A computing power network multi-factor fusion scheduling method and device

Abstract

The present invention discloses a multi-element fusion scheduling method and device for a computing power network, which relates to the field of computing power network resource scheduling. The method comprises the following steps: constructing a user objective function, converting a quantitative index into a target value U obtainable by computing power nodes, setting a coarse screening interval according to the target value U, constructing a system objective function, converting a system performance index into a target value S obtainable by computing power nodes, constructing a multi-objective fusion scheduling function based on NSGA II according to the user objective function and the system objective function, performing multi-objective coarse screening of computing power nodes using the multi-objective fusion scheduling function, obtaining all targets within the user coarse screening interval, performing personalized sorting based on all targets within the obtained user demand interval in combination with the user objectives and the system performance indicators, and selecting the optimal scheduling scheme according to the personalized sorting for computing power scheduling.

Description

Technical Field

[0001] The present invention discloses a method and a device, which relate to the field of computing power network resource scheduling, and specifically a computing power network multi-element fusion scheduling method and device. Background Art

[0002] The computing network unifies the underlying ubiquitous computing capabilities through a central system with orchestration and scheduling capabilities. In terms of user service capabilities, it needs to provide services to a variety of users such as universities, scientific research, industry, and individuals. Different users have personalized needs in terms of cost, service quality, latency, etc. In addition, it needs to meet user SLA requirements and performance indicators such as energy consumption and network load to achieve the goal of a green and low-carbon computing network. Current resource-related scheduling technologies are mainly based on local resource pools such as cloud data centers. The resource pools and user needs they face are relatively simple, making them difficult to apply to the massive computing nodes and complex user needs of the computing network. In addition, the current solution mainly considers user needs, without further considering the computing network's own performance indicators such as energy consumption and scheduling distance, as well as the user interaction experience capabilities of the scheduling, which makes it easy to encounter computing network scheduling bottlenecks. Summary of the Invention

[0003] In response to the problems of the prior art, the present invention provides a computing power network multi-factor fusion scheduling method and device, which accurately schedules computing power network resources through multi-objective coarse screening and personalized fine sorting strategies.

[0004] The specific scheme proposed by the present invention is:

[0005] The present invention provides a multi-factor fusion scheduling method for computing power network, constructs a user objective function, converts quantitative indicators into a target value U that can be obtained by computing power nodes, and sets a coarse screening interval according to the target value U.

[0006] Construct the system objective function and convert the system performance indicators into the target value S that can be obtained by computing power nodes.

[0007] Based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function.

[0008] Use the multi-objective fusion scheduling function to perform multi-objective coarse screening of computing nodes to obtain all the objectives within the user's coarse screening interval.

[0009] Combine user goals and system performance indicators to perform personalized sorting of all goals within the obtained user demand range, and select the optimal scheduling plan based on the personalized sorting to schedule computing power.

[0010] Furthermore, the construction of the user objective function in the computing power network multi-factor fusion scheduling method further includes:

[0011] According to the user SLA requirement indicators, convert the judgment type indicators into the node pre-screening rule F, and the judgment type indicators include computing power type, access network capacity, software environment, and computing power storage ratio.

[0012] Furthermore, in the multi-factor fusion scheduling method of a computing power network, perform computing power node screening according to the pre-screening rule F to form a computing power node pool for multi-objective rough screening. [[ID=*]]

[0013] Furthermore, in the multi-factor fusion scheduling method of a computing power network, perform multi-objective rough screening of computing power nodes by using the multi-objective fusion scheduling function, including: performing non-dominated sorting and solving based on the NSGA II algorithm according to U and S:

[0014] Step S1: Randomly generate initial solutions, and the number of initial solutions is greater than the number of set target values;

[0015] Step S2: Perform the first non-dominated sorting according to U. The non-dominated rank of non-dominated solutions is 1. After removing the non-dominated solutions with rank 1, a group with rank 2 is formed, and so on. Within the same first non-dominated sorting rank, perform the second non-dominated sorting according to S, and the sorting method is the same as that of U to form a Pareto set;

