A service resource allocation method for a multi-dimensional heterogeneous environment of a wisdom algorithm fusion network

Abstract

The application discloses a service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent algorithm fusion network, which comprises the following steps: receiving an algorithm network fusion service request ASR in a multi-dimensional heterogeneous environment; deploying virtual nodes and virtual links to physical nodes and physical links in the intelligent algorithm fusion network that meet constraint conditions to obtain a plurality of allocation schemes; taking each allocation scheme as a chromosome, and taking all the chromosomes as an initial population; randomly selecting a parameter in a segment length set and randomly pairing the chromosomes two by two; according to the parameter, dividing each pair of chromosomes into a plurality of gene segments, and calculating the fitness value of each gene segment; performing a segmented crossover and mutation operation on each pair of chromosomes to obtain new individuals; updating the population by using the new individuals, and then judging whether the iteration number is equal to a preset iteration number; if yes, selecting an individual with the highest fitness value in the population as a resource allocation scheme of the algorithm network fusion service request ASR; otherwise, returning to the parameter selection step.

Description

TECHNICAL FIELD

[0001] The present application relates to network resource allocation technology, in particular to a service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent algorithm fusion network. BACKGROUND

[0002] With the interconnection of multiple heterogeneous networks and the rapid development of artificial intelligence (AI) technology, especially the emergence of new algorithm network fusion services represented by AIGC (AI Generated Content), the demand for real-time audio and video generation, large-scale model training and inference, and other high-computing, low-latency computing and network interworking capabilities has shown explosive growth. To meet this demand, intelligent algorithm fusion networks, as a new network architecture that deeply integrates computing and network resources, have emerged. By efficiently scheduling distributed computing power, it collaborates with multi-dimensional heterogeneous resources such as computing, network, and storage to achieve deep integration of computing and network. Based on the service mechanism of on-demand resource allocation and dynamic scheduling, intelligent algorithm fusion networks can effectively support the diverse intelligent service needs of individual users, enterprise customers, and intelligent terminal devices in social media, marketing, education, and healthcare. It is of great significance to the construction of new digital infrastructure systems and the promotion of high-quality development of the digital economy.

[0003] However, the current mainstream centralized cloud architecture faces serious challenges in deploying algorithm network fusion services. Users call generative AI models or functional entities located in cloud data centers through the network to obtain services. This architecture exposes significant latency problems in practical applications, which not only affects user experience but also restricts the further development of latency-sensitive services. Another emerging solution is a distributed architecture based on edge computing. Using edge servers of ASP (algorithm network fusion service request), such as edge nodes equipped with high-performance computing devices (such as GPUs), intelligent algorithm fusion networks can shorten the physical distance between users and services by sinking computing resources to the vicinity of users, thereby achieving ultra-low latency, protecting privacy data, reducing bandwidth consumption, and improving overall energy efficiency, as shown in Figure 1 .

[0004] Part of the method by exhaustive or dynamic programming, etc. Systematic traversal of edge server, ensure to find the global optimal solution of the algorithm network integration service resource allocation, this kind of algorithm is usually based on mathematical model, strictly according to optimization theory to solve, can provide the optimal solution. Although this method can find the global optimal solution of the algorithm network integration service resource allocation, but its calculation complexity is very high, because service resource allocation often involves a large number of heterogeneous network (topology structure and scale flexible and variable), multi-dimensional heterogeneous resources (such as computing, storage, bandwidth, etc.), and complex constraint conditions (such as delay limit, etc.), the solution time complexity of exhaustive or dynamic programming is usually exponential, the calculation time and resource consumption rise rapidly, it is difficult to meet the real-time requirements in practical application. In addition, this method may not be able to run in large-scale scenarios due to excessive memory consumption or long calculation time, which weakens the response ability of the algorithm network integration service in the intelligent algorithm integration network, reduces the acceptance rate of the service and the user experience.

[0005] In order to reduce the solving time, another kind of method adopts the technical scheme based on experience and rules to quickly provide a feasible solution or suboptimal solution of resource allocation. The core of these methods is to use heuristic rules or evaluation functions to sort or prioritize the edge servers and resource requirements, thereby guiding the search of the solution space and avoiding the comprehensive traversal of all possible solutions by traditional methods.

