CDN (Content Delivery Network) node deployment method and system for realizing multi-dimensional resource quantification and complementary scheduling

Abstract

The invention discloses a CDN (Content Delivery Network) node deployment method and system for realizing multi-dimensional resource quantization and complementary scheduling, and relates to the technical field of content delivery networks, and the method comprises the following five steps: collecting domain name operation data and constructing a standardized feature vector; scientifically clustering domain names based on unit bandwidth CPU consumption intensity, IO consumption intensity and time sequence bandwidth characteristics; weighting, averaging and quantifying domain name class multi-dimensional resource requirements by taking the bandwidth as a weight; constructing a multi-constraint optimization model oriented to multi-dimensional resource complementation; and solving the model and generating a domain name level scheduling scheme. The system comprises a data acquisition module, a feature engineering module, a domain name classification module, a resource quantification module, a scheduling decision module and a deployment monitoring module. According to the method, scientific domain name classification and accurate resource quantification are realized, global collaborative optimization of CPU, IO, bandwidth and time sequence is achieved, node peak load is reduced, and the resource utilization rate is improved.

Description

Technical Field

[0001] This invention relates to the field of content delivery network technology, and in particular to a CDN node deployment method and system for realizing multi-dimensional resource quantification and complementary scheduling. Background Technology

[0002] In the field of Content Delivery Network (CDN) technology, with the rapid development of internet services, the number of domain names that a single CDN system needs to handle has grown from thousands in the early days to hundreds of thousands or even millions. Faced with such a large number of domain names, the computational complexity of directly performing global scheduling across all domains is extremely high, making it impossible to meet the real-time requirements of businesses. Therefore, aggregating massive amounts of domain names into a limited number of "domain groups" for dimensionality reduction scheduling is an inevitable choice to reduce computational complexity and achieve second-level response times.

[0003] However, existing domain name classification methods have fundamental flaws, making it impossible to accurately quantify domain name resources. Specifically: First, static grouping by customer / contract is the mainstream approach, based on business logic rather than domain resource consumption characteristics. Under the same customer, there may be heterogeneous businesses such as IO-intensive large file downloads and CPU-intensive dynamic API services, resulting in huge differences in resource characteristics within the group. Second, manual experience-based grouping relies on engineers' subjective judgment, with vague classification standards that cannot cope with the scale of millions and are difficult to adapt to dynamic business changes. Furthermore, random / round-robin aggregation has no scientific basis, and the grouping results are unrelated to resource consumption characteristics, constituting blind binding.

[0004] The aforementioned unscientific classification method leads to a high degree of heterogeneity in resource characteristics within domain name classes, rendering the quantitative results of key parameters such as CPU consumption intensity and I / O consumption intensity per unit bandwidth of domain name classes unusable. For example, the average resource consumption of a group of domain names that mix CPU-intensive and I / O-intensive domains is merely a mathematical result and cannot reflect the true resource requirements of the domain name class. Consequently, this quantitative data cannot be used as reliable input for scheduling models.

[0005] Due to the inability to accurately obtain quantitative indicators of multi-dimensional resources such as CPU and I / O for domain names, existing scheduling systems suffer from the following two core problems: Degenerates into single-dimensional scheduling: It can only allocate resources based solely on bandwidth, and cannot take into account the load balance of node CPU and IO, often resulting in one dimension of node resources being overloaded while the other dimension is idle. Unable to achieve resource complementarity and peak shifting: Unable to identify and combine CPU-intensive and IO-intensive domain name classes to achieve resource complementarity, and also unable to utilize the time sequence characteristics of domain name classes to achieve peak shifting deployment. Summary of the Invention

[0006] This invention proposes a CDN node deployment method and system for realizing multi-dimensional resource quantization and complementary scheduling, in order to solve the problems mentioned in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a CDN node deployment method for realizing multi-dimensional resource quantification and complementary scheduling, comprising the following steps: Step S1: Domain Name Operation Data Collection and Feature Vector Construction. First, the 24-hour period is divided into several time slices according to a preset time granularity. The time granularity ranges from 5 to 30 minutes to adapt to different business scenarios. Based on the divided time slices, real-time and historical bandwidth, CPU consumption, and IO consumption data of each domain name are collected from multiple source databases. Then, the CPU consumption intensity and IO consumption intensity corresponding to the unit bandwidth are calculated. The bandwidth data of each time slice are integrated to construct a high-dimensional bandwidth time series feature vector, which fully preserves the bandwidth fluctuation pattern and peak characteristics. Step S2: Domain-level multidimensional feature normalization. Given the differences in dimensions and numerical magnitudes among unit bandwidth CPU consumption intensity, unit bandwidth IO consumption intensity, and bandwidth time-series feature vectors, Min-Max or Z-Score normalization methods are used for domain-level unification. This domain-level normalization eliminates the magnitude differences between different domains, transforming multidimensional resource indicators into dimensionless standardized feature vectors while preserving the relative differences between features. Step S3: Domain name clustering and classification based on resource consumption characteristics. The clustering method is used to automatically group the standardized feature vectors. The silhouette coefficient is introduced into the clustering process to evaluate the classification quality. Domain names with high similarity in CPU consumption intensity, IO consumption intensity and bandwidth time series characteristics per unit bandwidth are aggregated into several domain name classes. The feature centroid of each domain name class is calculated. The centroid reflects the average resource consumption intensity and aggregated bandwidth time series characteristics of the domain names in the class. Step S4: Bandwidth-weighted multidimensional resource demand quantification. The cumulative bandwidth of all domains within each domain class is counted as the total bandwidth demand for that class. Using the proportion of each domain's bandwidth to the total bandwidth within the class as the weight, the CPU consumption intensity and IO consumption intensity per unit bandwidth of the domains within the class are weighted and averaged to obtain the comprehensive resource consumption intensity of the domain class. Combining the bandwidth data of each domain within the class in the corresponding time slice, the aggregated bandwidth time-series vector, CPU, and IO three-dimensional information of the domain class are formed to complete the resource demand quantification. Step S5: Construct a multi-dimensional resource complementary scheduling model, define binary allocation decision variables from domain name class to physical node, establish a constraint rule system, and the constraints include: complete allocation constraint of domain name class, bandwidth capacity constraint of physical node, CPU capability constraint, IO performance constraint and load stability constraint, and construct a system to suppress peak bandwidth load of nodes. Step S6: Model Solving and Scheduling Scheme Implementation. The intelligent resource scheduling engine uses linear programming to solve the scheduling model and obtain an approximate optimal allocation scheme from domain name class to physical node. Based on the bandwidth weight of the domain name within its domain name class, the class-level allocation scheme is decomposed into a domain-level scheduling list. The list is then sent to the physical nodes through the scheduling interface, and load monitoring is started simultaneously.

[0008] Furthermore, the present invention also includes a time-series bandwidth feature dimensionality reduction mechanism. The preset time granularity in step S1 is a configuration parameter with a value range of 5-30 minutes, corresponding to a number of time slices of 48-288. The preferred configuration is 15 minutes and divided into 96 time slices. The time-series bandwidth feature is represented as a 96-dimensional vector, and the time-series bandwidth vector can be dimensionality reduced by using discrete cosine transform or fast Fourier transform.

[0009] Furthermore, step S1 of the present invention also includes data cleaning and abnormal domain name identification and optimization steps, which involve using the 3σ principle to detect outliers in the original time series data and removing data points that exceed the mean ± 3σ range; filtering data records with missing rates exceeding a preset threshold; filling missing data with linear interpolation or the historical 7-day average; introducing trend consistency verification during the abnormal domain name identification process; and judging the consistency between the domain name bandwidth time series and domain names of the same business type by using the Pearson correlation coefficient.

[0010] Furthermore, in step S1 of the present invention, when the CPU consumption or IO consumption per unit bandwidth of a domain name exceeds three times the standard deviation of a preset threshold, it is marked as a domain name with abnormal resource consumption. When the fluctuation coefficient of the domain name's temporal bandwidth vector exceeds the threshold and exhibits non-periodic random fluctuation, it is marked as a domain name with abnormal bandwidth mode. Abnormal domain names are isolated or a special domain name class is established separately.