[0016] Step S3: For all the solutions obtained in S2, select them into the next-generation population in ascending order of non-dominated rank until the number of solution individuals in the first next-generation population is greater than the number of target values. Then, select the optimal solutions according to the crowding degree in the non-dominated rank of the first next-generation population until the number of solutions in the population is equal to the number of target values;

[0017] Step S4: Perform crossover and mutation operations based on the computing power node relationship table on the basis of the current solutions to generate new candidate solutions;

[0018] Step S5: Truncate the current solutions and retain the solution set P2 with the same number of solutions as the initial solutions,

[0019] Step S6: Repeat steps S2 - S4.

[0020] Furthermore, in the multi-factor fusion scheduling method of a computing power network, when performing mutation operations in step S4, preferentially perform mutation operations on the computing power nodes that commonly appear in the historical scheduling schemes:

[0021] Step S41: Generate a random integer r1, where 0 < r1 < the number of required computing power nodes;

[0022] Step S42: Randomly select a computing power node from the current solutions and obtain the computing power node relationship table R of the computing power node i(i=r1), the probability of the computing power node A and other computing power nodes related to A appearing at the same time is displayed through the computing power node relationship table. When A appears, other computing power nodes related to A are preferentially selected according to A's computing power node relationship table;

[0023] Step S43: Obtain a computing power node from other computing power nodes related to A through the roulette algorithm and perform a mutation operation.

[0024] Furthermore, in the described computing power network multi-factor fusion scheduling method, the solution of the first non-dominated sort in step S2 is binarized, including: for the target value U, if the current solution target value is greater than 0.95*U and less than 1.05*U, it is recorded as 1, otherwise it is 0, and the non-dominated sort is determined by using the solution after binarization processing.

[0025] Furthermore, in the described computing power network multi-factor fusion scheduling method, the solution set P2 is first sorted and screened according to the dimension of system performance indicators, and then personalized sorted according to the dimension of user goals to obtain the solution set P3, and the optimal scheduling plan is selected based on the solution set P3.

[0026] The present invention also provides a computing power network multi-factor fusion scheduling device, including a construction module, a coarse screening module, a fine sorting module and a selection module.

[0027] The construction module constructs the user objective function, converts the quantitative indicators into the target value U that can be obtained by computing power nodes, and sets the coarse screening interval according to the target value U; constructs the system objective function, converts the system performance indicators into the target value S that can be obtained by computing power nodes; based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function.

[0028] The coarse screening module uses the multi-objective fusion scheduling function to perform multi-objective coarse screening of computing nodes to obtain all targets within the user's coarse screening interval.

[0029] The fine sorting module combines user goals and system performance indicators to perform personalized sorting based on all goals within the obtained user demand range. The selection module selects the optimal scheduling plan for computing power scheduling based on the personalized sorting.

[0030] The benefits of the present invention are:

[0031] This paper proposes a multi-factor fusion scheduling method for computing power networks. Based on the NSGA II algorithm, it converts user needs and the computing power network system's own performance indicators into multiple objectives for selection and optimization. A multi-objective coarse screening-personalized fine ranking scheme is designed. During the coarse screening phase, the objective function solution and the learning group are intelligently optimized, effectively improving the efficiency and accuracy of the computing power network resource scheduling system. Through this multi-objective optimization scheme, user SLA requirements for computing power networks can be met in all dimensions, the system response efficiency is improved by more than 50%, and the computing efficiency is improved by 30% after the system has run for a certain period of time. In addition, the personalized recommendation of fine-ranked nodes effectively improves the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0035] The present invention provides a multi-factor fusion scheduling method for computing power network, constructs a user objective function, converts quantitative indicators into a target value U that can be obtained by computing power nodes, and sets a coarse screening interval according to the target value U.

[0036] Construct the system objective function and convert the system performance indicators into the target value S that can be obtained by computing power nodes.

[0037] Based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function.

