Self-adaptive micro-service deployment resource optimization method oriented to cloud side-end collaboration

Abstract

The invention discloses a self-adaptive micro-service deployment resource optimization method for cloud edge-end cooperation. The method comprises the following steps: 1) modeling a system model; 2) modeling a communication model; 3) modeling a user demand processing delay module; 4) deploying resource consumption according to a micro-service type required by user demand processing, and modeling a micro-service deployment resource consumption model and a user demand processing resource consumption model; 5) performing formulation definition on the micro-service deployment resource optimization objective function to complete heterogeneous constraint high-quality characterization; 6) aiming at the difference dimension of the multi-objective optimization function, finishing normalization processing of different dimensions; and 7) solving a multi-objective optimization problem by adopting a tabu search method to realize effective utilization of resources in a deployment process.The invention belongs to the technical field of system resource optimization of the Internet of Things, and solves the problem that the Internet of Things in the prior art has defects in the aspects of application space demand coupling, decentralized coordination and service difference.

Description

Technical Field

[0001] The present invention belongs to the technical field of resource optimization of Internet of Things systems, and relates to an adaptive microservice deployment resource optimization method for cloud-edge-end collaboration. Background Art

[0002] The edge computing paradigm offers greater possibilities for modern mobile intelligent applications. Meanwhile, the microservices architecture model, which breaks down a single, large service into multiple, fine-grained, small, independent microservices, provides greater flexibility for mobile intelligent applications. Microservices interact via lightweight communication mechanisms (such as HTTP / REST), offering advantages such as flexibility, maintainability, and scalability.

[0003] Currently, task offloading and resource scheduling are mainstream research areas in edge computing resource optimization applications, focusing on dynamic offloading, multi-objective optimization, and predictive models. Dynamic offloading determines whether to offload tasks to edge servers or the cloud based on current network status and computing resource availability. Multi-objective optimization simultaneously considers resource consumption, computing energy consumption, service latency, and computing resource utilization to optimize task offloading decisions. Predictive models leverage machine learning and data mining techniques to predict future resource needs and proactively offload tasks. Resource allocation aims to maximize resource utilization efficiency across the entire network or system and reduce resource consumption costs. Research focuses on load balancing, user demand management, resource elastic scaling, and multi-tenant management. By monitoring network or system load in real time, resource scheduling and task allocation are optimized based on load balancing, avoiding issues such as decreased network resource utilization and uneven network load. Elastic scaling dynamically adjusts the number of service instances based on real-time load to ensure optimal system performance and resource utilization. Multi-tenant management rationally allocates and isolates resources in a multi-tenant environment to ensure service quality for different tenants.

[0004] Despite extensive research focusing on optimizing microservice deployment, efficient deployment based on specific user needs remains challenging in practical applications. The diverse nature of user needs, the differentiated nature of service types, and the coupled nature of heterogeneous IoT application space requirements make dynamically adjusting service deployment strategies based on specific needs a challenge. In edge computing, edge server computing resources are limited. Therefore, achieving decentralized resource coordination and scheduling to meet user needs and optimize overall network resource utilization within these resource constraints is a key issue for ensuring user service quality, user experience quality, and system performance. User needs and network environments frequently change, and rapidly adapting to these changes to ensure service continuity and stability is also a challenge. Furthermore, ensuring service quality for each user in a multi-tenant environment and avoiding performance degradation caused by resource competition are also crucial issues. Summary of the Invention

[0005] The purpose of the present invention is to provide an adaptive microservice deployment resource optimization method for cloud-edge-end collaboration. While ensuring the collaboration of edge servers to realize resource optimization service deployment decisions, it solves the problems of the existing Internet of Things in terms of application space demand coupling, decentralized coordination and service differences.