[0006] Firstly, this kind of method highly depends on the design of heuristic rules, and the heuristic rules often need the support of the experience of domain experts. If the designer does not fully understand the requirements of the algorithm network integration service or the characteristics of the intelligent algorithm integration network, the rules may not be reasonable, thereby leading to the performance degradation of the resource allocation scheme and affecting the resource utilization rate. Secondly, the quality of the solution lacks stability, which may find a relatively optimal solution in a simple scenario, but easily fluctuates greatly in a complex scenario, and even deviates from the global optimal solution. In addition, this kind of method usually lacks mathematical theory support, and it is difficult to provide explicit performance guarantee or theoretical limit for the quality of the solution, which increases the uncertainty of the application. SUMMARY

[0007] In view of the above problems in the prior art, the service resource allocation method for the multi-dimensional heterogeneous environment of the intelligent algorithm integration network provided by the present application solves the problems of low resource utilization rate and low service acceptance rate of the existing method network.

[0008] In order to achieve the above invention purpose, the technical scheme adopted by the present application is:

[0009] A service resource allocation method for the multi-dimensional heterogeneous environment of the intelligent algorithm integration network is provided, which includes the following steps:

[0010] S1, receive an algorithm network fusion service request ASR in a multi-dimensional heterogeneous environment, the algorithm network fusion service request ASR includes a plurality of virtual nodes and a plurality of virtual links;

[0011] S2, deploy the virtual nodes and the virtual links to the physical nodes and the physical links in the intelligent algorithm fusion network that meet the constraint conditions, to obtain a plurality of allocation schemes including node mapping and link mapping;

[0012] S3, each allocation scheme is taken as an initial chromosome, and all chromosomes are taken as an initial population;

[0013] S4, a parameter is randomly selected in a segment length set, and the chromosomes are randomly paired; according to the parameter, each pair of chromosomes is divided into a plurality of gene segments, and the fitness value of each gene segment is calculated;

[0014] S5, according to the fitness of the gene segment, the segmented crossover and mutation operations are performed on each pair of chromosomes to obtain new individuals;

[0015] S6, the population is updated by using the new individuals, and then it is judged whether the iteration number is equal to a preset iteration number, if yes, step S7 is entered, otherwise, step S4 is returned;

[0016] S7, an allocation scheme composed of the node mapping and the link mapping corresponding to the individual with the highest fitness value in the population is selected as a resource allocation scheme of the algorithm network fusion service request ASR.

[0017] Further, step S2 further includes:

[0018] S21, a physical node meeting the constraint condition in the intelligent algorithm fusion network is randomly selected and one-to-one mapped with each virtual node to obtain a node mapping scheme;

[0019] S22, according to the physical network topology composed of the selected physical nodes, the k-shortest paths algorithm is used to calculate k shortest paths based on the physical network topology for each virtual link,

[0020] S23, all candidate paths with bandwidth meeting the demand of the virtual link are screened out from the k shortest paths, and a candidate path is randomly selected as the link mapping of the virtual link;

[0021] S24, the node mapping scheme and the corresponding link mapping are taken as an allocation scheme, steps S21 to S23 are repeated until a plurality of allocation schemes are obtained.

[0022] Further, the expression for calculating the fitness value of the gene segment is:

[0023]

[0024] Among them, S i A gene segment of any chromosome of the i-th computational network fusion service request ASR; (S i ) is S i The fitness value of α is the predetermined weight of the mapping reward and link cost ratio; β is the predetermined weight of the resource coordination efficiency; For S i Mapping rewards; For S i Link overhead; Resource allocation plan Resource coordination efficiency; n v and n p They are virtual nodes and physical nodes respectively; The set of all node mappings for the i-th computing-network convergence service request ASR; is a resource dimension set, C, G, M, S, and B are all resource dimensions; n v The sum of all resources required; n p The total capacity of all resources; A(n p ) is the resource capacity of the physical node on A; A is any dimension in the resource dimension set; A(n v ) is the virtual node n v The resource demand on A; δ is a constant.

[0025] Furthermore, the method of performing segmented crossover on a chromosome includes:

[0026] S51. Identify the best and worst gene segments in each chromosome based on the fitness values ​​of the gene segments;

[0027] S52: Determine whether to perform a crossover operation based on a preset crossover probability. If yes, proceed to step S53; otherwise, terminate the crossover operation.

[0028] S53. In each pair of chromosomes, each chromosome exchanges the best gene segment of the other chromosome with its corresponding segment, and exchanges its worst gene segment with the corresponding segment of the other chromosome;

[0029] S54: Determine whether there are duplicate physical nodes in the new individual obtained after the crossover. If so, proceed to step S55; otherwise, output the new individual.