[0038] Use the multi-objective fusion scheduling function to perform multi-objective coarse screening of computing nodes to obtain all the objectives within the user's coarse screening interval.

[0039] Combine user goals and system performance indicators to perform personalized sorting of all goals within the obtained user demand range, and select the optimal scheduling plan based on the personalized sorting to schedule computing power.

[0040] The method of the present invention converts the computing power requirements, hardware requirements, service quality and system performance indicators required by the user into an optimization function for multi-objective fusion scheduling, and optimizes and solves it through the NSGA II algorithm. Computing power scheduling is performed according to the solution obtained, and the computing power scheduling process is divided into a multi-objective coarse screening + personalized fine sorting process. In the multi-objective coarse screening stage, non-dominated sorting is performed by the NSGA-II algorithm according to the node relationship table, and a scheduling scheme in which all targets are within a small range of user requirements is obtained. In the personalized fine sorting stage, the candidate set is sorted according to each target to obtain the optimal solution under different dimensions, which can be fed back to the interactive interface for user selection, and the node relationship table can be iteratively updated according to the relationship between the computing power nodes during each fine sorting, and the experience of each personalized fine sorting is learned and fed back to the next multi-objective coarse screening stage. At the same time, node clustering relationships can be established based on network relationships, node types and historical scheduling information, and the search algorithm's exploration strategy for the solution space is optimized to further improve the solution speed.

[0041] In some embodiments of the method of the present invention, the process can be referred to as follows:

[0042] Construct the user objective function and convert the judgment type index into the node pre-screening rule F according to the user SLA requirement index l , including the required computing power type (CPU, GPU, MLU, DCU, etc.), the required access network capabilities, the required software environment, the computing power storage ratio, etc.; quantitative indicators are converted into target values ​​that can be calculated through the underlying computing power node attributes, including the required underlying data volume, the required total CPU computing power, the required total storage, the node distance dispersion (expressed by variance), the total cost, the average access delay, etc., recorded as U (User), where U i Represents the i-th target value. Define 0.95*U - 1.05*U as the coarse screening interval. SLA: Service-Level Agreement. Judgment type refers to the type of indicator that requires a yes or no answer.

[0043] Construct the system objective function and convert the system performance indicators into target values ​​that can be calculated through the properties of the underlying computing nodes, including service energy consumption, total node utilization, total computing network revenue, etc., recorded as S (System), where S i represents the i-th target value.

[0044] Based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function. The multi-objective fusion scheduling function is used to perform multi-objective coarse screening of computing nodes to obtain all the objectives within the user coarse screening interval.

[0045] According to the pre-screening rule F lPerform computing power node screening to form a computing power node pool for multi-objective rough screening.

[0046] Use the multi-objective fusion scheduling function to perform multi-objective rough screening of computing power nodes, including: performing non-dominated sorting and solving based on the NSGA II algorithm according to U and S:

[0047] Step S1: Randomly generate initial solutions, and the number of initial solutions is greater than the number of set target values;

[0048] Step S2: Perform the first non-dominated sorting according to U. The non-dominated rank of non-dominated solutions is 1. After removing the non-dominated solutions with rank 1, a group with rank 2 is obtained, and so on. Within the same first non-dominated sorting rank, perform the second non-dominated sorting according to S, and the sorting method is the same as that of U to form a Pareto set;

[0049] Step S3: For all solutions obtained in S2, select them into the next-generation population in ascending order of non-dominated rank until the number of solution individuals in the first next-generation population is greater than the number of target values. Then, select the optimal solutions according to the crowding degree in the non-dominated rank of the first next-generation population until the number of solutions in the population is equal to the number of target values;

[0050] Step S4: Perform crossover and mutation operations based on the computing power node relationship table on the basis of the current solutions to generate new candidate solutions;

[0051] Step S5: Truncate the current solutions and retain the solution set P2 with the same number of solutions as the initial solutions;

[0052] Step S6: Repeat steps S2 - S4.