[0006] The technical solution adopted by the present invention is an adaptive microservice deployment resource optimization method for cloud-edge-end collaboration, which includes the following steps: Step 1: Model the system model and complete the definition of user requirements and microservice types; Step 2: Model the communication model and obtain the transmission rate of the user transmission task; Step 3: Model the user demand processing delay module to obtain the total delay expression consumed by user demand processing in the network; Step 4: Based on the resource consumption of the microservice types required for user demand processing, model the resource consumption model of microservice deployment and user demand processing; Step 5: Formulate and define the objective function for optimizing microservice deployment resources to achieve high-quality characterization of heterogeneous constraints. Step 6: According to the different dimensions of the multi-objective optimization function, complete the normalization processing of different dimensions; Step 7: Use the tabu search method to solve the multi-objective optimization problem and achieve effective resource utilization during the deployment process.

[0007] The beneficial effects of this invention include the following: 1) Targeting diverse personalized service requirements and heterogeneous, ubiquitous, and resource-constrained IoT applications, it maximizes the effective utilization of system resources and ensures efficient processing of user requests. 2) By deeply integrating edge computing and microservices concepts, this approach analyzes the resource consumption caused by deploying microservices from the perspective of user request processing latency, taking into account edge server configuration and network bandwidth resources. 3) Considering mixed continuous and discrete high-dimensional variables, it introduces a tabu search-based resource optimization method for adaptive microservices-oriented service placement algorithms with cloud-edge-end collaboration (MSPA), achieving optimal microservice deployment solutions and improving resource utilization. Furthermore, it uses a memory structure to monitor search history, preventing local searches from repeatedly looping back to the same solution and avoiding being trapped in local optima. 4) Simulation results demonstrate that compared to existing conventional methods, this method achieves superior resource consumption and resource utilization performance across various experimental environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic diagram of the microservice mapping relationship targeted by the method of the present invention; Figure 2 It is the dynamic change of the number of user demand chains URC during the implementation of the method of the present invention; Figure 3 The effectiveness of delay guarantee of the method of the present invention is compared with that of three existing methods; Figure 4 It is a comparison of ES resource utilization between the method of the present invention and three existing methods; Figure 5 The comparison of bandwidth resource utilization between the method of the present invention and three existing methods is shown; Figure 6 It is a comparison of ES resource consumption between the method of the present invention and three existing methods; Figure 7 The bandwidth resource consumption comparison between the method of the present invention and three existing methods is shown in FIG. Figure 8 The figure shows the comparison of resource consumption of the number of activated ESs between the method of the present invention and three existing methods. DETAILED DESCRIPTION

[0009] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0010] Aiming at heterogeneous IoT application systems in edge computing environments, this paper designs a microservice deployment service architecture consisting of three layers (microservice layer, server layer, and user layer), such as Figure 1 As shown in the figure, the user layer is the initiator of demand and the driving force of the service architecture. The server layer is responsible for providing various services to meet the diverse needs of users. The server layer is composed of edge servers (ES) and cloud servers (CS). The microservice layer covers a variety of microservices running on CS and ESs (plural of ES).

[0011] Figure 1 Different rectangular bars represent different microservice types, while solid rectangular bars represent microservice types that have been deployed on the ES and may be currently executing service requests. The CS carries various types of microservices, and each ES carries different types and quantities of microservices, and the CS resource allocation is far superior to the ES. Therefore, the present invention assumes that the CS is responsible for all ESs within its communication coverage area and can provide different types of microservice ISO files to the ESs. When the ES receives the requested microservice ISO file, it chooses whether to deploy the corresponding type of microservice based on the current user needs.

[0012] The method of the present invention uses the concept of microservices, which decomposes a large single overall service into multiple independent, fine-grained small microservices, thereby realizing an adaptive service deployment mechanism for diverse user needs. These microservices can not only reduce the resource load of ES, but also improve the service provision performance level. Since ES is subject to its own resource limitations, each ES can only carry some microservices to provide service support. In view of the fact that each ES can provide services to multiple users at the same time in actual applications, the present invention introduces the concept of User Requirement Chain (URC), see Figure 2 , used to represent the demands received by the ES from multiple users at a given moment. Taking into account highly collaborative, integrated applications at the user, server, and microservice levels, and to promote better collaboration among the microservice, server, and user layers, and to achieve optimized, coordinated utilization of heterogeneous IoT application service resources and a service deployment mechanism, this paper designs a microservice provisioning mechanism that encompasses user demand chain service mapping, user demand integration, and the addition, removal, and release of microservices.

