Network resource optimization method based on multi-dimensional computing power scheduling

Abstract

This invention relates to the field of computing power network technology, specifically disclosing a network resource optimization method based on multi-dimensional computing power scheduling. This optimization method includes the following steps: S1: Establishing a multi-dimensional computing power detection terminal. This terminal covers all network computing power nodes, business domains, and terminal devices, collecting real-time computing power demand data, task data, and network resource occupancy data for each business domain across the entire domain, forming a comprehensive computing power demand dataset. This invention achieves real-time data collection of comprehensive computing power resources through a multi-dimensional computing power detection terminal. Combined with a computing power demand analysis model, it accurately calculates the real-time computing power demand for each business domain, avoiding resource allocation deviations in the traditional static allocation model. Furthermore, through a multi-level optimization mechanism of initial overall allocation, periodic review and optimization, and local dynamic fine-tuning, it effectively solves the problem of uneven utilization of computing power node resources.

Description

Technical Field

[0001] This invention relates to the field of computing power network technology, and in particular to a method for optimizing network resources based on multi-dimensional computing power scheduling. Background Technology

[0002] Computing power is the ability of a device to process data and complete computational tasks, which is what we often call "computing ability". You can think of it as the "brainpower" or "muscle" of the digital world. Like water and electricity, it is a core basic resource in the digital economy era.

[0003] With the rapid development of the digital economy, the demand for network computing resources in business scenarios such as 5G, industrial internet, and artificial intelligence has exploded, and computing networks are gradually becoming the core infrastructure supporting the operation of various businesses.

[0004] Currently, traditional computing power scheduling models mostly adopt static allocation or single-dimensional resource scheduling methods, allocating computing power only based on basic hardware indicators such as CPU and memory of computing nodes. This fails to fully consider the task characteristics, latency requirements, priority differences, and dynamic changes in global computing resources within the business domain, leading to a severe disconnect between computing resource allocation and actual business needs. Specifically, this manifests as insufficient computing power, excessive task latency, and even business interruptions during peak computing periods for some core businesses, while the resource utilization of some idle computing nodes remains at a consistently low level, resulting in significant waste of computing resources. Therefore, to address these issues, a network resource optimization method based on multi-dimensional computing power scheduling is proposed. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a network resource optimization method based on multi-dimensional computing power scheduling.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A network resource optimization method based on multi-dimensional computing power scheduling includes the following steps:

[0008] S1: Build a multi-dimensional computing power detection terminal. The multi-dimensional computing power detection terminal covers the entire network computing power nodes, business areas and terminal devices. It collects computing power demand data, task data and network resource usage data of each business area in real time to form a full-domain computing power demand dataset.

[0009] S2: Build a multi-dimensional computing power terminal, establish a computing power demand analysis model, and transmit the full-domain computing power demand dataset collected by the multi-dimensional computing power detection terminal to the multi-dimensional computing power terminal in real time. Through the computing power demand analysis model, comprehensively analyze the task volume, computational complexity, and latency requirements of each business area, and accurately calculate the real-time computing power value required by each business area in the entire domain.

[0010] S3: Computing power allocation. After the multi-dimensional computing power terminal completes the computing power calculation required by each field, it combines the idle computing power resources of the whole domain and the upper limit of node computing power capacity, and completes the initial overall allocation of computing power resources of the whole domain according to the basic computing power needs of each field, so as to realize the initial matching and implementation of computing power resources.

[0011] S4: Task classification, constructing a multi-dimensional task classification evaluation system, using task importance, business priority, task latency requirements, and task urgency as evaluation indicators, classifying operational tasks in all business areas across the entire domain into three levels of task priority: high, medium, and low.

[0012] S5: Computing power shortage optimization. It compares the total computing power demand of all tasks with the total available computing power of the network in real time. When a computing power shortage scenario occurs where the total computing power required by a task exceeds the total available computing power of the network, it prioritizes allocating sufficient computing power resources to high-level important tasks according to the preset task level priority, and performs computing power compression and delay scheduling on medium and low-level tasks to ensure the stable operation of core businesses.