[0011] Furthermore, the node CPU capability rule in step S5 of this invention is specifically expressed as follows: Where c is a domain name index, , The CPU consumption intensity per unit bandwidth for the corrected domain class c. , Using the time decay formula The correction was obtained. The historical unit bandwidth CPU consumption intensity is represented by α, which is the time decay coefficient with a value range of 0.8-0.95, and t is the number of days since the current time. This correction reflects the time evolution pattern of the domain name's CPU consumption characteristics.

[0012] Furthermore, the node I / O performance rule in step S5 of this invention is specifically expressed as follows: Where c is a domain name index, Allocate the amount of bandwidth to node n for domain name class c. The domain name class c quantized in step S4 has a unit bandwidth IO consumption intensity. 0.80 is the IO capacity of node n, and 0.80 is the safety threshold coefficient, reserving 20% ​​resource redundancy to cope with sudden IO demand.

[0013] Furthermore, the node timing load rule in step S5 of this invention is specifically expressed as follows: Where c is the domain name index, t is the time slice index, and T is the total number of time slices divided into 24 hours. The amount of bandwidth allocated to node n for domain class c in time slice t. Let n be the bandwidth capacity of node n, and 0.85 be the safety threshold coefficient. At the same time, a node load stability constraint is established, which is achieved by calculating the variance of the node's load in each time slice over 24 hours. The constraint is set within the threshold range of variance ≤ 0.05-0.15 to suppress large fluctuations in node load.

[0014] Furthermore, the present invention also includes a domain name class centroid periodic update mechanism, which re-acquires the latest operating data of all domain names on the entire network at a preset period of midnight every day, and repeatedly executes steps S1 to S3 to update the feature vector, class centroid, and resource requirement quantification parameters of the domain name class; at the same time, it has a new domain name fast access mechanism. When a new domain name is accessed, a standardized feature vector is constructed according to steps S1 and S2, and the Euclidean distance or cosine similarity between the vector and the existing K domain name class centroids is calculated. The new domain name is assigned to the domain name class with the highest similarity. When the similarity is lower than a preset threshold, a new domain name class is created without performing full clustering.

[0015] Furthermore, this invention also proposes a CDN node scheduling system for implementing the aforementioned CDN node deployment method that achieves multi-dimensional resource quantization and complementary scheduling, comprising: The data acquisition and preprocessing module is used to collect historical domain name operation data from the database, sample and perform cleaning and verification according to preset time granularity, support multi-source database access and realize unified data format conversion; The feature engineering module is used to extract the CPU consumption intensity, IO consumption intensity and timing bandwidth features of domain names, and generates standardized feature vectors by using Min-Max normalization or Z-Score normalization, and integrates feature dimensionality reduction function. The domain name scientific classification module is used to cluster standardized feature vectors based on the K-Means++ algorithm, generate several domain name classes and class centroids, and support rapid classification and creation of new classes based on the similarity between new domain name feature vectors and existing class centroids. The resource demand quantification module is used to quantify the unit bandwidth CPU consumption intensity and unit bandwidth IO consumption intensity of each domain name class using a bandwidth-weighted average method, and introduces a time decay factor to correct the resource demand parameters. The multi-dimensional resource scheduling decision module is used to construct the scheduling model of the constraint rules, solve for the allocation scheme from domain name class to physical node, and support dynamic optimization of model parameters; The deployment, execution, and monitoring module is used to expand the domain name allocation strategy into a domain name-level scheduling list for execution, and to monitor node load and peak-shaving effects. The time granularity, number of target categories, clustering algorithm type, and normalization method are all configurable parameters that can be flexibly adjusted according to business needs.

[0016] Compared with existing technologies, the beneficial effects of this invention are: (1) Automated classification of millions of domain names: This invention integrates the unit bandwidth CPU consumption intensity, IO consumption intensity and bandwidth time-series fluctuation characteristics based on the actual resource consumption characteristics of domain names to construct a multi-dimensional index system. Through clustering algorithm and contour coefficient quality verification, millions of heterogeneous domain names are automatically classified into thousands of standardized domain name classes. After classification, the resource consumption characteristics of domain names within the class are highly homogeneous. The average resource consumption data of the class centroid can accurately represent the overall characteristics within the class, laying a scientific classification foundation for subsequent resource scheduling.

[0017] (2) Accurate Quantification of Multidimensional Resource Requirements for Domain Names: For the classified domain name classes, this invention introduces a bandwidth-weighted average method. Using the bandwidth proportion of each domain name within the class as the weight, the CPU and IO consumption intensity per unit bandwidth is calculated in a weighted manner. Combined with the aggregated bandwidth time-series vector within the class, the core resource requirement indicators of the domain name class are accurately quantified. This quantification result fully reflects the comprehensive characteristics of domain name resource consumption within the class, avoiding inaccurate resource assessment caused by single-sample bias.

[0018] (3) Three-dimensional resource collaborative optimization scheduling: This invention constructs a three-dimensional constraint system of CPU, IO and bandwidth. Through a multi-dimensional resource complementary scheduling model, CPU-intensive and IO-intensive domain name classes are deployed in a complementary manner at the node level to achieve resource load balance. At the same time, combined with bandwidth time sequence characteristics, the load superposition of domain name classes with the same time sequence characteristics is avoided by staggered deployment, effectively suppressing the peak load of nodes, improving the 24-hour load stability, and taking into account both resource utilization and operational stability. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of a CDN global scheduling method based on multi-dimensional resource complementarity proposed in this invention; Figure 2 A bar chart showing the average similarity distribution within domain name classes; Figure 3 A line graph showing the timing comparison of node bandwidth load before and after scheduling; Figure 4 Grouped bar charts comparing multidimensional resource utilization before and after scheduling. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A CDN node deployment method for achieving multi-dimensional resource quantification and complementary scheduling includes the following steps: Step S1: Domain Name Operation Data Collection and Feature Vector Construction. First, the 24-hour period is divided into several time slices according to a preset time granularity. The time granularity ranges from 5 to 30 minutes to adapt to different business scenarios. Based on the divided time slices, real-time and historical bandwidth, CPU consumption, and IO consumption data of each domain name are collected from multiple source databases. Then, the CPU consumption intensity and IO consumption intensity corresponding to the unit bandwidth are calculated. The bandwidth data of each time slice are integrated to construct a high-dimensional bandwidth time series feature vector, which fully preserves the bandwidth fluctuation pattern and peak characteristics. Step S2: Domain-level multidimensional feature normalization. Given the differences in dimensions and numerical magnitudes among unit bandwidth CPU consumption intensity, unit bandwidth IO consumption intensity, and bandwidth time-series feature vectors, Min-Max or Z-Score normalization methods are used for domain-level unification. This domain-level normalization eliminates the magnitude differences between different domains, transforming multidimensional resource indicators into dimensionless standardized feature vectors while preserving the relative differences between features. Step S3: Domain name clustering and classification based on resource consumption characteristics. The clustering method is used to automatically group the standardized feature vectors. The silhouette coefficient is introduced into the clustering process to evaluate the classification quality. Domain names with high similarity in CPU consumption intensity, IO consumption intensity and bandwidth time series characteristics per unit bandwidth are aggregated into several domain name classes. The feature centroid of each domain name class is calculated. The centroid reflects the average resource consumption intensity and aggregated bandwidth time series characteristics of the domain names in the class. Step S4: Bandwidth-weighted multidimensional resource demand quantification. The cumulative bandwidth of all domains within each domain class is counted as the total bandwidth demand for that class. Using the proportion of each domain's bandwidth to the total bandwidth within the class as the weight, the CPU consumption intensity and IO consumption intensity per unit bandwidth of the domains within the class are weighted and averaged to obtain the comprehensive resource consumption intensity of the domain class. Combining the bandwidth data of each domain within the class in the corresponding time slice, the aggregated bandwidth time-series vector, CPU, and IO three-dimensional information of the domain class are formed to complete the resource demand quantification. Step S5: Construct a multi-dimensional resource complementary scheduling model, define binary allocation decision variables from domain name class to physical node, establish a constraint rule system, and the constraints include: complete allocation constraint of domain name class, bandwidth capacity constraint of physical node, CPU capability constraint, IO performance constraint and load stability constraint, and construct a system to suppress peak bandwidth load of nodes. Step S6: Model Solving and Scheduling Scheme Implementation. The intelligent resource scheduling engine uses linear programming to solve the scheduling model and obtain an approximate optimal allocation scheme from domain name class to physical node. Based on the bandwidth weight of the domain name within its domain name class, the class-level allocation scheme is decomposed into a domain-level scheduling list. The list is then sent to the physical nodes through the scheduling interface, and load monitoring is started simultaneously.