[0053] Furthermore, when performing the mutation operation in step S4, preferentially mutate the computing power nodes that co-occur in the historical scheduling scheme:

[0054] Step S41: Generate a random integer r1, where 0 < r1 < the required number of computing power nodes;

[0055] Step S42: Randomly select a computing power node from the current solutions and obtain the computing power node relationship table R of the computing power node i (i = r1), display the co-occurrence probability of the computing power node A and other computing power nodes related to A through the computing power node relationship table. When A appears, preferentially select other computing power nodes related to A according to the computing power node relationship table of A;

[0056] Step S43: Use the roulette wheel algorithm to obtain a computing power node among other computing power nodes related to A for mutation operation.

[0057] Furthermore, the solution of the first non-dominated sort in step S2 is binarized, including: for the target value U, if the current solution target value is greater than 0.95*U and less than 1.05*U, it is recorded as 1, otherwise it is 0, and the non-dominated sort is determined by using the solution after binarization.

[0058] In the non-dominated sorting judgment, for example, for solutions A and B, if the sum of the target values ​​of the binary solution of A is less than the sum of the target values ​​of B, then A must not dominate B. Otherwise, the dominance relationship is determined after comparing the target values ​​one by one.

[0059] Furthermore, during the second non-dominated sorting in step S2, sorting is performed in different dimensions according to the business definition, for example, the lower the energy consumption, the better, the higher the total revenue, the better, etc., and all dimensions are comprehensively considered. If a solution A is better than another solution B in all dimensions, then A is defined to dominate B.

[0060] Combine user goals and system performance indicators to perform personalized sorting based on all goals within the obtained user demand range, and select the optimal scheduling solution based on the personalized sorting to schedule computing power. Further, sort and filter according to the dimension of system performance indicators, and then perform personalized sorting according to the dimension of user goals to obtain the solution set P3. Based on the solution set P3, the optimal scheduling solution is selected.

[0061] For the candidate set P2, we first perform system screening according to the various dimensions of S, such as filtering solutions with too high energy consumption or too low computing network income.

[0062] Sort by each goal of U and take the best solution under each dimension, such as the lowest cost, lowest delay, shortest distance, etc., to obtain the candidate set P3.

[0063] The candidate set P3 is displayed on the front-end interface for personalized recommendation. The user selects the final solution, and then the relevant basic environment resources are called up and the application is deployed.

[0064] Post-analyze the solution selected by the user and build a node relationship knowledge table: For all computing power nodes i and j in the solution selected by the user, update the node relationship knowledge table: R ij =R ij +1, initially R = 1. The node relationship knowledge table can further represent the probability of node B appearing when node A appears, the probability of node C appearing, and so on. This is equivalent to establishing node clustering relationships, which can optimize the search algorithm's exploration strategy of the solution space and further improve the solution speed.

[0065] The method of the present invention combines user needs with computing network performance indicators to generate the objective function of NSGA II, which is then solved using a genetic algorithm. In the multi-objective coarse screening stage, the system uses set thresholds, a node relationship knowledge table, and the objective function solution to screen multiple scheduling solutions within the upper and lower ranges of the user's needs. The resulting series of scheduling solutions are then individually refined based on the satisfaction of each objective, yielding scheduling solutions with the lowest latency, lowest cost, and lowest distance for user selection. Cluster analysis is then performed on the scheduling solutions that meet the requirements, and a node relationship knowledge table is established, indicating the probability of node B appearing when node A appears, the probability of node C appearing, and so on. This relationship table can be used as a reference for mutation and crossover in the genetic algorithm using a roulette wheel algorithm. This achieves the comprehensive optimization of user needs and computing network performance indicators. By binarizing the objective function solution, the analysis rate of the supporting solution is accelerated; by using multi-objective optimization solutions, the user's SLA requirements can be guaranteed; by generating personalized recommendation solutions under multiple objectives, the user interactivity of the system is increased; by analyzing historical scheduling solutions and constructing a node relationship knowledge table, historical knowledge can be fully reused in the solution process, improving the system's evolutionary capabilities.