[0013] S6: Global computing power review and optimization. After completing the allocation of computing power and task scheduling in various fields, it collects real-time actual operation data such as task running speed, computing power utilization, and network latency of each computing power node and each business field to form computing power operation feedback data. Based on the feedback data, it performs periodic review and optimization of the global computing power allocation scheme to correct computing power allocation deviations.

[0014] S7: Computing power fine-tuning. Based on real-time feedback data of global computing power operation, it performs dynamic fine-tuning of computing power for local abnormal scenarios such as excess or insufficient computing power in a single domain or single node, and balances the resource load of each computing power node in real time to achieve refined and dynamic optimal configuration of global computing power resources.

[0015] Preferably, in S1, the dimensions of the computing power demand data collected by the multi-dimensional computing power detection terminal include: CPU utilization, memory usage, GPU utilization, network bandwidth usage, disk I / O rate of the computing power node, online status of the terminal device, computing power limit, current load rate, as well as task type, number of concurrent requests, and data transmission rate in the business domain.

[0016] Preferably, in S2, the computing power demand analysis model is a multi-feature fusion model based on deep learning. The model input is the global computing power demand dataset, and the output is the real-time computing power demand value of each business domain. The training samples of the model are historical computing power demand data and corresponding business operation status data. The training objective is to make the computing power demand prediction error rate less than 3%.

[0017] Preferably, in S3, the initial overall allocation of computing power includes: based on the basic computing power requirements of each business area, prioritizing the allocation of idle computing power resources to core business nodes, while reserving no less than 15% of the node computing power as an emergency computing power pool to cope with sudden high computing power demand scenarios.

[0018] Preferably, in S4, the evaluation index weights of the multi-dimensional graded evaluation system for tasks are configured as follows: task importance weight 40%, business priority weight 30%, task latency requirement weight 20%, and task urgency weight 10%. The comprehensive score of the task is calculated by weighted summation. A score ≥80 is considered a high-level task, 50-79 is considered a medium-level task, and <50 is considered a low-level task.

[0019] Preferably, in S5, the optimization methods for handling low- and medium-level tasks with scarce computing power include: using computing power downgrading for medium-level tasks to reduce computational complexity to 70%-80% of the original requirement; and using delayed scheduling for low-level tasks to postpone task execution to idle computing power periods, with the delay duration not exceeding the maximum latency threshold allowed by the business.

[0020] Preferably, in S6, the cycle for global computing power review and optimization is once per hour. The review content includes the computing power utilization deviation rate of each computing power node, the latency compliance rate of business tasks, and the network resource utilization rate. When the deviation rate exceeds 10%, the automatic correction of the computing power allocation scheme is triggered.

[0021] Preferably, in S7, the triggering condition for dynamic fine-tuning of computing power is: the computing power utilization rate of a single node exceeds 90% or is lower than 30%, or the latency compliance rate of a single domain business task is lower than 95%. The fine-tuning method is to schedule the computing power resources of the surplus nodes to the insufficient nodes through the computing power migration algorithm. The amount of computing power resources for each fine-tuning does not exceed 10% of the total computing power of the node.

[0022] Preferably, it also includes a computing power resource visualization module, which is used to display the resource status of computing power nodes across the entire domain, the computing power demand and allocation of various business areas, task running progress and computing power shortage warning information in real time. The visualization interface supports data filtering and querying by business area, computing power node and time dimension.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. This invention achieves real-time data collection of computing power resources across the entire domain through a multi-dimensional computing power detection terminal. Combined with a computing power demand analysis model, it accurately calculates the real-time computing power demand of each business area, avoiding resource allocation deviations in the traditional static allocation mode. Furthermore, through a multi-level optimization mechanism of initial overall allocation, periodic review and optimization, and local dynamic fine-tuning, it effectively solves the problem of uneven utilization of computing power node resources, improves the utilization rate of computing power resources across the entire domain, increases the activation rate of idle computing power resources, reduces the ineffective waste of computing power resources, and achieves the maximum utilization of computing power resources.