[0030] S55. Check whether there is a non-duplicate physical node in its parent generation. If so, proceed to step S56. Otherwise, delete the current new individual.

[0031] S56, randomly select a non-repeated physical node to replace any repeated physical node in the new individual, and output the updated new individual.

[0032] Further, the method for performing mutation operation on the chromosome after the crossover operation comprises:

[0033] A1, the non-optimal gene fragment part in the chromosome after the mutation operation is selected as a mutation optional range;

[0034] A2, according to a preset mutation probability, a mutation point is selected in the mutation optional range of each chromosome;

[0035] A3, by using a random remapping method, the physical nodes and physical links that meet the resource constraints are selected to be deployed at the mutation point in the physical network composed of the physical nodes of the intelligent algorithm fusion network, to obtain a new individual.

[0036] Further, between step S5 and step S6, an elite selection strategy is used to select K new individuals with the largest fitness value as the final new individuals from all the obtained new individuals.

[0037] Further, the constraint condition comprises:

[0038] The resource demand constraint of the algorithm network fusion service request ASR:

[0039]

[0040] Wherein, n v and n p are virtual nodes and physical nodes respectively; is a set composed of all node mappings of the i-th algorithm network fusion service request ASR; C(n v ) and C(n p ) are the CPU resource capacities of n v and n p respectively;

[0041] The GPU resource capacity constraint of the physical node and the virtual node:

[0042]

[0043] Wherein, G(n v ) and G(n p ) are the GPU resource capacities of n v and n p respectively;

[0044] The memory resource demand constraint of the physical node and the virtual node:

[0045]

[0046] where M(n v ) and M(n p ) are the corresponding memory resource requirements of n v and n p , respectively;

[0047] Storage resource requirement constraints of physical nodes and virtual nodes:

[0048]

[0049] where S(n v ) and S(n p ) are the corresponding storage resource requirements of n v and n p , respectively;

[0050] Bandwidth resource capacity constraints of physical links and virtual links:

[0051]

[0052] where e p and e v are physical links and virtual links, respectively; E p is the set of all physical links in the intelligent computing and convergence network; is the set of all virtual links mapped to physical link e p ; B(e v ) and B(e p ) are the corresponding bandwidth resource requirements of e v and e p , respectively;

[0053] Delay constraints of all physical links to which a virtual link is mapped:

[0054]

[0055] where, is the set of all physical links to which e v is mapped; D(e p ) and D(e v ) are the delay of e p and the maximum allowed delay of e v , respectively.

[0056] Further, the constraint conditions further include a reward function constraint and a link overhead constraint, respectively:

[0057]

[0058]

[0059] where, is the i-th algorithmic network fusion service request ASR; is the reward function of ; ω c , ω g , ω m , ω s and ω b are the weights of C(n v ), G(n v ), M(n v ), S(n v ) and B(e v ) respectively; is the virtual node set of the i-th algorithmic network fusion service request ASR; N p is the physical node set in the algorithmic fusion network; is the virtual communication link set of the i-th algorithmic network fusion service request ASR; is the mapping solution set of to the undirected graph G p composed of the algorithmic fusion network; is the link cost function of . The mapping solution includes node mapping and link mapping, and the solution or mapping solution mentioned below is a complete chromosome, that is, an allocation scheme.

[0060] Compared with the prior art, the beneficial effects of the present application are:

[0061] 1. Low deployment and operation cost. The present scheme is suitable for general server clusters and GPU clusters and does not depend on specific hardware or software platforms. Its implementation process can fully utilize existing software and hardware resources and is simple to interface with existing edge computing systems. It only needs to receive service requests and perceive the topology and resource information of the underlying network, without the need to replace server configurations and network configurations or invest in additional hardware facilities. The optimization results output by the method can be accessed by the algorithmic fusion network controller for execution, thereby significantly reducing the complexity of deployment and operation.

[0062] 2. High solving efficiency. The present scheme significantly improves the exploration efficiency and convergence speed of the algorithm and can quickly generate high-quality solutions. Through the combination of chromosome solution space segmentation technology, crossover mutation strategy and elite selection mechanism, the present scheme exhibits excellent performance in solving complex problems. It can comprehensively consider link cost, reward and resource utilization rate to ensure that the solution is comprehensive optimal in various key indicators. At the same time, its efficient optimization mechanism also enables it to handle complex problems with multiple constraints and multiple objectives, making it a fast and efficient solving method.

