaluation results.

[0005] 2) Multi-level inaccuracy: There is a lack of comprehensive consideration of multi-level computing power characteristics, and evaluation is based only on the characteristics of a single level, which makes it difficult to accurately reflect the performance of the entire computing power network.

[0006] 3) Insufficient heterogeneity processing: There are differences in hardware, architecture, and performance between different nodes. Existing methods perform poorly in handling these heterogeneities, affecting the accuracy of the evaluation.

[0007] 4) Low model efficiency: Existing model methods are highly complex and it is difficult to achieve efficient evaluation while ensuring accuracy.

[0008] In view of the above problems, the present invention proposes a real-time evaluation method based on hierarchical analysis method, which realizes efficient and accurate evaluation of the performance of multi-layer computing power network through the application of multi-level feature analysis and integrated learning model. Summary of the invention

[0009] This invention proposes a method for real-time evaluation of multi-layer computing network performance, aiming to solve the problems of insufficient real-time performance, inaccurate multi-level evaluation, insufficient heterogeneity processing and low model integration efficiency in the prior art. The specific implementation steps are as follows:

[0010] Step 1: Establish a comprehensive evaluation index system:

[0011] 1. Analyze the computing power network from the whole to the part, and divide it into several major categories of computing power features X i , such as processor, memory, network, disk, etc., as the first-level indicators.

[0012] 2. Further subdivide the first-level indicators into more specific second-level indicators, such as processor usage, memory utilization, etc. If the second-level indicators are still not sufficient to describe the computing power characteristics in detail, they can be further subdivided into multi-level indicators to ensure comprehensive coverage of computing power characteristics. Ensure that the various characteristics of the computing power network can be fully and carefully described, providing a basis for subsequent data analysis and model training.

[0013] Step 2: Data collection and preprocessing:

[0014] 1. According to the established comprehensive evaluation index system, collect the operation data of each node in the computing power network and extract the computing power feature vector X = {x 1 ,x 2 ,x 3 ,...,x n}, operation data includes but is not limited to basic information such as CPU utilization, memory utilization, disk I / O rate, network bandwidth utilization, etc. If conditions permit, higher-level computing power characteristics should also be collected, such as characteristic data at the city-level computing power network level. These high-level characteristics include computing power quality and efficiency, computing power scale, storage power scale, storage power performance, network capacity quality and efficiency, and basic network conditions.

[0015] 2. Calculate the power feature vector X = {x 1 ,x 2 ,x 3 ,...,x n}Cleaning and normalization are performed to eliminate outliers and noise, ensure the quality and consistency of the data, and provide a reliable data foundation for model training and evaluation for subsequent analysis and modeling.

[0016] Step 3: Integrated model training and optimization: Use Stacking's heterogeneous learning method to integrate models to build a computing power node computing power prediction model;

[0017] 1. Define the eigenvector X i , X i =[x i1 ,x i2 ,...,x in ], where x ij Represents the jth feature vector of the ith server.

[0018] 2. Use Stacking heterogeneous ensemble learning to predict the computing power of each server. Assume that there are M base learners, denoted as {h 1 ,h 2 ,...,h M}, use k-fold cross validation (such as k=5) to train and evaluate the model to ensure the stability and generalization ability of the model.

[0019] 3. Divide the training set D into k mutually exclusive subsets D 1 ,D 2 ,...,D N , and then perform k-fold cross validation on each base learner.

[0020] 4. Perform meta-learner training, combine the prediction results of all base learners into a new feature vector, and use the new feature vector to train the meta-learner and obtain the final prediction value.

[0021] Step 4: Real-time performance prediction and evaluation:

[0022] 1. Conduct real-time performance evaluation on computing network area D. First, the regional scope of computing network D is layered, such as regions, provinces, cities, and even server computing node layers, and performance evaluation is performed on each layer separately;

[0023] 2. Server computing power prediction: Through the trained integrated model, the computing power of all server computing nodes in the computing power network area D is evaluated to calculate the computing power value of server i

[0024] 3. Hierarchical computing power summary: Through the weighted average method, the computing power nodes are gradually added from the lower layer to the upper layer, and gradually summarized from small areas to large areas, layer by layer, and finally the overall performance evaluation result of the entire computing power network D is obtained.