[0025] 2. This invention constructs a multi-dimensional hierarchical evaluation system for tasks. By classifying tasks into different levels, it achieves priority allocation of computing resources. In scenarios where computing power is scarce, it prioritizes the computing power needs of high-level core businesses. At the same time, it adopts computing power compression, delay scheduling and other processing methods for medium and low-level tasks to avoid problems such as excessive latency and operation interruption of core businesses due to insufficient computing power.

[0026] 3. In this invention, the global computing power review and optimization module realizes the periodic automatic correction of the computing power allocation scheme. Combined with the local computing power dynamic fine-tuning mechanism, it can quickly respond to dynamic scenarios such as changes in computing power node load and fluctuations in business demand. It solves the problem that the traditional scheduling mode cannot adapt to changes in computing power resources in real time. The response time of computing power scheduling is shortened, which can effectively cope with sudden high computing power demand scenarios and improve the overall resilience of the computing power network. Attached Figure Description

[0027] Figure 1 This is a flowchart of the network resource optimization method based on multi-dimensional computing power scheduling proposed in this invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0029] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", 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.

[0030] Reference Figure 1This embodiment uses a city-level 5G computing network platform as an application scenario. This platform covers 50 computing nodes across 12 districts and counties, carrying over 200 operational tasks across four major business areas: communications, government affairs, industrial internet, and smart cities. The computing power requirements of each business area vary significantly, and computing resource shortages frequently occur during peak periods. Under the traditional static computing power scheduling model, the computing power utilization rate is only 65%, and the core business latency compliance rate is only 82%, indicating significant resource waste and business stability issues. This embodiment employs the network resource optimization method based on multi-dimensional computing power scheduling of the present invention. The specific implementation process is as follows:

[0031] S1: Establish a multi-dimensional computing power detection terminal. In this embodiment, the multi-dimensional computing power detection terminal adopts a distributed deployment architecture. A computing power acquisition module is deployed on each computing power node, a business data acquisition module is deployed in each business domain, and a device status acquisition module is deployed on the terminal device side. The computing power node acquisition module collects CPU utilization, memory usage, GPU utilization, network bandwidth usage, and disk I / O rate data in real time, at a frequency of 1 time / second. The business data acquisition module collects task type, concurrency, data transmission rate, and latency requirements data for each business domain, at a frequency of 5 times / second. The terminal device status acquisition module collects device online status, computing power limit, and current load rate data, at a frequency of 1 time / second. All collected data is transmitted to the data center in real time via the MQTT protocol, forming a full-domain computing power demand dataset. Data storage uses the time-series database InfluxDB, with a data retention period of 30 days, facilitating subsequent computing power demand analysis and post-mortem optimization. In this embodiment, the platform deploys a total of 50 computing power node acquisition modules, 12 business domain acquisition modules, and more than 2,000 terminal device acquisition modules, achieving full coverage acquisition of computing power resources and business data across the entire domain, with no data acquisition blind spots.

[0032] S2: Build a multi-dimensional computing power terminal and establish a computing power demand analysis model. The multi-dimensional computing power terminal is deployed in the platform's core data center. The computing power demand analysis model adopts a multi-feature fusion deep learning model based on LSTM. The model input includes computing power node status data, business task data, and terminal device data from the global computing power demand dataset. The model's hidden layer contains three layers of LSTM units, with 128 units per layer. The output layer is the real-time computing power demand value for each business domain. Model training uses historical computing power data from the platform over the past six months. After training, the model is deployed on the TensorFlow Serving framework of the computing power terminal to achieve real-time analysis and calculation of computing power demand. The calculation time for a single batch of computing power demand does not exceed 100ms. In this embodiment, the computing power demand analysis model stably controls the prediction error rate of computing power demand for the four major business domains within 2.5%, accurately calculating the real-time required computing power value for each business domain, providing a reliable basis for subsequent computing power allocation.