[0063] ​3. Strong scalability. This solution can adapt to both small edge node clusters and large distributed server clusters without additional complex environment configuration. In addition, this solution allows flexible adjustment of parameters (such as segment length) according to different application scenarios to adapt to optimization tasks of various scales, and can be easily integrated into existing edge computing systems.

[0064] 4. High resource utilization. This solution can minimize resource waste and prevent resource fragmentation, thereby improving the overall utilization of physical resources. This efficient resource management capability is particularly important for scenarios in the intelligent computing and networking network where resources are limited and demand is dynamic.

[0065] 5. Wide application scenarios. This solution is not only suitable for service deployment and resource allocation in the intelligent computing and networking network, but can also be extended to other complex optimization scenarios, such as large-scale task scheduling, network function virtualization (NFV) mapping, cloud computing resource scheduling, etc. Its framework and method ideas have wide generality and portability. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 An illustration of service deployment in the intelligent computing and networking network.

[0067] Figure 2 A flowchart of the service resource allocation method for the multi-dimensional heterogeneous environment of the intelligent computing and networking network.

[0068] Figure 3 An example diagram of node mapping, edge mapping, and chromosome segmentation.

[0069] Figure 4 An illustration of the segment-based crossover operation.

[0070] Figure 5 An illustration of the crossover and mutation operations. DETAILED DESCRIPTION

[0071] The specific embodiments of the present application are described below to facilitate understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments. For those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application as defined in the appended claims, and all inventions utilizing the concept of the present application are within the scope of protection.

[0072] In this solution, the intelligent computing and networking network is represented as an undirected graph G p = (N p , E p ), where N p represents the set of edge physical servers, and E p represents the set of physical links. Np Any physical node n in p There is a resource attribute set {C(n p ),G(n p ),M(n p ),S(n p ),B(n p )}, representing the CPU, GPU, memory, storage and bandwidth resource capacities respectively; E p Any physical link e in p , its bandwidth capacity is B(e p ), the link delay is D(e p These resource dimensions are the main dimensions of computing-network integration services.

[0073] Since the computing-network fusion service needs to be distributed and deployed on edge servers, the upcoming i-th ASR is also represented as a virtual undirected graph Virtual node collection and virtual link sets They represent microservices or service function entities and the network topology within the service respectively.

[0074] Each virtual node n in v All have resource requirements {C(u v ),G(u v ),M(u v ),S(u v ),B(u v )}. Each virtual link e in v Both have bandwidth requirements B(e v ) and the maximum allowable delay D(e v ). Resource allocation for edge AIGC services can be viewed as a sequential online decision problem, where ASRs continuously arrive and leave, occupying the resources of the intelligent computing fusion network within a specific time. We define yes to G p The mapping solution of the node is The node mapping is one-to-one, one virtual node corresponds to one physical node, and the link mapping is Link mapping is one-to-many. A virtual link may require multiple physical links to be connected end to end. It is e v The set of physical links mapped to, defines All mapped to physical links p A collection of virtual links.

[0075] refer to Figure 2 ,Figure 2 A service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent algorithm fusion network is shown. Figure 2 As shown, the method S includes steps S1-S7.

[0076] In step S1, an algorithm network fusion service request ASR under a multi-dimensional heterogeneous environment is received, and the algorithm network fusion service request ASR includes a plurality of virtual nodes and a plurality of virtual links.

[0077] In step S2, the virtual nodes and the virtual links are deployed to physical nodes and physical links in the intelligent algorithm fusion network that meet the constraint conditions, to obtain a plurality of allocation schemes including node mapping and link mapping.

[0078] In this scheme, the node mapping indicates that each functional entity (such as a virtual machine or a service image container) in the virtual network should be deployed to which physical node, and the link mapping indicates that the logical link in the virtual network should be composed of which physical link path, which can be referred to the legend shown in the upper part of Figure 3 The rule of node mapping is that each virtual node is one-to-one mapped with a physical node, and the demand of the virtual node must be met, and the rule of link mapping requires that the physical path mapped by each virtual link can meet the topology connectivity and bandwidth delay constraints. The node numbers of the physical network and the virtual network can be in any order, but once determined in the initialization phase, the subsequent process will not be changed.