[0024] The present invention also includes a time-series bandwidth feature dimensionality reduction mechanism. The preset time granularity in step S1 is a configuration parameter with a value range of 5-30 minutes, corresponding to a number of time slices of 48-288. The preferred configuration is 15 minutes and divided into 96 time slices. The time-series bandwidth feature is represented as a 96-dimensional vector, and the time-series bandwidth vector can be dimensionality reduced by using discrete cosine transform or fast Fourier transform.

[0025] Step S1 of this invention also includes data cleaning and abnormal domain name identification and optimization steps. The original time series data is subjected to outlier detection using the 3σ principle, and data points exceeding the mean ± 3σ range are removed. Data records with missing rates exceeding a preset threshold are filtered out. Missing data are filled using linear interpolation or the historical 7-day average. Trend consistency verification is introduced in the abnormal domain name identification process, and the consistency between the domain name bandwidth time series and domain names of the same business type is judged by the Pearson correlation coefficient.

[0026] In step S1 of this invention, when the CPU consumption or IO consumption per unit bandwidth of a domain name exceeds three times the standard deviation of a preset threshold, it is marked as a domain name with abnormal resource consumption. When the fluctuation coefficient of the domain name's temporal bandwidth vector exceeds the threshold and exhibits non-periodic random fluctuation, it is marked as a domain name with abnormal bandwidth mode. Abnormal domain names are isolated or a special domain name class is established separately.

[0027] The node CPU capability rule in step S5 of this invention is specifically expressed as follows: Where c is a domain name index, , The CPU consumption intensity per unit bandwidth for the corrected domain class c. , Using the time decay formula The correction was obtained. The historical unit bandwidth CPU consumption intensity is represented by α, which is the time decay coefficient with a value range of 0.8-0.95, and t is the number of days since the current time. This correction reflects the time evolution pattern of the domain name's CPU consumption characteristics.

[0028] The node I / O performance rule in step S5 of this invention is specifically expressed as follows: Where c is a domain name index, Allocate the amount of bandwidth to node n for domain name class c. The domain name class c quantized in step S4 has a unit bandwidth IO consumption intensity. 0.80 is the IO capacity of node n, and 0.80 is the safety threshold coefficient, reserving 20% ​​resource redundancy to cope with sudden IO demand.

[0029] The node timing load rule in step S5 of this invention is specifically expressed as follows: Where c is the domain name index, t is the time slice index, and T is the total number of time slices divided into 24 hours. The amount of bandwidth allocated to node n for domain class c in time slice t. Let n be the bandwidth capacity of node n, and 0.85 be the safety threshold coefficient. At the same time, a node load stability constraint is established, which is achieved by calculating the variance of the node's load in each time slice over 24 hours. The constraint is set within the threshold range of variance ≤ 0.05-0.15 to suppress large fluctuations in node load.

[0030] This invention also includes a domain name class centroid periodic update mechanism, which re-acquires the latest operating data of all domain names on the entire network at a preset period of midnight every day, and repeats steps S1 to S3 to update the feature vector, class centroid, and resource requirement quantification parameters of the domain name class; at the same time, it has a new domain name fast access mechanism. When a new domain name is accessed, a standardized feature vector is constructed according to steps S1 and S2, and the Euclidean distance or cosine similarity between the vector and the existing K domain name class centroids is calculated. The new domain name is assigned to the domain name class with the highest similarity. When the similarity is lower than a preset threshold, a new domain name class is created without performing full clustering.

[0031] This invention also proposes a CDN node scheduling system for implementing the aforementioned CDN node deployment method that achieves multi-dimensional resource quantization and complementary scheduling, comprising: The data acquisition and preprocessing module is used to collect historical domain name operation data from the database, sample and perform cleaning and verification according to preset time granularity, support multi-source database access and realize unified data format conversion; The feature engineering module is used to extract the CPU consumption intensity, IO consumption intensity and timing bandwidth features of domain names, and generates standardized feature vectors by using Min-Max normalization or Z-Score normalization, and integrates feature dimensionality reduction function. The domain name scientific classification module is used to cluster standardized feature vectors based on the K-Means++ algorithm, generate several domain name classes and class centroids, and support rapid classification and creation of new classes based on the similarity between new domain name feature vectors and existing class centroids. The resource demand quantification module is used to quantify the unit bandwidth CPU consumption intensity and unit bandwidth IO consumption intensity of each domain name class using a bandwidth-weighted average method, and introduces a time decay factor to correct the resource demand parameters. The multi-dimensional resource scheduling decision module is used to construct the scheduling model of the constraint rules, solve for the allocation scheme from domain name class to physical node, and support dynamic optimization of model parameters; The deployment, execution, and monitoring module is used to expand the domain name allocation strategy into a domain name-level scheduling list for execution, and to monitor node load and peak-shaving effects. The time granularity, number of target categories, clustering algorithm type, and normalization method are all configurable parameters that can be flexibly adjusted according to business needs.

[0032] The following two examples further illustrate the specific implementation of this system: Example 1: Implementation of Multidimensional Resource Quantization and Scheduling in a Medium-Sized CDN Network This embodiment focuses on a medium-sized CDN network covering 30 provincial-level administrative regions in China, containing 30 physical edge nodes, and carrying 500,000 domain names. It elaborates on the complete implementation process of a CDN node deployment method that achieves multi-dimensional resource quantification and complementary scheduling, and verifies the technical feasibility and advantages through full-process implementation.

[0033] Step 1: Domain Name Operation Data Collection and Feature Vector Construction In the data collection phase, a multi-source database adapted to CDN business scenarios was selected as the data source. Operational metrics data for 500,000 domains over the past 7 days were read, with core metrics including bandwidth, CPU consumption, and IO consumption at various time points. To balance data granularity and computational efficiency, a sampling frequency of 5 minutes was set, resulting in 288 sampling points per day, ensuring complete capture of daily changes in the domains' operational status.

[0034] Data cleaning and verification are crucial steps to ensure the accuracy of subsequent analysis, and a multi-layered verification mechanism is employed: First, outlier detection is performed using the 3σ principle. The mean and standard deviation of each indicator are calculated, and data points exceeding the range of mean plus or minus 3 times the standard deviation are removed, effectively filtering out extreme data caused by sudden failures. Next, a data integrity threshold is set to filter domain data with a missing rate exceeding 10% to avoid incomplete data affecting the overall analysis quality; for sporadic missing data, linear interpolation or the average of the same period in the past 7 days is used to fill in the missing data to ensure data continuity.

[0035] After cleaning, data from 485,000 valid domain names were retained, providing a high-quality data source for feature construction.

[0036] Feature extraction and standardization are performed as follows: The CPU consumption and I / O consumption per unit bandwidth for each domain are calculated. CPU consumption per unit bandwidth is the ratio of CPU consumption to bandwidth, expressed as a percentage per Mbps; I / O consumption per unit bandwidth is the ratio of I / O consumption to bandwidth, expressed as MB per second per Mbps. When constructing the time-series bandwidth vector, the 24 hours are divided into 96 time slices, each 15 minutes long. The average bandwidth value for each domain in each time slice is calculated, forming a 96-dimensional time-series bandwidth vector. This vector fully characterizes the daily access cycle pattern and peak distribution features of the domain.

[0037] Given the differences in units and magnitudes of CPU consumption per unit bandwidth, I / O consumption per unit bandwidth, and time-series bandwidth vectors, a Min-Max normalization method is used to normalize all features to the 0-1 range, generating a normalized feature vector for each domain name. The normalization formula is as follows: in, The values ​​are normalized, and X is the original feature value. This is the minimum value of the feature. This represents the maximum value of the feature. Standardization eliminates the order-of-magnitude differences between different domain names while preserving the relative differences between features, providing a high-quality data foundation that meets the algorithm requirements for subsequent clustering analysis.