[0033] S3: Computing power allocation, completing the initial overall allocation of computing power resources across the entire domain. After the multi-dimensional computing power terminal completes the computing power calculation for each business area, it combines the idle computing power resources of 50 computing power nodes across the entire domain and the node computing power capacity limit to conduct the initial overall allocation according to the basic computing power requirements of each area. In this embodiment, the communication business area is the core business, with basic computing power requirements accounting for 40% of the platform's total computing power; the government affairs business area accounts for 25%; the industrial internet business area accounts for 20%; and the smart city business area accounts for 15%. During the computing power allocation process, idle computing power resources are prioritized for allocation to core business nodes such as communication and government affairs. At the same time, 15% of the computing power resources are reserved for each computing power node as an emergency computing power pool to cope with sudden high computing power demand scenarios. For example, if the total computing power of a certain district's computing power nodes is 100 TOPS, the computing power allocated to the communication business is 40 TOPS, the government affairs business is 25 TOPS, the industrial internet business is 20 TOPS, the smart city business is 15 TOPS, and 10 TOPS is reserved as an emergency computing power pool. After the initial computing power allocation is completed, the computing power resource allocation plan is distributed to each computing power node through the platform's computing power management system, realizing the initial matching and implementation of computing power resources. In this embodiment, after the initial computing power allocation is completed, the average utilization rate of computing power resources across the entire domain increases to 75%, which is 10 percentage points higher than the traditional model.

[0034] S4: Task classification, constructing a multi-dimensional task classification evaluation system. In this embodiment, the evaluation indicators and weights of the multi-dimensional task classification evaluation system are configured as follows: task importance weight 40%, business priority weight 30%, task latency requirement weight 20%, and task urgency weight 10%. Task importance is categorized based on the core nature of the business domain: voice calls and data transmission in telecommunications are worth 100 points; administrative approvals and data queries in government affairs are worth 90 points; equipment control and data acquisition in industrial internet are worth 85 points; and video surveillance and environmental monitoring in smart cities are worth 70 points. Business priority is based on the platform's preset business levels: Level 1 priority is worth 100 points, Level 2 is worth 80 points, and Level 3 is worth 60 points. Task latency requirements are based on the maximum permissible latency: tasks with latency ≤100ms are worth 100 points, tasks with latency between 100ms and 500ms are worth 80 points, and tasks with latency >500ms are worth 60 points. Task urgency is based on the execution time window: tasks requiring immediate execution are worth 100 points, tasks requiring execution within one hour are worth 80 points, and tasks requiring execution within 24 hours are worth 60 points. A weighted summation of the overall task score is used to calculate the overall score: tasks with a score ≥80 are high-level tasks, 50-79 are medium-level tasks, and <50 are low-level tasks. In this embodiment, the platform is divided into 80 high-level tasks, 70 medium-level tasks, and 50 low-level tasks, forming a three-tiered task priority hierarchy, which provides a basis for subsequent optimization in the face of computing power shortages.

[0035] S5: Optimization for computing power shortages, addressing scenarios with insufficient computing resources. In this embodiment, during peak platform periods, the total computing power requirement for all tasks is 4500 TOPS, while the total available computing power across the platform is only 4000 TOPS, resulting in a computing power shortage scenario. In this case, based on task priority, sufficient computing power resources are allocated to high-priority tasks first, while computing power compression and delayed scheduling are implemented for low- and medium-priority tasks. Specifically, high-level communication voice calls and government data query tasks are given priority in acquiring computing resources, with an allocation of 2000 TOPS. Mid-level industrial internet data collection tasks undergo a downgraded computing power approach, reducing computational complexity to 75% of the original requirement, decreasing the required computing power from 1200 TOPS to 900 TOPS. Low-level smart city video surveillance tasks are handled with delayed scheduling, postponing non-real-time video analysis tasks to idle computing power at night, reducing the required computing power from 800 TOPS to 300 TOPS. Through these measures, the total computing power requirement is controlled within 3200 TOPS, lower than the platform's available total computing power of 4000 TOPS, effectively solving the computing power shortage problem. In this embodiment, after the computing power shortage optimization, the latency compliance rate for high-level core services reaches 99.5%, while the operational status of mid- and low-level tasks is not significantly affected, ensuring the stable operation of core services.