[0079] In an embodiment of the present application, step S2 further includes:

[0080] S21, randomly selecting physical nodes in the intelligent algorithm fusion network that meet the constraint conditions and one-to-one mapping with each virtual node to obtain a node mapping scheme;

[0081] S22, according to the physical network topology composed of the selected physical nodes, using the k-shortest paths algorithm to calculate k shortest paths based on the physical network topology for each virtual link,

[0082] S23, in the k shortest paths, screening all candidate paths that meet the bandwidth demand of the virtual link, and randomly selecting a candidate path as the link mapping of the virtual link;

[0083] S24, taking the node mapping scheme and the corresponding link mapping as an allocation scheme, repeating steps S21-S23 until a plurality of allocation schemes are obtained.

[0084] In step S3, each allocation scheme is taken as an initial chromosome, and all chromosomes are taken as an initial population. Then, a segment length set and other hyperparameters (hyperparameters related to genetic algorithm) are initialized. The segment length set is a constant set, for example, {3, 5, 7, 9}, which is a hyperparameter newly added in the algorithm. The value of the segment length should be referred to the size of the virtual network. For example, when the virtual network is small, the size of the segment length set is small, for example, {3, 4, 5}. When the virtual network is large, the value of the segment length set can be {3, 6, 9, 12}. The specific value can be determined by experience or randomly.

[0085] In step S4, a parameter is randomly selected from the segment length set, and chromosomes are randomly paired two by two. According to the parameter, each pair of chromosomes is divided into multiple gene segments, and the fitness value of each gene segment is calculated.

[0086] The individual fitness value can evaluate the quality of a complete solution (resource allocation scheme), and the fitness based on segmentation evaluates the quality of part of the genes in a chromosome. In order to quantify the pros and cons of the multi-dimensional resource allocation scheme, the total amount of resources required by a virtual node is defined as follows: wherein represents a set of different resource dimensions. Similarly, the total resource capacity of a physical node is:

[0087] In implementation, the expression for calculating the fitness value of the gene segment is preferably as follows:

[0088]

[0089] wherein, S i is a gene segment of any chromosome of the i th algorithmic network integration service request ASR; (S i ) is the fitness value of S i ; α is a predetermined weight of the mapping reward and the link overhead ratio; β is a predetermined weight of the resource coordination efficiency; is the mapping reward of S i ; is the link overhead of S i ; is the resource coordination efficiency of the resource allocation scheme ; n v and n p are virtual nodes and physical nodes, respectively; is a set composed of the mapping of all nodes of the i th algorithmic network integration service request ASR; is a set of resource dimensions, and C, G, M, S, and B are resource dimensions; is the total amount of all resources required by n v ; is np The total capacity of all resources; A(n p ) is the resource capacity of the physical node on A; A is any dimension in the resource dimension set; A(n v ) is the virtual node n v The resource demand on A; δ is a constant.

[0090] By maximizing the fitness value, this scheme can optimize the reward of the current mapping while minimizing the link overhead and resource imbalance, thereby improving long-term benefits and resource utilization.

[0091] In order to facilitate the understanding of "dividing each pair of chromosomes into multiple gene segments according to parameters", the following Figure 3 For detailed explanation:

[0092] The basic idea of ​​chromosome segmentation is to group related nodes and link mappings into the same segment. In the invention, the segments are assumed to be continuous and uniform. All virtual nodes are treated equally, allowing arbitrary ordering within the chromosome. Given an ASR with N nodes, a set of discrete segment lengths L = {l1, l2, ..., l k}. In each iteration, a segment length l is randomly selected t ∈L. Starting from the first virtual node, this scheme divides the continuous node mapping sequence and its associated link mapping that meet the length requirement in each chromosome into segments Any remaining sequence that is not long enough is also classified as a chromosome segment. For example:

[0093] A chromosome has nine nodes, numbered starting from 1: 123456789. Edges can be considered as tuples, such as (1,2) representing an edge between virtual nodes 1 and 2. Assuming the chosen parameter (i.e., segment length) is 4, then the first gene segment includes the mappings for nodes 1234 and their associated edges; the second segment similarly includes 5678 and their associated edges; and the last segment, although not long enough, is still considered a chromosome segment, consisting of node 9 and its associated edges.

[0094] It is important to note that the node mappings in the chromosome segmentation scheme do not overlap with each other, but they may share some link mappings because the nodes at both ends of the link can be assigned to different segments. Subsequently, the quality of these segments is evaluated to guide the subsequent crossover and mutation operations.