[0038] A trend consistency check is introduced during the abnormal domain name identification process. The Pearson correlation coefficient is used to determine the consistency between the domain name's bandwidth time series and that of domain names with the same business type. The correlation coefficient formula is as follows: in, This represents the bandwidth value for domain name i in time slice t. This represents the average bandwidth value of domains with the same business type in time slice t. The average bandwidth value for domain name i. is the overall average bandwidth value for domains of the same business type, r is the correlation coefficient, and T is the total number of time slices.

[0039] When a domain name's unit bandwidth CPU consumption or IO consumption exceeds a threshold, its bandwidth pattern exhibits non-periodic random fluctuations, or its correlation coefficient absolute value is ≤0.3, it is marked as an abnormal domain name. In this embodiment, a total of 15,000 abnormal domain names were identified, and they were isolated and a special domain name class was established separately to avoid interfering with the classification and scheduling of normal domain names.

[0040] Step Two: Scientific Classification of Domain Names Based on Resource Consumption Characteristics Domain name classification is the core prerequisite for accurate quantification. Unlike traditional grouping methods based on customer relationships, human experience, or randomness, this embodiment uses scientific clustering based on the actual resource consumption characteristics of domain names to ensure homogeneity of domain name characteristics within a category.

[0041] The K-Means++ algorithm was selected for clustering. This algorithm effectively reduces the sensitivity of clustering results to initial values ​​and minimizes the risk of getting trapped in local optima by optimizing the initial centroid selection strategy. Considering the domain size and business requirements, the target number of domain classes was set at 1000, which is 0.2% of the total number of domains on the entire network, falling within a reasonable range of 0.1%-1%, and meeting the constraint of no less than 500 and no more than 2000. The clustering process uses standardized feature vectors as input. The core idea is to group domains with closer distances into one class by calculating the Euclidean distance between feature vectors, ensuring that domains within the same class have high similarity in terms of CPU consumption intensity, IO consumption intensity, and temporal bandwidth characteristics.

[0042] Cluster convergence control employs a triple criterion: iteration is stopped when the relative rate of change of the sum of squared errors within two consecutive iterations is less than 0.001, the cluster affiliation of all sample points no longer changes, or the number of iterations reaches 100.

[0043] In this embodiment, when the iteration reaches the 32nd iteration, the relative change rate of the sum of squared errors within two consecutive iterations decreases to 0.0008, satisfying the convergence condition, and the iteration stops. After clustering is completed, the centroid of each domain name class is calculated. The centroid represents the typical resource consumption characteristics of that domain name class in the normalized feature space, including average CPU consumption intensity, average I / O consumption intensity, and typical temporal bandwidth characteristics.

[0044] To verify the classification effect, the silhouette coefficient was introduced as a clustering quality assessment index. The silhouette coefficient was calculated as follows: ,in This represents the average distance between sample i and other samples in the same cluster, i.e., the average distance of sample i within its own cluster, reflecting the cluster compactness. This is the average distance between sample i and the nearest heterogeneous cluster, which is the minimum value among the average distances from sample i to all other clusters, reflecting the inter-cluster separation. Let be the silhouette coefficient of sample i. Statistics show that the average silhouette coefficient of all samples in this embodiment is 0.68, which is greater than the judgment threshold of 0.6, indicating a good clustering effect. Meanwhile, the resource characteristic similarity of domain names within the same category reaches over 0.82, far exceeding the 0.45 similarity achieved by traditional customer grouping. This high similarity ensures that the average resource consumption within the category accurately represents the true characteristics of that category, laying a solid foundation for subsequent accurate quantification.

[0045] Step 3: Quantification of multidimensional resource requirements based on bandwidth-weighted average This step achieves accurate quantification of multi-dimensional resource requirements for domain names based on the classification results. The specific process is as follows: Calculate the total bandwidth requirement for each domain class and sum the bandwidth of all domains within each domain class. Taking domain class 1 as an example, this class contains 500 domains, and the total bandwidth is calculated to be 50,000 Mbps.

[0046] The CPU consumption intensity per unit bandwidth of each domain class is calculated using a bandwidth-weighted average method. The core principle is that domains with larger bandwidth have more stable and reliable resource consumption data, so they are given greater weight. By using the stable data of large-bandwidth domains to suppress the noisy data of small-bandwidth domains, accurate quantitative parameters can be obtained.

[0047] The specific calculation steps are as follows: The first step is to calculate the product of the CPU consumption and the bandwidth of each domain in the domain class. The second step is to sum up the products of all domain names within the class to obtain the weighted total CPU consumption. The third step is to divide the weighted total CPU consumption by the total bandwidth of that class to obtain the CPU consumption intensity per unit bandwidth for that class. This can be expressed mathematically as follows: ,in, CPU consumption per unit bandwidth for domain names. The CPU consumption of the i-th domain within the class. Let N be the bandwidth of the i-th domain within the class, and N be the total number of domains within the class. This represents the total bandwidth for this domain class.

[0048] Taking domain class 1 as an example, its total bandwidth is 50,000 Mbps. The sum of the product of CPU consumption and bandwidth of each domain in the class is 15,000 percent Mbps. By calculation, the CPU consumption per unit bandwidth is 15,000 percent Mbps divided by 50,000 Mbps, which is 0.3 percent per Mbps. This value can be converted into 8G bandwidth that a 24-core CPU can carry at full load, which is highly consistent with the actual operating data.

[0049] The calculation of I / O consumption intensity per unit bandwidth incorporates a time decay factor for dynamic correction. The correction formula is as follows: in, This represents the corrected I / O consumption intensity per unit bandwidth. Based on the bandwidth I / O consumption intensity per unit, CV is the coefficient of variation correction factor, where CV is the coefficient of variation of bandwidth time series for domain class. Taking domain class 1 as an example, if the historical unit bandwidth I... O If the consumption intensity is 0.32 percentage points per Mbps and there are 3 days from the current time, then the corrected value is 0.32 multiplied by 0.9 cubed, which results in 0.233 percentage points per Mbps. This correction reflects the time evolution of domain name IO consumption characteristics, avoiding overload or waste caused by allocating resources based on outdated data.

[0050] The unit bandwidth I / O consumption intensity of each domain class is calculated using the same bandwidth-weighted averaging method as CPU consumption intensity. The specific calculation steps are as follows: First, for each domain in the domain class, calculate the product of its I / O consumption and its bandwidth; second, sum the products of all domains in the class to obtain the weighted total I / O consumption; third, divide the weighted total I / O consumption by the total bandwidth of the class to obtain the unit bandwidth I / O consumption intensity of that class. The mathematical expression is as follows: ,in, For domain-level bandwidth I / O consumption intensity, For the I / O consumption of the i-th domain within the class, Let N be the bandwidth of the i-th domain within the class, and N be the total number of domains within the class. This represents the total bandwidth of the domain class. Taking domain class 1 as an example, the sum of the product of the IO consumption and bandwidth of each domain within the class is 37500 MB·Mbps per second, and the total bandwidth is 50000Mbps. The calculated IO consumption per unit bandwidth is 37500 MB·Mbps per second divided by 50000Mbps, which gives 0.75 MB per second per Mbps.

[0051] Calculate the standardized time-series bandwidth vector for each domain class, and statistically analyze the average bandwidth of each domain class in each time slice to form a 96-dimensional time-series bandwidth vector for that class, thus fully preserving the intraday access pattern characteristics of domains within the class.

[0052] To verify the quantification effect, the deviation between the quantified data and the online monitoring data was compared. The results showed that the deviation rate of CPU consumption intensity per unit bandwidth was less than 3%, and the deviation rate of IO consumption intensity per unit bandwidth was less than 4%. This indicates that the quantified data can accurately predict how much CPU and IO resources will be consumed by deploying a certain domain name, providing a reliable data foundation for subsequent multi-dimensional resource complementary scheduling.