[0013] The adaptive microservice deployment resource optimization method for cloud-edge-device collaboration of the present invention is based on the above principles and is implemented according to the following steps: Step 1: Model the system model based on the application requirements of heterogeneous IoT, complete the definition of user requirements and microservice types. The specific process is: Define the parameters of the system model: use Represents the topology of a heterogeneous IoT network, where On behalf of the user, M stands for ES, stands for CS, Represents the physical link between ESs; at any given point in time, each ES may receive requests from multiple users; To simplify the description, define It is the set of URCs received by ES at a certain moment, and defines the five-tuple Represents all URCs related to any ES received in a certain time window, where: and Identify the channels through which users need to access the network ES, which is the network entrance and exit; Represents all user requirements (UR) contained in the URC at a certain moment. Each UR is represented by , Indicates what type of microservice is required for each user's needs; represents the set of links between URs; It refers to the maximum tolerable delay, which is determined by the processing delay of each user demand. In heterogeneous IoT application services with an integrated service architecture of microservices and edge computing, each URC can access the network through a different ES. To provide instant service, the URC often selects the nearest ES as the network access point. Given that user needs need to be processed in real time, this step stipulates that the network access point and exit of each URC must be the same ES when accessing the network.

[0014] The modeling of the system model is now completed.

[0015] Step 2: For the link communication between the user and the ES, a communication model is built. Non-orthogonal multiple access technology is used to establish communication and obtain the transmission rate of the user transmission task. The specific process is as follows: Communication modeling: In order to better simulate the communication connection between users and ES, and taking into account the rational use of channel resources, non-orthogonal multiple access technology is used to establish communication between users and ES. The set of users is represented as , the set of ES is represented as , the set of subcarriers is expressed as ; Let ES and users use a single antenna device, for each ES, the users it serves can share subchannel, and and ; The entire network bandwidth resource is , which is evenly divided into orthogonal multiple access sub-channels and are used by all ESs; any ES and its first The maximum number of users that a subchannel can serve is and ; Any ES The maximum transmission power is ; In order to capture the changes in IoT application requirements and service deployment at different times, the system running time is divided into different time slots ,but Time Slot ES The set of users served by all sub-channels of , express Time slot access ES the number of users; Taking into account the channel fading, Any ES in a time slot and its first Users served on a subchannel The channel coefficient between is defined as: (1) in, is the small-scale Rayleigh fading, It is a large-scale channel fading caused by path loss and shadow fading; exist Time slot, ES In its ES The superimposed signal transmitted by the user served on each sub-channel is expressed as: (2) in, yes Time Slot ES Assigned to Subchannels to users Provide the signals required for service transmission transmission power; is a binary variable reflecting Time Slot ES and users Inter-channel allocation decision; express Time Slot ES Allocate Subchannels for users Provide service support, otherwise ; Inter-signal interference is defined as two types: interval interference and intra-interval interference. From ES No. The signal received by each sub-channel is formulated as: (3) Note: The display of formula (3) includes three kinds of interference.

[0016] in, is a user In ES No. The Gaussian white noise received by each sub-channel satisfies , is the noise power; ES of the time slot No. The power allocated to each sub-channel; ES of the time slot No. transmission signals allocated to sub-channels; In the non-orthogonal multiple access communication mode, before extracting its own information, each user will decode with the user with relatively weaker channel coefficient in the cell; in the process of serial interference cancellation, the intra-cell interference will be treated as noise, and the inter-cell interference will also be eliminated through serial interference; therefore, Time slot users In ES No. The channel gain of the sub-channels needs to satisfy ,in, Indicates that the first decoding user needs to decode the previous user information first, and then decode his own information, then there is Time slot users In ES No. The signal to interference plus noise ratio of each subchannel is: (4) Note: The display of formula (4) includes two kinds of interference.

[0017] Accordingly, Time slot users In ES No. The transmission rate of the sub-channel transmission task is formulated as: (5) Among them, the entire network bandwidth resource is , which is evenly divided into orthogonal multiple access sub-channels and are used by all ESs.