[0036] S6: Global Computing Power Review and Optimization to Correct Computing Power Allocation Deviations. In this embodiment, the global computing power review and optimization cycle is once per hour, and the review process is automatically executed by the platform's computing power management system. First, real-time data such as task execution speed, computing power utilization, and network latency of each computing power node are collected to form computing power operation feedback data. Then, statistical analysis is performed on the computing power utilization deviation rate, business task latency compliance rate, and network resource occupancy rate of each computing power node. For example, if the review finds that the computing power utilization rate of a certain computing power node is 95%, exceeding the preset reasonable range of 80%-90%, with a deviation rate exceeding 10%, and the latency compliance rate of the industrial internet tasks carried by this node is only 90%, lower than the preset 95%, then the automatic correction of the computing power allocation scheme is triggered, migrating some of the smart city business computing power resources of this node to nearby idle computing power nodes. After the correction, the computing power utilization rate of this node drops to 85%, and the task latency compliance rate increases to 98%. Simultaneously, the review process also analyzed the prediction errors of the computing power demand analysis model. For business domain data with high error rates, incremental training was performed on the model to optimize model parameters and improve the accuracy of subsequent computing power demand predictions. In this embodiment, through periodic review and optimization, the overall computing power allocation deviation rate was controlled within 5%, significantly improving the accuracy of computing power resource allocation.

[0037] S7: Computing power fine-tuning to achieve dynamic balancing of local computing power resources. In this embodiment, the triggering condition for dynamic computing power fine-tuning is: the computing power utilization rate of a single node exceeds 90% or is lower than 30%, or the latency compliance rate of a single domain business task is lower than 95%. For example, if the computing power utilization rate of an edge computing node is 98%, and the latency compliance rate of the communication data transmission task it carries is only 92%, the computing power fine-tuning mechanism is triggered. Through the computing power migration algorithm, idle computing power resources of neighboring computing power nodes are scheduled to this node. The amount of computing power resources adjusted each time is 8% of the total computing power of the node. After the fine-tuning, the computing power utilization rate of this node drops to 88%, and the task latency compliance rate increases to 99%. At the same time, for scenarios where the computing power utilization rate of some computing power nodes is lower than 30%, their idle computing power resources are scheduled to nodes with tight computing power, achieving resource load balancing of computing power nodes across the entire domain. In this embodiment, the response time of computing power fine-tuning is no more than 5 seconds, which can quickly correct local computing power resource configuration deviations and achieve refined and dynamic optimal configuration of computing power resources across the entire domain.