[0053] Step 4: Construct a multi-dimensional resource complementary scheduling model Based on the accurate parameters obtained from step three quantization, a multi-dimensional resource complementary scheduling model is constructed to achieve collaborative optimization of CPU, IO, and bandwidth resources.

[0054] Define the decision variable as assign[c,n], which represents the amount of bandwidth allocated to physical node n for domain name class c. Here, c takes values ​​from 1 to 1000, representing 1000 domain name classes; n takes values ​​from 1 to 30, representing 30 physical edge nodes. The scale of the decision variable is 30000, which is suitable for large-scale scheduling requirements.

[0055] Establish a resource constraint rule system, covering four core constraints: The domain name allocation rules ensure that the bandwidth requirements of each domain name class are fully met. For each domain name class c, within each scheduling cycle, the sum of the bandwidth allocated to all physical nodes for that class must equal the current total bandwidth requirement for that class. The mathematical expression is as follows: ,in, Allocate the amount of bandwidth to node n for domain name class c. Let N be the total bandwidth requirement for domain class c, and N be the total number of physical nodes.

[0056] Node bandwidth capacity rules are implemented to avoid node bandwidth overload. For each physical node n, the sum of bandwidth allocated to that node for all domain names shall not exceed 85% of that node's bandwidth capacity, with a 15% redundancy reserved to handle sudden traffic surges. The mathematical expression is as follows: ,in, Allocate the amount of bandwidth to node n for domain name class c. Let C be the bandwidth capacity of node n, C be the total number of domain name classes, and 0.85 be the bandwidth security threshold coefficient.

[0057] The node CPU capacity rule uses dynamically adjusted consumption intensity to ensure constraint accuracy. For each physical node n, the total CPU consumption of all domain names on that node shall not exceed 85% of the node's CPU capacity.

[0058] Node IO performance rules ensure balanced IO resource load. For each physical node n, the total IO consumption of all domain names on that node shall not exceed 80% of the node's IO capacity, with a 20% redundancy reserved to cope with sudden IO demands.

[0059] Node load stability constraints ensure that node time-series load fluctuations are controllable. For each physical node n, the variance of its load rate in each time slice within 24 hours must be controlled within a preset threshold range to avoid drastic fluctuations in node load over time.

[0060] An optimization objective function is constructed, defined as minimizing the sum of peak loads across all physical nodes in the network to achieve time-series load balancing. The optimization objective function is as follows: Single-objective optimization objective function formula in Let N be the peak load of node n, and N be the total number of physical nodes. Formula for the objective function of multi-objective optimization: ,in, Let n be the peak load of node n. For delay weighting coefficients, Allocate the amount of bandwidth to node n for domain name class c. The transmission delay allocated to node n for domain class c, where C is the total number of domain classes and N is the total number of physical nodes.

[0061] Multi-dimensional collaborative optimization is achieved through a dual mechanism: On the one hand, constrained by node CPU capacity rules and node IO performance rules, the scheduling engine deploys CPU-intensive and IO-intensive domain name classes in a complementary manner. For example, it deploys domain name classes with high CPU consumption and low IO consumption on the same node as domain name classes with low CPU consumption and high IO consumption, thereby achieving complementary balance in the resource dimension. On the other hand, with the optimization goal of minimizing the sum of node peak values, the scheduling engine deploys different domain name classes during peak access periods in a staggered manner, thereby achieving stable load in the time dimension.

[0062] The effectiveness of staggered deployment can be illustrated with an example: Domain Class 1 has a bandwidth of 500Mbps in time slice 10 and 1200Mbps in time slice 50; Domain Class 2 has a bandwidth of 1200Mbps in time slice 10 and 240Mbps in time slice 50. Deploying both domain classes on the same node results in a total bandwidth load of 500Mbps + 1200Mbps = 1700Mbps in time slice 10 and 1200Mbps + 240Mbps = 1440Mbps in time slice 50. The peak access times for the two domain classes are staggered, with Domain Class 2 contributing more bandwidth during the morning peak and Domain Class 1 contributing more during the afternoon peak, avoiding excessive bandwidth load accumulation within a single period and achieving a stable and balanced time-series load on the node.

[0063] Step 5: Model Solving and Generation of Domain-Level Scheduling Scheme The intelligent optimization solution engine is invoked to solve the scheduling model. The solution algorithm uses linear programming, which is an aggregate of algorithms. It employs branch and bound and simplex algorithms as its framework, supplemented by various heuristic algorithms. This algorithm can find an approximate optimal solution for a large number of decision variables within a finite time, adapting to the real-time scheduling requirements of CDN. In this embodiment, the number of decision variables is 1000 multiplied by 30, equaling 30000, and the solution time is controlled within one minute, fully meeting the real-time scheduling requirements of large-scale CDN systems.

[0064] After solving, the optimal bandwidth allocation (assign[c,n]) ​​for each domain class to each physical node is obtained. Taking domain class 1 as an example, its total bandwidth is 50000Mbps. The solution is that the bandwidth allocated to node 1 is 10000Mbps, the bandwidth allocated to node 2 is 7500Mbps, the bandwidth allocated to node 3 is 5000Mbps, and the remaining bandwidth is allocated to the remaining 27 nodes according to node capacity and load balancing requirements, ensuring that the constraints of each node are met.

[0065] Based on the solution results, the domain name allocation scheme is expanded into a domain-level scheduling list: For each domain name class c, the proportion of allocated bandwidth to each node relative to the total bandwidth of that class is calculated, and the bandwidth of each domain name within the class is allocated to each node in the same proportion. Taking domain name class 1 as an example, the proportion of bandwidth allocated to node 1 relative to the total bandwidth of that class is 10000Mbps divided by 50000Mbps, which equals 20%. Therefore, 20% of the bandwidth of each domain name within that class is allocated to node 1; the proportion allocated to node 2 is 15%, so 15% of the bandwidth of each domain name is allocated to node 2, and so on, to complete the bandwidth allocation for all domain names.

[0066] The scheduling list is converted into a configuration command format that the system can recognize and sent to each physical node for execution through the scheduling interface. At the same time, the load monitoring module is started to collect bandwidth, CPU, and IO load data of each node in real time, providing data support for scheduling effect verification and subsequent optimization.

[0067] Step Six: Verification and Comparative Analysis of Scheduling Effects To verify the effectiveness of the technology, the resource utilization of 30 physical nodes before and after scheduling was compared. The statistical data is shown in Table 1 below.

[0068] Table 1 Comparison of physical node resource utilization before and after scheduling

[0069] Data analysis results, as shown in Table 1, intuitively present the core optimization effects of physical node resource utilization before and after scheduling. Regarding resource balance, the CPU utilization variance decreased from 0.125 to 0.068, a reduction of 45.6%; the IO utilization variance decreased from 0.132 to 0.079, a reduction of 40.2%. This significant decrease in variance indicates a substantial reduction in resource consumption differences among nodes, completely changing the unbalanced state of some nodes being overloaded and others idle before the implementation of this solution. In terms of peak load control, the average node peak load decreased from 71Gbps to 62Gbps, a reduction of 12.7%, reflecting the synergistic effect of time-series peak shifting and multi-dimensional complementary scheduling, effectively avoiding a surge in node pressure caused by load superposition. Overall, the data confirms the effectiveness of this method in optimizing resource distribution and alleviating peak pressure, providing strong support for the stable and efficient operation of the CDN network.

[0070] Step 7: Dynamic Maintenance and Continuous Optimization To adapt to the dynamic changes in domain name operation characteristics, a mechanism for periodic updates and rapid access of new domain names should be established to ensure the continuous effectiveness of the scheduling strategy.

[0071] The periodic update mechanism executes at a fixed time of 2 AM daily, re-acquiring the latest operational data of all 500,000 domains across the network, and repeating steps one through three to update the domain classification results and resource demand quantification parameters, ensuring that scheduling decisions are always based on the latest domain operational characteristics. The update process employs an incremental optimization strategy. By calculating the rate of change of domain resource consumption characteristics, for domains with a change rate below 10%, historical classification results are retained; only for domains with a change rate exceeding 10% are the feature extraction and clustering process re-executed. This strategy reduces the update time from 2 hours for a full update to less than 30 minutes, significantly improving update efficiency.