[0095] Figure 3The middle and lower part of the chromosome includes six nodes, the segment length is set to 2, the segment recombination and the edges associated with the nodes are set, under the segment rule, the segment is uniquely determined, but because a virtual edge has two nodes, the two nodes can be divided into different segments, at this time, the edge mapping needs to be added to the corresponding segment.

[0096] In step S5, according to the fitness of the gene segment, the segment crossover and mutation operation is performed on each pair of chromosomes to obtain new individuals.

[0097] In an embodiment of the present application, the method for performing segment crossover on chromosomes comprises:

[0098] S51, according to the fitness value of the gene segment, the optimal and worst gene segments in each chromosome are identified;

[0099] S52, according to the preset crossover probability, it is judged whether to perform the crossover operation, if yes, step S53 is entered, otherwise the crossover operation is ended;

[0100] S53, in each pair of chromosomes, the best gene segment in the other chromosome is exchanged with the corresponding segment, and the worst gene segment is exchanged with the corresponding segment of the other chromosome;

[0101] S54, it is judged whether there is a repeated physical node in the new individual obtained after the crossover, if yes, step S55 is entered, otherwise the new individual is outputted;

[0102] S55, it is judged whether there is a non-repeated physical node in its parent, if yes, step S56 is entered, otherwise the current new individual is deleted;

[0103] S56, a non-repeated physical node is randomly selected to replace any repeated physical node in the new individual, and the updated new individual is outputted.

[0104] The above-mentioned method is used for crossover, which can improve the exploration efficiency by using high-quality genes in complex optimization scenarios. The determination of the crossover range depends on the quality evaluation of the segments of different chromosomes. In order to facilitate the understanding of the crossover process, Figure 4 An example of the crossover operation is given, in which each chromosome includes 10 nodes, the segment length is 2, and each chromosome is divided into 5 gene segments.

[0105] In the implementation, the method for performing mutation operation on the chromosomes after the crossover operation is preferably:

[0106] A1, the non-optimal gene segment part in the chromosome after the mutation operation is selected as the mutation optional range;

[0107] A2, selecting the mutation points in the mutation optional range of each chromosome according to the preset mutation probability;

[0108] A3, using the random remapping method, selecting the physical nodes and physical links satisfying the resource constraints to deploy at the mutation points in the physical network composed of the physical nodes of the intelligent algorithm fusion network, to obtain a new individual.

[0109] In order to facilitate the understanding of the crossover, Figure 5 The mutation example is shown, which is based on Figure 4 The upper two chromosomes in the figure are the initial chromosomes before the crossover, the middle two are the chromosome diagrams after the crossover using the steps S51-S56 of the scheme, and the last two are the chromosome structures obtained after the mutation based on the crossover.

[0110] In order to improve the efficiency of the crossover and mutation operations, the quality of the gene fragments is evaluated (the fitness value of the gene fragments is calculated) before the crossover and mutation are performed. The quality evaluation process comprehensively considers the rewards, link overheads and resource coordination efficiency of the fragments, to ensure that high-quality gene fragments are preferentially retained in the genetic operation.

[0111] In order to ensure that a high-quality resource allocation scheme is obtained subsequently, between step S5 and step S6, an elite selection strategy is used to select the K new individuals with the largest fitness values from all the new individuals obtained as the final new individuals.

[0112] In step S6, the new individuals are used to update the population, and then it is judged whether the iteration number is equal to the preset iteration number. If yes, step S7 is entered, otherwise step S4 is returned.

[0113] In step S7, the node mapping and link mapping corresponding to the individual with the highest fitness value in the population are selected to form an allocation scheme for the resource allocation scheme of the algorithm network fusion service request ASR.

[0114] In an embodiment of the present application, the constraint conditions include:

[0115] The resource demand constraint of the algorithm network fusion service request ASR:

[0116]

[0117] Wherein, n v and n p are virtual nodes and physical nodes respectively; is a set composed of all node mappings of the i-th algorithm network fusion service request ASR; C(n v ) and C(n p ) are n v and np Corresponding CPU resource capacity;

[0118] GPU resource capacity constraints of physical nodes and virtual nodes:

[0119]

[0120] Wherein, G(n v ) and G(n p ) are the corresponding GPU resource capacities of n v and n p respectively;

[0121] Memory resource demand constraints of physical nodes and virtual nodes:

[0122]

[0123] Wherein, M(n v ) and M(n p ) are the corresponding memory resource demands of n v and n p respectively;

[0124] Storage resource demand constraints of physical nodes and virtual nodes:

[0125]

[0126] Wherein, S(n v ) and S(n p ) are the corresponding storage resource demands of n v and n p respectively;

[0127] Bandwidth resource capacity constraints of physical links and virtual links:

[0128]

[0129] Wherein, e p and e v are physical links and virtual links respectively; E p is the set of all physical links in the intelligent algorithm fusion network; is the set of all virtual links mapped to the physical link e p ; B(e v ) and B(e p ) are the corresponding bandwidth resource demands of e v and e p respectively;

[0130] Delay constraints of all physical links to which the virtual link is mapped:

[0131]

[0132] wherein, is e v a set of all physical links mapped to; D(e p ) and D(e v ) are the delay of e p and the maximum allowed delay of e v , respectively;

[0133] The constraints also include reward function constraints and link cost constraints, which are:

[0134]

[0135] wherein, is the ith artificial service request ASR; is the reward function of ; ω c , ω g , ω m , ω s and ω b are the weights of C(n v ), G(n v ), M(n v ), S(n v ) and B(e v ), respectively; is a set of virtual nodes of the ith artificial service request ASR; N p is a set of physical nodes in the intelligent computing and converging network; is a set of virtual communication links of the ith artificial service request ASR; is the mapping solution set of to the undirected graph G p formed by the intelligent computing and converging network; is the link cost function of .

[0136] The scheme can further improve the service reception rate, system reward, reduce system cost and resource fragmentation under the above constraints.

[0137] The resource allocation scheme provided in the scheme has simple deployment environment and can run on a general server without relying on specific hardware environment or special equipment, and has strong environmental adaptability. The method can be deployed at any location of the intelligent algorithm and network fusion network, and the nodes in the intelligent algorithm and network fusion network can be interconnected according to any physical topology structure, as long as the communication interworking between all nodes is ensured, the operation requirement of the method can be met. At the same time, the topology structure and resource information of the intelligent algorithm and network fusion network need to be perceived, including the key resource states such as computing power, storage capacity, network bandwidth, to provide comprehensive data support for the method.

[0138] In addition, the method also needs to have the function of receiving algorithm and network fusion service request, which can process the service request from the user or the application, including the information such as the logical topology structure of the resource demand of the algorithm and network fusion service. After the method is run, the output result, that is, the optimal resource mapping scheme of the algorithm and network fusion service, can be directly connected to the intelligent algorithm and network fusion network controller, and the controller can execute specific service mirror deployment operation according to the mapping scheme, including mapping the virtual service mirror to the physical node and mapping the virtual link to the physical link, so as to realize the decoupling with the specific deployment operation.

[0139] In summary, the algorithm and network fusion service resource allocation method provided in the scheme can maximize the reduction of resource fragmentation, reduce the service deployment cost and improve the long-term average income through reasonable resource scheduling and load balancing strategy. At the same time, the scheme can adapt to the heterogeneous and variable service demand and resource state in the edge environment, realize the fast solution scheme and improve the efficiency. The scheme has good expansibility and can adapt to the growth of network node number, the change of service demand and the iteration of future technology, so as to provide long-term stable technical support and operation optimization ability for the algorithm and network fusion service.