[0038] The above description is only a preferred embodiment 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 network resource optimization method based on multi-dimensional computing power scheduling, characterized in that, The optimization method includes the following steps: S1: Build a multi-dimensional computing power detection terminal. The multi-dimensional computing power detection terminal covers the entire network computing power nodes, business areas and terminal devices. It collects computing power demand data, task data and network resource usage data of each business area in real time to form a full-domain computing power demand dataset. S2: Build a multi-dimensional computing power terminal, establish a computing power demand analysis model, and transmit the full-domain computing power demand dataset collected by the multi-dimensional computing power detection terminal to the multi-dimensional computing power terminal in real time. Through the computing power demand analysis model, comprehensively analyze the task volume, computational complexity, and latency requirements of each business area, and accurately calculate the real-time computing power value required by each business area in the entire domain. S3: Computing power allocation. After the multi-dimensional computing power terminal completes the computing power calculation required by each field, it combines the idle computing power resources of the whole domain and the upper limit of node computing power capacity, and completes the initial overall allocation of computing power resources of the whole domain according to the basic computing power needs of each field, so as to realize the initial matching and implementation of computing power resources. S4: Task classification, constructing a multi-dimensional task classification evaluation system, using task importance, business priority, task latency requirements, and task urgency as evaluation indicators, classifying operational tasks in all business areas across the entire domain into three levels of task priority: high, medium, and low. S5: Computing power shortage optimization. It compares the total computing power demand of all tasks with the total available computing power of the network in real time. When a computing power shortage scenario occurs where the total computing power required by a task exceeds the total available computing power of the network, it prioritizes allocating sufficient computing power resources to high-level important tasks according to the preset task level priority, and performs computing power compression and delay scheduling on medium and low-level tasks to ensure the stable operation of core businesses. S6: Global computing power review and optimization. After completing the allocation of computing power and task scheduling in various fields, it collects real-time actual operation data such as task running speed, computing power utilization, and network latency of each computing power node and each business field to form computing power operation feedback data. Based on the feedback data, it performs periodic review and optimization of the global computing power allocation scheme to correct computing power allocation deviations. S7: Computing power fine-tuning. Based on real-time feedback data of global computing power operation, it performs dynamic fine-tuning of computing power for local abnormal scenarios such as excess or insufficient computing power in a single domain or single node, and balances the resource load of each computing power node in real time to achieve refined and dynamic optimal configuration of global computing power resources.

2. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S1, the computing power demand data collection dimensions of the multi-dimensional computing power detection terminal include: CPU utilization, memory usage, GPU utilization, network bandwidth usage, disk I / O rate of computing power nodes, online status of terminal devices, computing power limit, current load rate, as well as task type, concurrency, and data transmission rate in the business domain.

3. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S2, the computing power demand analysis model is a multi-feature fusion model based on deep learning. The model input is the global computing power demand dataset, and the output is the real-time computing power demand value of each business domain. The training samples of the model are historical computing power demand data and corresponding business operation status data. The training objective is to make the computing power demand prediction error rate less than 3%.

4. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S3, the initial overall allocation of computing power includes: based on the basic computing power needs of each business area, prioritizing the allocation of idle computing power resources to core business nodes, while reserving no less than 15% of the node computing power as an emergency computing power pool to cope with sudden high computing power demand scenarios.

5. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S4, the evaluation index weights of the multi-dimensional graded evaluation system for tasks are configured as follows: task importance weight 40%, business priority weight 30%, task latency requirement weight 20%, and task urgency weight 10%. The comprehensive score of the task is calculated by weighted summation. A score ≥80 is a high-level task, 50-79 is a medium-level task, and <50 is a low-level task.

6. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S5, the optimization methods for handling low- and medium-level tasks with scarce computing power include: for medium-level tasks, computing power is downgraded to reduce the computational complexity to 70%-80% of the original requirement; for low-level tasks, delayed scheduling is used to postpone the task execution time to the idle period of computing power, and the delay time does not exceed the maximum latency threshold allowed by the business.

7. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S6, the global computing power review and optimization cycle is once per hour. The review includes the computing power utilization deviation rate of each computing power node, the latency compliance rate of business tasks, and the network resource utilization rate. When the deviation rate exceeds 10%, the automatic correction of the computing power allocation scheme is triggered.

8. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, In S7, the triggering conditions for dynamic fine-tuning of computing power are: the computing power utilization rate of a single node exceeds 90% or is lower than 30%, or the latency compliance rate of a single domain business task is lower than 95%. The fine-tuning method is to schedule the computing power resources of the surplus nodes to the insufficient nodes through the computing power migration algorithm. The amount of computing power resources for each fine-tuning does not exceed 10% of the total computing power of the node.

9. The network resource optimization method based on multi-dimensional computing power scheduling according to claim 1, characterized in that, It also includes a computing resource visualization module, which is used to display the resource status of computing nodes across the entire domain, the computing power demand and allocation of various business areas, task running progress and computing power shortage warning information in real time. The visualization interface supports data filtering and querying by business area, computing node, and time dimension.

Images