[0072] The new domain name fast access mechanism is designed for newly added domain names. When a new domain name is accessed by the CDN system, its operational data is first collected according to step one. After cleaning and verification, a standardized feature vector is generated according to step two. The weighted cosine similarity between this feature vector and the centroids of 1000 existing domain names is calculated. The similarity formula is: ,in Let k be the k-th dimension value of the new domain name feature vector. The value of the k-th dimension of the centroid eigenvector is... The weight of the k-th feature is determined by calculating the contribution of each feature to the classification result during the clustering process; the higher the contribution, the greater the weight. M is the total number of feature dimensions, and cosθ is the weighted cosine similarity value between the centroids of the new domain and the target domain, ranging from -1 to 1. The closer the value is to 1, the higher the similarity between the two features. If the maximum similarity is greater than or equal to 0.75, the new domain is assigned to that domain class; if all similarities are less than 0.75, a new domain class is created and the class centroid list is updated. This mechanism eliminates the need to re-execute full clustering, reducing the time for new domain access from 2 hours in traditional methods to 5 minutes, significantly improving operational efficiency while ensuring the stability of the classification system.

[0073] Step 8: Scalability Verification To verify the scalability of the technology, tests were conducted on scenarios involving expansion of the number of nodes and the size of domain names: In the node expansion test, 10 new nodes were added to the existing 30 physical nodes, bringing the total number of nodes to 40. The expansion process only required adding constraints such as bandwidth capacity, CPU capacity, and IO capacity for the 10 new nodes to the scheduling model; no modification to the core model framework or solution logic was necessary. Solution tests showed that even with the addition of nodes increasing the number of decision variables to 40,000, the solution time remained within 1.2 minutes, essentially the same as the original solution time, demonstrating the technology's excellent node expansion capabilities.

[0074] In the domain name scaling test, the number of domain names was increased from 500,000 to 1 million. After performing the classification according to steps one through three, the number of domain name classes remained at around 1,000, and the size of the decision variable did not increase significantly with the increase in the number of domain names. The solution results show that the solution time after doubling the domain name size is 1.1 minutes, which is only slightly different from the solution time at the original size. This proves that the scaling achieved through scientific classification ensures that the scheduling performance is not significantly affected by the increase in the number of domain names, and has the ability to support ultra-large-scale CDN networks.

[0075] Example 2: Implementation of Multidimensional Resource Quantization and Scheduling in Large-Scale CDN Networks This embodiment focuses on a large-scale CDN network covering major regions globally, containing 100 physical edge nodes, and carrying 2 million domain names. It illustrates the process of scaling up the application of the technology and verifies its feasibility and stability in ultra-large-scale scenarios.

[0076] Step 1: Domain Name Operation Data Collection and Feature Vector Construction Data collection employs a distributed data collection architecture, adapting to the data aggregation needs of globally distributed CDN nodes. It collects operational metrics data for nearly 14 days from 2 million domains, with core metrics including bandwidth, CPU consumption, and IO consumption. Considering the regional differences in domain access across a large-scale network, the sampling frequency is set to once every 5 minutes, resulting in 288 sampling points per day. Simultaneously, it records the regional distribution data of domain access, providing support for subsequent regional adaptation in scheduling.

[0077] Data cleaning and verification employed a multi-layered mechanism consistent with Example 1, using the 3σ principle to remove outliers and filtering domain names with a missing rate exceeding 10%. Missing data was filled using linear interpolation or the historical 14-day average. In addition to using Pearson correlation coefficients for trend consistency verification, abnormal domain name identification also incorporated regional access characteristics. Domains with over 30% access from non-target service regions were marked as having regionally abnormal access. After cleaning, 1.92 million valid domain names were retained, and 80,000 abnormal domain names were identified and isolated.

[0078] During feature extraction and standardization, the calculation methods for CPU consumption and IO consumption per unit bandwidth are consistent with those in Example 1. When constructing the time-series bandwidth vector, 24 hours are divided into 96 time slices, and the peak bandwidth of each domain name in each time slice is statistically analyzed to form a 96-dimensional time-series bandwidth vector. Simultaneously, regional access ratio features are integrated to construct a multi-dimensional feature system. The standardization process uses the Z-Score standardization method, with the formula: ,in, Here, X represents the normalized values, and X represents the original eigenvalues. The mean of this feature. The standard deviation of this feature is used to adapt to the standardization requirements of large-scale data distributions.

[0079] Dimensionality reduction of time-series bandwidth features employs the discrete cosine transform method to reduce the computational complexity of high-dimensional data. The dimensionality reduction formula is as follows: ,in, The k-th eigenvalue after dimensionality reduction. This represents the original value of the t-th dimension of the time-series bandwidth vector. The discrete cosine transform kernel function. hour , hour Where T is the total number of time slices, and k is the feature dimension after dimensionality reduction. In this embodiment, the 96-dimensional time series vector is reduced to 20 dimensions. The effectiveness of the dimensionality reduction is verified by the reconstruction error, which is expressed by the formula: Where E is the reconstruction error, This represents the original value of the t-th dimension of the time-series bandwidth vector. Let t be the value of the t-th dimension of the reconstructed temporal bandwidth vector, where T is the total number of time slices. The calculated reconstruction error is 0.035, which is less than the 0.05 threshold, indicating that the core features are still fully preserved after dimensionality reduction, thus reducing the burden and improving the efficiency of subsequent clustering calculations.

[0080] Step Two: Scientific Classification of Domain Names Based on Resource Consumption Characteristics The clustering algorithm still uses the K-Means++ algorithm. Considering the scale of 2 million domain names, the target number of domain name classes is set at 1500, which is 0.75% of the total number of domain names on the entire network, within a reasonable range of 0.1%-1%, and meets the constraint of not less than 500 and not more than 2000. The clustering input is a standardized feature vector, including dimensionality-reduced time-series bandwidth features, CPU consumption per unit bandwidth, IO consumption per unit bandwidth, and geographic access features, ensuring that the classification results take into account multi-dimensional resource consumption characteristics and business scenario requirements.

[0081] Cluster convergence control adopted the same triple criterion as in Example 1. By the 45th iteration, the relative change rate of the sum of squared errors within two consecutive iterations decreased to 0.0009, satisfying the convergence condition. After clustering, the centroid of each domain name class was calculated. The centroid includes average CPU consumption intensity, average I / O consumption intensity, typical time-series bandwidth characteristics, and typical regional access characteristics, comprehensively characterizing the core characteristics of domain names within the class.

[0082] Clustering quality assessment shows that the average silhouette coefficient of all samples is 0.65, which is greater than the judgment threshold of 0.6, indicating good clustering effect. The resource characteristic similarity of domain names within the same category reaches above 0.80, and the regional access characteristic similarity reaches above 0.85, ensuring a high degree of homogeneity of domain names within the category in terms of resource consumption and business scenarios, laying the foundation for subsequent accurate quantification and scheduling.

[0083] Step 3: Quantification of multidimensional resource requirements based on bandwidth-weighted average Total bandwidth requirement is calculated by summing the bandwidth of all domains within each domain class. For example, domain class 500 contains 1200 domains and has a total bandwidth of 120,000 Mbps.

[0084] The CPU consumption intensity per unit bandwidth is calculated using a bandwidth-weighted average method, with the formula consistent with Example 1. The weighted total CPU consumption for domain class 500 is 36,000 percentage points per Mbps, and the total bandwidth is 120,000 Mbps. Therefore, the CPU consumption per unit bandwidth is calculated as 36,000 percentage points per Mbps divided by 120,000 Mbps, resulting in 0.3 percentage points per Mbps. A time decay factor is introduced for correction, with a time decay coefficient α of 0.92. Five days from the current time, the historical CPU consumption intensity per unit bandwidth is 0.33 percentage points per Mbps. The corrected value is 0.33 * 0.92. 5 The result is 0.33 * 0.659 = 0.217 percentage points per Mbps.