[0025] Compared with the prior art, the advantages of this method are:

[0026] High real-time performance: Through the collection and processing of real-time data, the performance changes of the computing power network can be reflected in a timely manner.

[0027] Multi-level precise evaluation: By comprehensively considering the computing power characteristics at different levels, the evaluation results are more accurate and reliable.

[0028] Heterogeneity processing: Ability to handle the heterogeneity of different hardware and architectures to improve the universality of evaluation.

[0029] Model efficiency: Use stacking heterogeneous ensemble learning models to improve prediction performance while maintaining high computational efficiency.

[0030] Through the above method, the present invention can effectively improve the real-time and accuracy of multi-layer computing power network performance evaluation, and provide an important reference for the optimal configuration and scheduling of computing power resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 :Overall system flow chart

[0032] Demonstrate the process of the entire computing power network performance evaluation system.

[0033] It includes the connection between various modules such as establishing an indicator system, collecting data, data preprocessing, computing power prediction, and result summary.

[0034] Figure 2 :Hierarchical structure diagram of comprehensive evaluation index system

[0035] Explain the hierarchical structure from primary indicators to multi-level indicators.

[0036] The first-level indicators include processor, memory, network, disk, GPU, energy consumption, failure rate, etc.

[0037] This is further broken down into secondary and tertiary metrics, such as processor utilization, memory utilization, etc.

[0038] Figure 3 :Data collection and preprocessing flow chart

[0039] Demonstrate the steps of data collection, cleaning and normalization.

[0040] Data is collected from various computing nodes.

[0041] The data cleaning process removes outliers and normalization ensures data consistency.

[0042] Figure 4 :Characteristics of city computing power nodes

[0043] Display the performance indicators of urban computing power nodes in terms of computing power quality and efficiency, computing power scale, storage capacity scale, storage performance, network capacity quality and efficiency, and basic network conditions.

[0044] Figure 5 :Stacking integrated model training flow chart

[0045] Describe the training and prediction process of the base learners and meta-learners, including the k-fold cross-validation steps.

[0046] The prediction results of the base learners are combined into a new feature vector for training the meta-learner.

[0047] Figure 6 :Real-time performance prediction and evaluation flow chart

[0048] Demonstrate the steps from regional stratification, server computing node evaluation, to the summary of overall performance evaluation results.

[0049] The hierarchical structure ranges from regional level, provincial and municipal level to server node level.

[0050] The computing power values ​​are aggregated layer by layer using the weighted average method.

[0051] Figure 7 :Schematic diagram of computing power network layering

[0052] Shows the hierarchical structure from city level, regional level to server node level.

[0053] The connection relationship between each level and the method of summarizing the computing power value.

[0054] Figure 8 :Schematic diagram of integrated model structure

[0055] Shows the structure of the Stacking ensemble model, including the relationship between the base learner and the meta learner.

[0056] Detailed structures of different base learners (such as decision trees, support vector machines, neural networks, etc.). DETAILED DESCRIPTION

[0057] like Figure 1 As shown, the specific steps of the technical solution of the present invention are:

[0058] Step 1: Establish a comprehensive evaluation index system, such as Figure 2 As shown:

[0059] Analyze the computing power network, from the whole to the parts, and divide the computing power characteristics from top to bottom X i For server computing nodes, the first-level indicator X i Including: processor, memory, network, disk, GPU, energy consumption, failure rate, on this basis, further subdivided into the first-level indicators X i , the first-level indicators are divided into more specific second-level indicators x i Among them, the secondary indicators of the processor include: processor usage, processor main frequency, processor cache size, etc.; the secondary indicators of the memory include: memory utilization, memory capacity, memory access speed, etc.; the secondary indicators of the network include network bandwidth utilization, network delay, packet loss rate, etc.; the secondary indicators of the disk include: disk I / O rate, disk capacity, disk read and write delay, etc.; the secondary indicators of the GPU include: GPU utilization, GPU memory utilization, etc.; the secondary indicators of energy consumption include: average energy consumption, peak energy consumption, minimum energy consumption, etc.; the secondary indicators of the failure rate include: average failure rate, peak failure rate, minimum failure rate, etc.

[0060] If a more detailed description is needed, it can be further subdivided into three-level indicators. For example, processor utilization can be further subdivided into average utilization, peak utilization, minimum utilization, etc., to ensure comprehensive coverage of computing power characteristics.