[0018] From then on, the modeling of the communication model in the process of user demand processing is completed.

[0019] Step 3: Model the user demand processing delay module for user demand processing and obtain the total delay expression consumed by user demand processing in the network. The specific process is: Considering the entire process from user demand generation to processing and returning results, the "user demand processing delay module" is divided into four parts: user demand transmission delay, user demand waiting delay, user demand execution delay, and user demand propagation delay; define the triple is the user demand feature, where Indicates the size of user demand, Indicates the amount of computation required by the user, This reflects the maximum latency constraints required by users. The details of these four parts are as follows: a) User demand transmission delay: Time slots, user needs The time cost of transmitting data to the corresponding ES over the network, and the time cost of returning the result to the corresponding user over the network after the ES processes the user's request, are collectively expressed as: , among which Indicates the size of the result after ES executes and processes the user's request; b) User demand waiting delay: Time slot, when the user demand arrives at the corresponding ES, because there are other user demands that have arrived earlier and have not been processed, the user demand needs to queue up for processing. In order to simulate the delay of user demand waiting for processing, the queuing theory is used for analysis; let the arrival rate of user demand be Satisfying the Poisson distribution, the service rate of ES satisfies the parameter The exponential distribution of ES service capacity is ,in represents the number of ES computing resources, and the user demand waiting time is expressed as ; c) User demand execution delay: Time slot, the time cost of user demand processing by ES defines the ES computing capacity as , then the user demand execution delay is: ; d) User demand propagation delay: Time slots, user needs The time cost of the execution result propagating through the communication medium in the network is expressed as ,in Indicates the link length between the user and the corresponding ES, is the propagation speed of the signal in the communication medium between the user and the ES, Indicates the link length for transferring user demand between ESs, is the propagation speed of the signal in the communication medium between ESs, In summary, the overall expression for defining the user demand processing delay is: (6) Since then, the modeling of the "user demand processing delay module" has been completed.

[0020] Step 4: Based on the resource consumption of the microservice types required for user demand processing, model the resource consumption model of microservice deployment resource consumption and user demand processing resource consumption, and obtain the resource consumption expressions of CPU resources, memory resources, bandwidth resources and the number of activated ES resources consumed during the deployment process. Considering the service specificity, that is, each user demand can only be handled by one type of microservice, we define express The time slot user demand is processed by the type of microservice on that ES; at the same time, in order to ensure real-time response to user demand, set The time slot user needs to access the network at any time through the same ES. ; Under normal circumstances, The resources consumed by deploying a type of microservice on a certain ES will not only include the basic memory resources for deploying a certain type of microservice on ES and basic CPU resources , also includes the ES memory resources consumed by starting a microservice to execute a user requirement and CPU resources ; Defining variables express Does the time slot user demand need to be of type If the microservice The time slot user demand needs to be of type Microservices, then ;otherwise ; Defining variables express Is an ES in an idle state in the time slot activated to serve the user's needs? The slot idle state ES is activated, then ;otherwise ; Defining variables express Whether a certain type of microservice is deployed on a certain ES in the time slot. If a specific type of microservice is deployed on the ES, then ;otherwise ; For any user demand, there will be a corresponding physical link from the generation to the result return. The link mapping relationship generated by each user demand in the time slot is defined as If user needs are met by ES If the user demand is successfully received and can be processed, the link mapping relationship is ;otherwise ; Every user requirement in any URC When all user requirements in the URC are mapped to a physical link, there is also a corresponding physical link between them. Define the variable express Physical link mapping between different user requirements in time slots, if different user requirements and Mapping on physical link On, then ;otherwise ; Although ES cannot deploy all microservice types at the same time due to its configuration, it can still deploy a small number of microservices to meet network service needs; in addition, User requirements contained in a URC in a time slot There may be a need for the same microservice or some user needs of the microservice type is deployed on the same ES, define express Some user demands of a certain URC in a time slot are served by the same ES, that is, these user demands are confined to the same ES, in which case there will be no additional link resource overhead; in order to better distinguish Whether the time slot is connected to different or the same ES in the same URC is defined as follows: (7) At the same time, in order to ensure Each user demand in a time slot can be mapped to the corresponding physical link, that is, each user demand can be processed in a timely manner. The following constraints are defined:

[0021] Based on the above analysis of microservice deployment resource consumption, the CPU resource consumption and memory consumption of deploying any microservice in ES and serving user needs can be expressed as follows: (8) At the same time, the expression of link resource consumption required by each user is: (9) in, Defines the URC and Link bandwidth consumption; It represents the bandwidth consumption caused by physical link mapping between users and ES; To better serve diverse user needs, when competition for network resources is intense, it is necessary to activate idle ESs. Therefore, the expression defining the activation consumption of ESs is: (10) in, Is a constant related to system settings.

[0022] Since then, the modeling of microservice deployment resource consumption and user demand processing resource consumption model has been completed.

[0023] Note: The models built in the first four steps are independent of each other, but in the process of dealing with user needs, several models need to go through the process, and each process will affect each other.

[0024] Step 5: Formulate and define the objective function for optimizing microservice deployment resources during user demand processing to achieve high-quality characterization of heterogeneous constraints. Deployment resource optimization problem modeling (here is the modeling of the entire resource optimization objective function).

[0025] The issues that need to be addressed during service deployment are: 1) the resource consumption (CPU and memory) of deploying a certain type of microservice on ES; 2) the resource consumption (CPU and memory) of the microservice deployed on ES to serve a user demand when that demand needs to be processed; 3) the bandwidth resource consumption caused by deploying a certain type of microservice; and 4) the resource consumption of activating an idle ES.

[0026] In view of the above problems and problem modeling description, the microservice deployment resource optimization problem is defined as the following multi-objective optimization problem: (11) The above formula shows that there are four optimization objective functions, namely CPU, memory, link bandwidth, and total resource consumption of activated ES; represents the solution space of the optimization problem, is a feasible solution in the solution space of the optimization problem; Reflects the mapping relationship between user needs, microservices, and ES; C1 indicates that a user need can only be processed by one microservice on ES; C2 reflects user needs Is the type required? Microservices; C3 indicates whether the idle ES is activated to serve user needs; C4 represents whether a certain type is deployed on a certain ES Microservices; C5 indicates that the bandwidth resource consumption caused by mapping user needs on the physical link must not exceed the available bandwidth resource limit of the current link. C6 indicates that the delay cost caused by deploying services to process user needs must not exceed the maximum delay constraint limit of the user needs. ; C7 represents the link resource consumption caused by user demand and the corresponding ES link shall not exceed the maximum bandwidth resource limit between each other ; C8 means that the total CPU resource consumption caused by microservice deployment must not exceed the maximum CPU resource limit available to ES itself ; C9 indicates that the total memory resource consumption caused by microservice deployment must not exceed the maximum memory resource limit available to ES itself .

[0027] This step aims to solve the resource consumption problem caused by the deployment of microservices in heterogeneous IoT applications, that is, to minimize the objective function of formula (11); however, observing formula (11), it can be seen that the formula contains four personalized objectives and the four objective dimensions are different, so it is necessary to introduce a normalization method to further improve formula (11).

[0028] Step 6: For the different dimensions of the multi-objective optimization function, a normalization method is used to complete the normalization of different dimensions. The specific process is: Normalization of multi-objective optimization problems, For the CPU, memory, bandwidth, and total resource consumption of activated ES in the optimization objective function of formula (11), there are dimensional differences and value range differences between the four objectives. In order to solve this difference problem, this step introduces a normalization method for scalarization processing. Therefore, formula (11) is transformed into formula (12), and the expression is: (12) in, It is the parameter of the weighted summation of the normalization method, and also represents the weight of different optimization objectives. The weight value can be randomly adjusted according to different actual application scenarios; Represents the four optimization objectives in the optimization function of formula (11); In addition, in order to ensure that the solution values of the four different optimization objectives are within the same dimensional range, this step uses the maximum-minimum normalization method to scalarize the solution value of each objective in formula (11), that is, normalize the solution value of each objective to between (0, 1), so the four optimization objectives are re-expressed as: , , , , in, , Reflects the i The solution value of the objective function is minimized; similarly, ,in Representative i The solution value that maximizes the objective function.