[0085] The unit bandwidth IO consumption intensity is calculated using the same bandwidth-weighted averaging method. The weighted total IO consumption for domain class 500 is 96,000 MB / s, and the total bandwidth is 120,000 Mbps. Therefore, the unit bandwidth IO consumption is calculated as 96,000 MB / s divided by 120,000 Mbps, resulting in 0.8 MB / s per Mbps. Considering the impact of bandwidth fluctuations, the time-series variation coefficient for this type of bandwidth is 0.2. After introducing a fluctuation correction factor, the calculation yields: .

[0086] The temporal bandwidth vector is constructed by statistically analyzing the peak bandwidth of each domain name class in each time slice, forming a 96-dimensional temporal bandwidth vector. At the same time, it is associated with regional access characteristics. Its core service regions are automatically matched and generated through probe data in actual projects, thus providing an accurate basis for regional adaptation in subsequent scheduling.

[0087] Quantitative analysis shows that the deviation rates of CPU and I / O consumption intensity per unit bandwidth are 2.8% and 3.5%, respectively, which meet the requirements for precise scheduling.

[0088] Step 4: Construct a multi-dimensional resource complementary scheduling model The decision variable is defined as assign[c,n], which represents the amount of bandwidth allocated to physical node n for domain name class c. Here, c takes values ​​from 1 to 1500, representing 1500 domain name classes; and n takes values ​​from 1 to 100, representing 100 physical nodes.

[0089] The resource constraint rule system adds a regional constraint rule to the existing implementation of Example 1, ensuring that domain names are deployed on physical nodes within their core service region, thus reducing cross-regional transmission latency. Specifically, for domain name class c whose core service region is a certain area, it is only allowed to be assigned to physical nodes within that region, avoiding network latency caused by cross-regional deployment.

[0090] Other constraints are the same as in Example 1: the domain name class complete allocation rule ensures that the bandwidth requirements of each domain name class are fully met; the node bandwidth capacity rule limits the total bandwidth of a node to no more than 85% of its capacity; the node CPU capacity rule adopts a dynamically adjusted consumption intensity to constrain the total CPU consumption to no more than 85% of its capacity. In this example, the load stability threshold is set to 0.10.

[0091] Step 5: Model Solving and Generation of Domain-Level Scheduling Scheme The intelligent optimization solution engine is invoked to solve the scheduling model. The solution algorithm uses linear programming, which is an aggregate of a series of algorithms. It uses branch and bound and simplex as the algorithm framework and various heuristic algorithms as auxiliary algorithms. It is suitable for decision variables of 1500 multiplied by 100 equals 150000. The solution task is distributed to multiple computing nodes for parallel processing through task splitting. The solution time is controlled within 3 minutes, which meets the scheduling time requirements of large-scale CDN.

[0092] Taking domain name class 500 as an example, its total bandwidth is 120,000 Mbps, and the core service region is East China. The solution is to allocate 60% of the bandwidth to 20 nodes in East China, 30% of the bandwidth to 15 nodes in North China, and 10% of the bandwidth to 10 nodes in South China, which satisfies the regional constraints and achieves load balancing.

[0093] The domain-level scheduling list is generated in the same way as in Example 1. The bandwidth of each domain name in the class is allocated to the corresponding node according to the node allocation ratio. After being converted into configuration instructions, it is sent to 100 physical nodes for execution. At the same time, the distributed monitoring system is started to collect the load and transmission delay data of each node in real time.

[0094] Step Six: Verification and Comparative Analysis of Scheduling Effects The scheduling effect was verified by comparing the resource utilization and performance indicators of 100 nodes under the old and new schemes. The statistical data is shown in Table 2 below.

[0095] Table 2 Comparison of Physical Node Resource Utilization and Performance Indicators between the New and Old Solutions

[0096] Table 2 presents the significant optimization effects of the new solution on the physical nodes of the large-scale CDN network in three core dimensions: resource balance, peak load, and transmission performance. Regarding resource load balance, CPU utilization variance decreased from 0.138 to 0.072, a reduction of 47.8%, and IO utilization variance decreased from 0.145 to 0.085, a reduction of 41.4%. The significant decrease in variance indicates a substantial reduction in resource consumption differences among nodes, completely changing the unbalanced state of some nodes being overloaded and others idle before scheduling. In terms of peak load control, the average peak load per node decreased from 73Gbps to 62Gbps, a 15% reduction, demonstrating the effectiveness of time-series staggered scheduling. By avoiding peak access times for different domain names, it prevented a surge in node pressure caused by load aggregation. Transmission performance optimization is particularly outstanding, with the average transmission latency decreasing from 45ms to 28ms, a reduction of 37.8%. This is attributed to the synergistic effect of geographical constraints and load balancing, reducing cross-regional transmission losses and node overload latency, significantly improving the user experience. The comprehensive improvement of the four categories of indicators fully demonstrates the scientific nature and practicality of the multi-dimensional resource scheduling model of this invention, and provides strong support for the efficient and stable operation of large-scale CDN networks.

[0097] Step 7: Dynamic Maintenance and Continuous Optimization The periodic update mechanism is executed at 3:00 AM every day, using a distributed data collection and incremental update strategy to reacquire the latest data of 2 million domain names. Only 120,000 domain names with resource characteristic change rates exceeding 10% are reclassified, and the update time is controlled within 1 hour.

[0098] The new domain name rapid access mechanism is optimized for large-scale scenarios. It employs feature vector caching and parallel computing technologies. When a new domain name is accessed, class centroid data is retrieved from the cache, and similarity is calculated in parallel, with the access time controlled within 8 minutes. In this embodiment, approximately 100,000 new domain names are added each month. This mechanism ensures that new domain names are quickly integrated into the scheduling system, guaranteeing service continuity.

[0099] Step 8: Scalability Verification In the node expansion test, adding 50 nodes to bring the total to 150 only required adding node constraint parameters to the model, increasing the solution time to 1.5 minutes, while still meeting the scheduling time requirements. In the domain name expansion test, increasing the number of domain names to 3 million, with 1800 domain name classes after classification, the solution time was 1.8 minutes, demonstrating that the technology still has good scalability in ultra-large-scale scenarios.

[0100] In summary, the two embodiments verified the feasibility of the technology in medium-sized and large-scale CDN network scenarios, respectively. Through core processes such as scientific classification, precise quantification, and multi-dimensional optimization, the technology improves resource utilization, reduces peak load, and optimizes transmission performance. It also has good dynamic adaptability and scalability, providing an efficient solution for CDN node deployment.

[0101] Reference Figure 2 This figure visually demonstrates the advantages of the clustering method of this invention. Traditional customer-based grouping and manual experience-based grouping result in intra-cluster similarities of only 0.42 and 0.47, respectively, indicating heterogeneous domain name resource characteristics within each cluster and making accurate quantification impossible. This invention, through the K-Means++ algorithm, combines CPU, IO, and time-series bandwidth features for clustering, achieving an average intra-cluster similarity of 0.83. This results in highly homogeneous domain name resource consumption within each cluster, providing a reliable foundation for subsequent bandwidth-weighted quantization and solving the core pain point of "intra-cluster heterogeneity" in traditional grouping.

[0102] Reference Figure 3 This figure highlights the scheduling effect of time-series peak shifting and resource complementarity. The old solution experienced severe load fluctuations, peaking at 71Gbps, with the midday and evening peaks overlapping, leading to overload risks. This invention, by matching domain names with different time-series characteristics for peak-shifting deployment and complementing CPU / IO-intensive domain name classes, reduces the peak load to 62Gbps, significantly reducing fluctuations. The 24-hour load is more stable, avoiding load overlap in a single period, improving node operational stability, and validating the effectiveness of the multi-dimensional scheduling model.

[0103] Reference Figure 4 This diagram illustrates the core value of multi-dimensional resource complementarity. The old solution suffers from "one-dimensional overload and one-dimensional idleness" problems; for example, node 1 has 78% CPU utilization but only 35% IO, while node 2 has 80% IO utilization but only 32% CPU utilization. This invention constructs a multi-dimensional constraint model to deploy heterogeneous domain names in a complementary manner. After scheduling, the CPU and IO utilization of all nodes are concentrated in a balanced range of 55%-65%, avoiding resource waste and overload risks. It achieves coordinated optimization of CPU, IO, and bandwidth, solving the imbalance problem of traditional one-dimensional scheduling.