[0066] The present invention also provides a computing power network multi-factor fusion scheduling device, including a construction module, a coarse screening module, a fine sorting module and a selection module.

[0067] The construction module constructs the user objective function, converts the quantitative indicators into the target value U that can be obtained by computing power nodes, and sets the coarse screening interval according to the target value U; constructs the system objective function, converts the system performance indicators into the target value S that can be obtained by computing power nodes; based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function.

[0068] The coarse screening module uses the multi-objective fusion scheduling function to perform multi-objective coarse screening of computing nodes to obtain all targets within the user's coarse screening interval.

[0069] The fine sorting module combines user goals and system performance indicators to perform personalized sorting based on all goals within the obtained user demand range. The selection module selects the optimal scheduling plan for computing power scheduling based on the personalized sorting.

[0070] The information interaction, execution process, etc. between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention. For specific contents, please refer to the description in the embodiment of the method of the present invention and will not be repeated here.

[0071] Similarly, the device of the present invention, based on the NSGA II algorithm, converts user needs and the performance indicators of the computing network system itself into multiple objectives for selection and optimization. It designs a multi-objective coarse screening and personalized fine ranking scheme. In the coarse screening stage, the objective function solution and the learning group are intelligently optimized, effectively improving the efficiency and accuracy of the computing network resource scheduling system. Through this multi-objective optimization scheme, it is possible to meet user SLA requirements for the computing network in various dimensions, improve system response efficiency by more than 50%, and improve computing efficiency by 30% after the system has been running for a certain period of time. In addition, the personalized recommendation of fine ranking nodes effectively improves the user experience.

[0072] It should be noted that not all steps and modules in the above-mentioned processes and device structures are required, and certain steps or modules can be omitted according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or may be implemented by certain components in multiple independent devices.

[0073] The above-described embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A multi-factor fusion scheduling method for computing power network, characterized by Construct a user objective function, convert the quantization class metrics into the target value U that can be obtained by the calculation of the computing power nodes, set a coarse screening interval according to the target value U, and convert the judgment type metrics into node pre-screening rules F according to the user SLA requirement metrics. The judgment type metrics include computing power type, access network capacity, software environment, and computing power storage ratio. Construct a system objective function, convert the system performance metrics into the target value S that can be obtained through the calculation of the computing power node attributes. The system performance metrics include service energy consumption, total node utilization rate, and total computing power network income. Based on NSGA II, construct a multi-objective fusion scheduling function according to the user objective function and the system objective function. Use the multi-objective fusion scheduling function to perform multi-objective coarse screening of the computing power nodes, including: performing non-dominated sorting and solving based on U and S according to the NSGA II algorithm. Step S1: Randomly generate initial solutions, and the number of initial solutions is greater than the set number of target values. Step S2: Perform the first non-dominated sorting according to U. The non-dominated level of the non-dominated solutions is 1. After removing the non-dominated solutions with level 1, a group with level 2 is formed, and so on. Within the same first non-dominated sorting level, perform the second non-dominated sorting according to S, and the sorting method is the same as that of U to form a Pareto set. Step S3: For all the solutions obtained in S2, select them into the next-generation population in ascending order of non-dominated levels until the number of solution individuals in the first next-generation population is greater than the number of target values. Then, select the optimal solutions according to the crowding degree in the non-dominated level of the first next-generation population until the number of solutions in the population is equal to the number of target values. Step S4: Perform crossover and mutation operations based on the computing power node relationship table on the basis of the current solutions to generate new candidate solutions. When performing the mutation operation, preferentially perform the mutation operation on the computing power nodes that commonly appear in the historical scheduling schemes. Step S41: Generate a random integer r1, where 0 < r1 < the number of required computing power nodes. Step S42: Randomly select a computing power node from the current solution and obtain the computing power node relationship table R of the computing power node i (i=r1), the probability of computing node A and other computing nodes related to A appearing at the same time is displayed through the computing node relationship table. When A appears, other computing nodes related to A are preferentially selected according to A's computing node relationship table; Step S43: Obtain a computing power node from other computing power nodes related to A through the roulette wheel algorithm for mutation operation. Step S5: Truncate the current solutions, and retain the solution set P2 with the same number of solutions as the initial solutions. Step S6: Repeat steps S2 - S4. Obtain all the targets within the user's coarse screening interval. Combine the user objectives and the system performance metrics, perform personalized sorting based on all the targets within the obtained user requirement interval, and select the optimal scheduling scheme for computing power scheduling according to the personalized sorting.