[0029] The key issue in the maximum-minimum normalization method is determining the maximum and minimum values for each objective in the optimization problem. To accurately evaluate the maximum-minimum values for each optimization objective, this step uses an approximate algorithm to estimate the total resource consumption of CPU, memory, bandwidth, and the total resource consumption of the activated ES. To obtain the maximum resource consumption for each optimization objective, each user request in any URC is distributed among the corresponding microservices on different ESs for processing. As a result, it is observed that link bandwidth resources tend to be maximized while ES memory and CPU resources tend to be minimized. Similarly, each user request in any URC is processed by analyzing the types of microservices running on ESs across the entire network, and the URC is distributed as a whole to the corresponding microservices on an ES for processing. Conclusion: Link bandwidth resource consumption tends to be minimized, while ES memory and CPU resource consumption tend to be maximized.

[0030] Step 7: Use the tabu search method to solve the multi-objective optimization problem in Equation (12) to achieve efficient resource utilization during the deployment process. The specific process is as follows: To further improve the solution quality, a resource optimization method for adaptive microservices-oriented service placement algorithms with cloud-edge-end collaboration (MSPA) is introduced. Taboo search is a meta-heuristic search method widely used to solve combinatorial optimization problems. It has also been proven to be an effective method for solving microservice deployment and scheduling sub-problems. Unlike traditional local search methods (such as hill climbing), the key idea of tabu search is to use a memory structure as a means to monitor the search history to prevent the local search process from repeatedly looping back to the same solution and avoiding getting stuck in the local optimum.

[0031] The specific process of the MSPA algorithm is: First, generate an initial solution , and then the initial solution Set as current solution And the most well-known solution ,Notice, Indicates all Deployment and scheduling solutions for all URs in the cloud; Next, we iterate and generate The list of neighbor solutions (denoted as ) to perform the search process, and then make an action among these neighbor solutions Improvement, if there is no non-improved solution, the best neighbor solution in the list is still accepted to avoid the search process falling into the local optimum; In addition, in each iteration, the action corresponding to the accepted solution is marked as Tabu and stored in the taboo list Tabu, which will prevent a solution marked as taboo from being accepted in a specific number of iterations to prevent the search process from looping back to the previously accepted solution and falling into the local optimum; However, if a solution, its action is in the Tabu list, and the solution is better than Better, it can still be acceptable, a situation known as the inhalation standard; Subsequently, in each iteration, the accepted solution becomes the next iteration's , and update ;use represents the solution accepted in each iteration; Finally, the search process is repeated until There is no improvement in a given number of consecutive iterations.

[0032] During the execution of the MSPA algorithm, there are two key resource optimization scheduling schemes: rescheduling scheme and new microservice deployment scheme. The details are as follows: A) Rescheduling plan: First, start from the current Find the shortest path among the servers of URs in the URs; then, find the best microservice deployment instance from the list of all deployed microservice types. If the best microservice deployment instance leads to the lowest network link traffic scheduling resource consumption and the new microservice deployment instance has better performance than the old microservice deployment instance, the new microservice deployment instance is defined as the best; since the rescheduling of microservice deployment will affect the deployment scheduling of other microservices, all microservices will be rescheduled in a new round. If the new microservice deployment instance meets the UR maximum delay constraint, the UR is received and processed; otherwise, a new round of microservice deployment instances is rearranged; if all microservice deployments have been rescheduled and the UR maximum delay constraint is not met, proceed to the next step of the new microservice deployment plan.