[0104] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A CDN node deployment method for achieving multi-dimensional resource quantification and complementary scheduling, characterized in that, Includes the following steps: Step S1: Domain Name Operation Data Collection and Feature Vector Construction. First, the 24-hour period is divided into several time slices according to a preset time granularity. The time granularity ranges from 5 to 30 minutes to adapt to different business scenarios. Based on the divided time slices, real-time and historical bandwidth, CPU consumption, and IO consumption data of each domain name are collected from multiple source databases. Then, the CPU consumption intensity and IO consumption intensity corresponding to the unit bandwidth are calculated. The bandwidth data of each time slice are integrated to construct a high-dimensional bandwidth time series feature vector, which fully preserves the bandwidth fluctuation pattern and peak characteristics. Step S2: Domain-level multidimensional feature normalization. Given the differences in dimensions and numerical magnitudes among unit bandwidth CPU consumption intensity, unit bandwidth IO consumption intensity, and bandwidth time-series feature vectors, Min-Max or Z-Score normalization methods are used for domain-level unification. This domain-level normalization eliminates the magnitude differences between different domains, transforming multidimensional resource indicators into dimensionless standardized feature vectors while preserving the relative differences between features. Step S3: Domain name clustering and classification based on resource consumption characteristics. The clustering method is used to automatically group the standardized feature vectors. The silhouette coefficient is introduced into the clustering process to evaluate the classification quality. Domain names with high similarity in CPU consumption intensity, IO consumption intensity and bandwidth time series characteristics per unit bandwidth are aggregated into several domain name classes. The feature centroid of each domain name class is calculated. The centroid reflects the average resource consumption intensity and aggregated bandwidth time series characteristics of the domain names in the class. Step S4: Quantify the multi-dimensional resource demand with bandwidth weighting, and calculate the cumulative bandwidth of all domains within each domain class as the total bandwidth demand for that class; Using the proportion of each domain's bandwidth to the total bandwidth within the class as a weight, the CPU consumption intensity and I / O consumption intensity per unit bandwidth of the domains within the class are weighted and averaged to obtain the comprehensive resource consumption intensity of the domain class. Combined with the bandwidth data of each domain in the class at the corresponding time slice, the aggregated bandwidth time-series vector, CPU, and I / O three-dimensional information of the domain class are formed to complete the quantification of resource requirements. Step S5: Construct a multi-dimensional resource complementary scheduling model, define binary allocation decision variables from domain name class to physical node, establish a constraint rule system, and the constraints include: complete allocation constraint of domain name class, bandwidth capacity constraint of physical node, CPU capability constraint, IO performance constraint and load stability constraint, and construct a system to suppress peak bandwidth load of nodes. Step S6: Model Solving and Scheduling Scheme Implementation. The intelligent resource scheduling engine uses linear programming to solve the scheduling model and obtain an approximate optimal allocation scheme from domain name class to physical node. Based on the bandwidth weight of the domain name within its domain name class, the class-level allocation scheme is decomposed into a domain-level scheduling list. The list is then sent to the physical nodes through the scheduling interface, and load monitoring is started simultaneously.

2. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, It also includes a time-series bandwidth feature dimensionality reduction mechanism. The preset time granularity in step S1 is a configuration parameter with a value range of 5-30 minutes, corresponding to a number of time slices of 48-288. The preferred configuration is 15 minutes and divided into 96 time slices. The time-series bandwidth feature is represented as a 96-dimensional vector, and the time-series bandwidth vector can be dimensionality reduced by using discrete cosine transform or fast Fourier transform.

3. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, Step S1 also includes data cleaning and abnormal domain name identification and optimization steps. The original time series data is subjected to outlier detection using the 3σ principle, and data points exceeding the mean ± 3σ range are removed. Data records with missing rates exceeding a preset threshold are filtered out. Missing data are filled using linear interpolation or the historical 7-day average. Trend consistency verification is introduced during the abnormal domain name identification process, and the consistency between the domain name bandwidth time series and domain names of the same business type is judged by the Pearson correlation coefficient.

4. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 3, characterized in that, In step S1, when the CPU consumption or IO consumption per unit bandwidth of a domain exceeds three times the standard deviation of a preset threshold, it is marked as a domain with abnormal resource consumption. When the temporal bandwidth vector fluctuation coefficient of a domain exceeds the threshold and exhibits non-periodic random fluctuation, it is marked as a domain with abnormal bandwidth mode. Abnormal domains are isolated or a special domain class is created separately.

5. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, The node CPU capability rule in step S5 is specifically expressed as follows: Where c represents the domain name index. Allocate the amount of bandwidth to node n for domain name class c. The CPU consumption intensity per unit bandwidth for the corrected domain class c. Where is the CPU capacity of node n, and 0.85 is the safety threshold coefficient. Using the time decay formula The correction was obtained. The historical unit bandwidth CPU consumption intensity is represented by α, which is the time decay coefficient with a value range of 0.8-0.95, and t is the number of days since the current time. This correction reflects the time evolution pattern of the domain name's CPU consumption characteristics.

6. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, The node I / O performance rules in step S5 are specifically expressed as follows: Where c is a domain name index, Allocate the amount of bandwidth to node n for domain name class c. The domain name class c quantized in step S4 has a unit bandwidth IO consumption intensity. 0.80 is the IO capacity of node n, and 0.80 is the safety threshold coefficient, reserving 20% ​​resource redundancy to cope with sudden IO demand.

7. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, The node timing load rule in step S5 is specifically expressed as follows: Where c is the domain name index, t is the time slice index, and T is the total number of time slices divided into 24 hours. The amount of bandwidth allocated to node n for domain class c in time slice t. Let n be the bandwidth capacity of node n, and 0.85 be the safety threshold coefficient. At the same time, a node load stability constraint is established, which is achieved by calculating the variance of the node's load in each time slice over 24 hours. The constraint is set within the threshold range of variance ≤ 0.05-0.15 to suppress large fluctuations in node load.

8. The CDN node deployment method for realizing multi-dimensional resource quantization and complementary scheduling according to claim 1, characterized in that, It also includes a periodic update mechanism for domain name centroids, which re-acquires the latest operational data of all domain names on the entire network at a preset cycle every morning, and repeats steps S1 to S3 to update the feature vectors, centroids, and resource requirement quantification parameters of the domain name classes; it also has a fast access mechanism for new domain names. When a new domain name is accessed, a standardized feature vector is constructed according to steps S1 and S2, and the Euclidean distance or cosine similarity between the vector and the existing K domain name class centroids is calculated. The new domain name is assigned to the domain name class with the highest similarity. When the similarity is lower than a preset threshold, a new domain name class is created without performing full clustering.

9. A CDN node scheduling system, used to implement the CDN node deployment method for achieving multi-dimensional resource quantization and complementary scheduling as described in any one of claims 1-8, characterized in that, include: The data acquisition and preprocessing module is used to collect historical domain name operation data from the database, sample and perform cleaning and verification according to preset time granularity, support multi-source database access and realize unified data format conversion; The feature engineering module is used to extract the CPU consumption intensity, IO consumption intensity and timing bandwidth features of domain names, and generates standardized feature vectors by using Min-Max normalization or Z-Score normalization, and integrates feature dimensionality reduction function. The domain name scientific classification module is used to cluster standardized feature vectors based on the K-Means++ algorithm, generate several domain name classes and class centroids, and support rapid classification and creation of new classes based on the similarity between new domain name feature vectors and existing class centroids. The resource demand quantification module is used to quantify the unit bandwidth CPU consumption intensity and unit bandwidth IO consumption intensity of each domain name class using a bandwidth-weighted average method, and introduces a time decay factor to correct the resource demand parameters. The multi-dimensional resource scheduling decision module is used to construct the scheduling model of the constraint rules, solve for the allocation scheme from domain name class to physical node, and support dynamic optimization of model parameters; The deployment, execution, and monitoring module is used to expand the domain name allocation strategy into a domain name-level scheduling list for execution, and to monitor node load and peak-shaving effects. The time granularity, number of target categories, clustering algorithm type, and normalization method are all configurable parameters that can be flexibly adjusted according to business needs.

Images