2. A computing power network multi-factor fusion scheduling method according to claim 1, characterized in that Perform computing power node screening according to the pre-screening rules F to form a computing power node pool for multi-objective coarse screening.

3. The method for fusion scheduling of computing power network multi-factors according to claim 1 is characterized by: Perform binary processing on the solutions during the first non-dominated sorting in step S2, including: for the target value U, if the current solution target value is greater than 0.95*U and less than 1.05*U, it is recorded as 1, otherwise it is 0. Use the solutions after binary processing to judge the non-dominated sorting.

4. A computing power network multi-factor fusion scheduling device, characterized by It includes a construction module, a coarse screening module, a refined sorting module, and a selection module. The construction module constructs the user's objective function, converts the quantization class metrics into the objective value U that can be obtained by the calculation of the computing power nodes, sets the rough screening interval according to the objective value U, and converts the judgment type metrics into the node pre-screening rule F according to the user SLA requirement metrics. The judgment type metrics include computing power type, access network capacity, software environment, and computing power storage ratio. The construction module constructs the system objective function, converts the system performance metrics into the objective value S that can be obtained by the calculation of the computing power node attributes. The system performance metrics include service energy consumption, total node utilization rate, and total computing power network income. Based on NSGA II, a multi-objective fusion scheduling function is constructed according to the user objective function and the system objective function. The rough screening module uses the multi-objective fusion scheduling function to perform multi-objective rough screening of the computing power nodes, including: performing non-dominated sorting and solving based on U and S according to the NSGA II algorithm. Step S1: Randomly generate the initial solutions, and the number of initial solutions is greater than the number of set target values. Step S2: Perform the first non-dominated sorting according to U. The non-dominated rank of the non-dominated solutions is 1. After removing the non-dominated solutions with rank 1, a group with rank 2 is formed, and so on. Within the same first non-dominated sorting rank, perform the second non-dominated sorting according to S, and the sorting method is the same as that of U, forming the Pareto set. Step S3: For all the solutions obtained in S2, select them into the next-generation population in ascending order of the non-dominated rank until the number of solution individuals in the first next-generation population is greater than the number of target values. Then, select the optimal solutions according to the crowding degree in the non-dominated rank of the first next-generation population until the number of solutions in the population is equal to the number of target values. Step S4: Perform crossover and mutation operations on the basis of the current solutions according to the computing power node relationship table to generate new candidate solutions. When performing the mutation operation, preferentially mutate the computing power nodes that commonly appear in the historical scheduling schemes. Step S41: Generate a random integer r1, where 0 < r1 < the number of required computing power nodes. Step S42: Randomly select a computing power node from the current solution and obtain the computing power node relationship table R of the computing power node i (i=r1), the probability of computing node A and other computing nodes related to A appearing at the same time is displayed through the computing node relationship table. When A appears, other computing nodes related to A are preferentially selected according to A's computing node relationship table; Step S43: Obtain one computing power node from the other computing power nodes related to A through the roulette wheel algorithm for mutation operation. Step S5: Truncate the current solutions, and retain the solution set P2 with the same number of solutions as the initial solutions. Step S6: Repeat steps S2 - S4. Obtain all the targets within the user's rough screening interval. The fine sorting module performs personalized sorting based on all the targets within the obtained user demand interval by combining the user objectives and the system performance metrics. The selection module selects the optimal scheduling scheme according to the personalized sorting for computing power scheduling.

Images