[0061] Step 2: Data collection and preprocessing, the process is as follows Figure 3 As shown:

[0062] 1. According to the established comprehensive evaluation index system, collect the operation data of each node in the computing power network and extract the computing power feature vector X = {x 1 ,x 2 ,x 3 ,...,x n When collecting data, a distributed data collection system can be used to collect the operating data of each computing power node in real time. If conditions permit, higher-level computing power characteristics should also be collected, such as characteristic data at the city-level computing power network level. These high-level characteristics include computing power quality and efficiency, computing power scale, storage power scale, storage power performance, network capacity quality and efficiency, and basic network conditions, such as Figure 4 By comprehensively analyzing these multi-level and multi-dimensional data, the present invention can achieve a comprehensive and accurate evaluation of the performance of the computing power network, especially in the performance evaluation of the city-level computing power network, providing more in-depth and broad insights.

[0063] 2. Calculate the power feature vector X = {x 1 ,x 2 ,x 3 ,...,x n} Clean and normalize the data to eliminate outliers and noise for subsequent analysis and modeling.

[0064] Data cleaning includes the following two steps: 1) Missing value processing: Detect missing values ​​in the data and directly delete incomplete records. If some data records are important, they cannot be deleted. The average value filling method is used to process missing values. 2) Outlier processing: Use the Z-score standard score method to detect outliers. By measuring the distance between the data point and the average value, calculate the Z-score of the data, set a specific threshold t, and when the Z-score of the data>t, the data is judged as an outlier and the records containing the outlier are directly deleted. The Z-score calculation formula is as follows:

[0065]

[0066] Among them, X i is the data value, μ is the mean value of the data, and σ is the standard deviation of the data.

[0067] The normalization process is to convert the data to the same scale. Min-Max normalization can be used. The formula is as follows:

[0068]

[0069] Among them, x is the data value, x min is the minimum value of the data, x max is the maximum value of the data.

[0070] Step 3: Integrated model training and optimization: Use Stacking's heterogeneous learning method to integrate models to build a computing power node computing power prediction model. The structural diagram is as follows: Figure 5 As shown;

[0071] 1. Define the eigenvector X i , X i =[x i1 ,x i2 ,...,x in ], where x ij Represents the jth feature vector of the ith server.

[0072] 2. Use Stacking heterogeneous ensemble learning to predict the computing power of each server. Assume that there are M base learners, denoted as {h 1 ,h 2 ,...,h M}, use k-fold cross validation (such as k=5) to train and evaluate the model to ensure the stability and generalization ability of the model.

[0073] 3. Divide the training set D into k mutually exclusive subsets D 1 ,D 2 ,...,D N , and then perform k-fold cross validation on each base learner. In each validation: use k-1 subsets for training, the remaining subset for validation, and save the prediction results of the base learner on the validation set. Assume that the i-th fold training set is D -i =DD i , the validation set is D i , then the base learner h j The predicted value at the i-th fold is:

[0074]

[0075] Summarize the prediction results of all folds to get the complete prediction results of the base learner:

[0076]

[0077] 4. Meta-learner training, combining the prediction results of all the above base learners into a new feature vector z i

[0078]

[0079] Use the new eigenvector z i , train the meta-learner and get the final prediction value:

[0080]

[0081] Step 4: Real-time performance prediction and evaluation based on weighted average. The process is as follows: Figure 6 As shown:

[0082] 1. Computing network area stratification: Real-time performance evaluation of computing network area D, such as Figure 7 As shown, the regional scope of the computing power network D is first layered, such as regions, provinces, cities, and even server computing power node layers, and performance evaluation is performed on each layer;

[0083] 2. Server computing power prediction: Through the trained integrated model, the computing power of all server computing nodes in the computing power network area D is evaluated to calculate the computing power value of server i like Figure 8 shown.