Claims

1. A service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network, characterized in that: Including steps: S1. Receive a computing-network convergence service request (ASR) in a multi-dimensional heterogeneous environment, where the computing-network convergence service request (ASR) includes multiple virtual nodes and multiple virtual links. S2. Deploy virtual nodes and virtual links to physical nodes and physical links that meet the constraints in the intelligent computing fusion network, and obtain multiple allocation schemes including node mapping and link mapping; S3. Each allocation scheme is regarded as an initial chromosome and all chromosomes are regarded as the initial population; S4. Randomly select a parameter from the segment length set and randomly pair the chromosomes. According to the parameter, each pair of chromosomes is divided into multiple gene segments, and the fitness value of each gene segment is calculated. S5. According to the fitness of the gene segments, perform segmented crossover and mutation operations on each pair of chromosomes to obtain new individuals; S6. Update the population using the new individual, and then determine whether the number of iterations is equal to the preset number of iterations. If so, proceed to step S7, otherwise return to step S4; S7. Selecting the allocation scheme consisting of the node mapping and link mapping corresponding to the individual with the highest fitness value in the population as the resource allocation scheme for the computing network convergence service request ASR; Methods for performing segmented crossover on chromosomes include: S51. Identify the best and worst gene segments in each chromosome based on the fitness values ​​of the gene segments; S52: Determine whether to perform a crossover operation based on a preset crossover probability. If yes, proceed to step S53; otherwise, terminate the crossover operation. S53. In each pair of chromosomes, each chromosome exchanges the best gene segment of the other chromosome with its corresponding segment, and exchanges its worst gene segment with the corresponding segment of the other chromosome; S54: Determine whether there are duplicate physical nodes in the new individual obtained after the crossover. If so, proceed to step S55; otherwise, output the new individual. S55. Check whether there is a non-duplicate physical node in its parent generation. If so, proceed to step S56. Otherwise, delete the current new individual. S56. Randomly select a non-duplicate physical node to replace any duplicate physical node in the new individual, and output the updated new individual; The expression for calculating the fitness value of a gene fragment is: in, A gene segment of any chromosome of the i-th computational network fusion service request ASR; for The fitness value of is the predetermined weight of the mapping reward and link cost ratio; predefined weights for resource coordination efficiency; for Mapping rewards; for Link overhead; Resource allocation plan resource coordination efficiency; and They are virtual nodes and physical nodes respectively; The set of all node mappings for the i-th computing-network convergence service request ASR; is a resource dimension set, All are resource dimensions; for The sum of all resources required; for The sum of all resource capacities; is the resource capacity of the physical node on A; A is any dimension in the resource dimension set; For virtual nodes The resource demand on A; is a constant.

2. The service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network according to claim 1 is characterized in that: Step S2 further comprises: S21. Randomly select a physical node that meets the constraint conditions in the intelligent computing fusion network and perform one-to-one mapping with each virtual node to obtain a node mapping solution; S22. Based on the physical network topology formed by the selected physical nodes, k shortest paths based on the physical network topology are calculated for each virtual link using the k-shortest paths algorithm. S23. Among the k shortest paths, all candidate paths whose bandwidth meets the virtual link requirement are screened out, and a candidate path is randomly selected as the link mapping of the virtual link. S24: Take the node mapping scheme and the corresponding link mapping as an allocation scheme, and repeat steps S21 to S23 until multiple allocation schemes are obtained.

3. The service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network according to claim 1 is characterized in that: Methods for performing mutation operations on chromosomes after crossover operations include: A1. The non-optimal gene fragments in the chromosome after the mutation operation are selected as the mutation range; A2. Select a mutation point within the mutation range of each chromosome based on the preset mutation probability; A3. Using the random remapping method, in the physical network composed of the physical nodes of the intelligent computing fusion network, physical nodes and physical links that meet the resource constraints are selected and deployed at the mutation point to obtain new individuals.

4. The service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network according to claim 1 is characterized in that: Also included between step S5 and step S6: Among all the new individuals obtained, an elite selection strategy is adopted to select the K new individuals with the largest fitness values ​​as the final new individuals.

5. The service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network according to any one of claims 1 to 4, characterized in that: The constraints include: Resource requirements for ASR requests for computing and network convergence services: in, and They are virtual nodes and physical nodes respectively; The set of all node mappings for the i-th computing-network convergence service request ASR; and They are and Corresponding CPU resource capacity; GPU resource capacity constraints for physical and virtual nodes: in, and They are and Corresponding GPU resource capacity; Memory resource requirements for physical and virtual nodes: in, and They are and The corresponding memory resource requirements; Storage resource requirements for physical and virtual nodes: in, and They are and The corresponding storage resource requirements; Bandwidth resource capacity constraints for physical links and virtual links: in, and They are physical links and virtual links respectively; It is the collection of all physical links in the intelligent computing converged network; For all mappings to physical links A collection of virtual links; and They are and The corresponding bandwidth resource requirements; Delay constraints of all physical links to which the virtual link is mapped: in, for The set of all physical links mapped to; and They are The delay and The maximum allowed delay.

6. The service resource allocation method for a multi-dimensional heterogeneous environment of an intelligent computing fusion network according to claim 5 is characterized in that: The constraints also include reward function constraints and link cost constraints, which are: in, Request ASR for the i-th computing-network fusion service; for The reward function of 、 、 、 and They are 、 、 、 and The weight of The set of virtual nodes that request ASR for the i-th computing-network fusion service; It is a collection of physical nodes in the intelligent computing fusion network; The set of virtual communication links for the i-th computing-network fusion service request ASR; for To the undirected graph composed of intelligent computing fusion network The mapping solution set of ; for Link cost function.

Images