[0084] 3. Hierarchical computing power summary: Through the weighted average method, the computing power nodes are gradually added from the lower layer to the upper layer, and gradually summarized from small areas to large areas, layer by layer, and finally the overall performance evaluation result of the entire computing power network D is obtained. For example:

[0085] Suppose there are N A servers, the weight of server i is w i Then the computing power value of the region is Y A for:

[0086]

[0087] Assume that there are N sub-areas in area B, and the computing power of the i-th sub-area is Y i , with weight w i , then the computing power value of area B is Y B for:

[0088]

[0089] The weight w iThe subjective weighting method is used to determine the importance of each server. First, determine the importance factors. That is, according to the importance of the server, computing power, historical performance and collected computing power characteristics, determine the importance factors of each server. Second, assign initial weights. That is, assign initial weights to each server based on the above factors. The allocation of initial weights can be done by expert scoring. Suppose there are N servers, and the importance scores of each server are S 1 ,S 2 ,...,S N , then the initial weight w i for: Finally, adjust the weights. That is, adjust the weights regularly based on the monitoring data and feedback in actual operation. The adjustment method can use the exponential smoothing method to enable the weights to dynamically reflect the current status and historical performance of each server.

[0090]

[0091] Among them, w i (t) is the weight of the i-th server at time t, is the weight prediction value of the i-th server at time t+1, α is the smoothing coefficient (0

Claims

1. A real-time performance evaluation method for computing power network based on hierarchical analysis method, characterized by The real-time performance of the computing network is predicted and evaluated using the trained Stacking heterogeneous integration model based on the data of the computing nodes collected in real time, which includes at least the following steps: Step 1: Establish a comprehensive evaluation index system: 1) Establish a comprehensive evaluation index system, using a top-down analysis method, from the whole to the parts; 2) Analyze the computing power network from top to bottom and divide the computing power nodes into several major categories of computing power characteristics as the first-level indicators; 3) These broad categories of computing power characteristics are further divided into more specific computing power characteristics, and the first-level indicators are further divided into more specific second-level indicators. If the second-level indicators are still not sufficient to describe the computing power characteristics in detail, they can be further divided into multiple levels of indicators to ensure comprehensive coverage of the computing power characteristics; Step 2: Data collection and preprocessing: 1) Collect the operation data of each node in the computing power network according to the established comprehensive evaluation index system; 2) Clean and normalize the collected data to eliminate outliers and noise; Step 3: Integrated model training and optimization: Use Stacking's heterogeneous learning integrated model to build a computing power node computing power prediction model; Step 4: Real-time performance prediction and evaluation: 1) Conduct real-time hierarchical performance evaluation of computing power network areas; 2) Use the Stacking heterogeneous ensemble learning model to predict the computing power value of each server, evaluate the computing power of the server computing nodes in the computing power network area, and calculate the computing power value; 3) Through the weighted summation method, the data is gradually aggregated from small areas to large areas, layer by layer, and finally the overall performance evaluation result of the entire computing power network is obtained.

2. The real-time performance evaluation method of computing power network based on hierarchical analysis method according to claim 1 is characterized in that The real-time performance of the computing power network is predicted and evaluated using the trained Stacking heterogeneous integrated model through the data of the computing power nodes collected in real time. The step 3 uses the Stacking heterogeneous learning method integrated model to construct a computing power prediction model for the computing power nodes, and at least includes the following steps: 1) Define the eigenvector X i , X i =[x i1 ,x i2 ,...,x in ], where x ij Represents the jth feature vector of the ith server. 2) Use Stacking heterogeneous ensemble learning to predict the computing power of each server. Assume that there are M base learners, denoted as {h1,h2,...,h M }, use k-fold cross validation (such as k=5) to train and evaluate the model to ensure the stability and generalization ability of the model. 3) Divide the training set D into k mutually exclusive subsets D1, D2, ..., D N , and then perform k-fold cross validation on each base learner. In each validation: use k-1 subsets for training, the remaining subset for validation, and save the prediction results of the base learner on the validation set. Assume that the i-th fold training set is D -i =DD i , the validation set is D i , then the base learner h j The predicted value at the i-th fold is: Summarize the prediction results of all folds to get the complete prediction results of the base learner: 4) Meta-learner training, combining the prediction results of all the above base learners into a new feature vector z i Use the new eigenvector z i , train the meta-learner and get the final prediction value: 。

Citation Information

Cited By

  • Network quality evaluation method for computing power network

    CN119135555A

  • A network quality evaluation method for computing power network

    CN119135555B

  • Transformer health prediction system under multi-dimensional monitoring data fusion

    CN119150245A

  • Efficient GPU computing power server resource optimization method and system

    CN120147101A

  • Calculation power thermal evaluation method based on multi-dimensional data analysis and decision making system thereof

    CN120181518A