[0033] B) New Microservice Deployment Scheme: This scheme aims to gradually deploy microservice instances required by URs on the network and reschedule them to further improve UR completion time until the UR maximum latency constraint is met. During implementation, new microservice instances are deployed only when necessary to meet the UR maximum latency constraint, minimizing the overhead incurred by newly deployed microservices. This scheme proceeds as follows: First, all URs are sorted in descending order based on their link latency. For each UR, a new microservice is deployed. To locate a new server for the new microservice deployment instance, the shortest path between various servers is determined based on UR communication and latency. The server with the most available resources along this shortest path is then selected as the optimal candidate for microservice hosting. If this server has sufficient resources to host the microservice, the microservice is deployed and the URs for the requested microservice type are checked to see if they meet the maximum latency constraint. If so, the corresponding URs are accepted and the algorithm terminates. Otherwise, the worst server is selected for microservice hosting. Finally, the above process is repeated until the microservice hosted on the server meets the UR's maximum latency constraint.

[0034] Specifically, the MSPA algorithm proposed in this step consists of five parts, which are described in detail as follows: 1) Initial solution: Generate the initial solution based on the greedy heuristic algorithm , which aims to minimize the UR completion time without considering new microservice deployments; 2) Neighborhood Structure: In this MSPA algorithm, two actions are defined to generate neighboring solutions for each iteration. The first is the UR order exchange action, which aims to exchange the scheduling order of URs in S, because the scheduling order between URs has a significant impact on their completion time. The second is the new microservice deployment action, which aims to deploy the corresponding microservice instances required by a specific UR. Relying solely on the deployed microservice instances cannot meet the UR maximum latency constraint. The following explains the process of generating neighboring solutions based on these two actions: One is UR sequence exchange: given the current solution set of the current user demand chain URC , first randomly select The two URs in The order of the two selected URs is swapped in [1]. Finally, two adjacent solutions are generated for this order swap. The first and second solutions are generated without considering the new microservice deployment and rescheduling of all URs. The principle behind these two rescheduling strategies is to ensure that the search process does not miss good adjacent solutions.

[0035] The second is the new microservice deployment: given the current solution First, randomly select URs that do not meet the maximum delay constraint; next, select the microservice with the longest waiting time for virtual link traffic scheduling, and select the server node with the most available resources as the microservice host from the shortest path link corresponding to the currently scheduled microservice. This shortest path is based on the sum of communication and waiting time; finally, similar to the previous step (UR sequential exchange action), two adjacent solutions are generated for this step. For the first solution, all URs are rescheduled without considering the new microservice deployment plan, while for the second solution, all URs are rescheduled without considering the rescheduling plan and the new microservice deployment plan.

[0036] 3) Tabu list: In each iteration, the accepted neighbor solutions corresponding to an action can be marked and stored in the tabu list for a given number of Iteration (i.e., taboo term) to avoid the search process repeatedly visiting the same solution and avoiding falling into local optimality; define a tuple represents the UR sequence exchange action stored in the tabu list, Indicates that two URs exchange order; a tuple is also defined Represents a new microservice deployment action, where and Indicates that it is the first UR deployed a type of Microservices.

[0037] 4) Neighborhood evaluation and selection: In each iteration, the neighborhood solution is evaluated according to the objective function, defined as Equation (12). In the next iteration, the neighbor solution with the best benefit and whose action is not in the taboo list Tabu is always selected. As the new current solution ; However, if the neighbor solution has the best benefit and the action is in the tabu list Tabu Outperforms the best-known solutions , then the neighbor solution is still selected , in order to meet the expected standards.

[0038] 5) Termination criteria: If the best known solution The MSPA algorithm terminates if there is no improvement for a given number of consecutive iterations. It is worth noting that the number of iterations can be adjusted to ensure an appropriate trade-off between execution time and solution quality. The MSPA algorithm repeats the above steps until the optimal solution is generated, that is, the objective function (12) is optimized and the optimal solution under the current environmental state is obtained.

[0039] Experimental verification: As we all know, network load is in a dynamic state over time. In order to simulate network load changes as much as possible and better analyze the performance of the proposed algorithm MSPA, it is assumed that in each time window, for any user, only one user demand can be generated. At the same time, the total number of all user demands that the entire system can generate in each time window is known.

[0040] The change in the number of user demands in each time window can be replaced by the change in the transit network traffic in the TOTEM project. By compressing the time scale, the total change in the number of user demands in different time windows of the entire system is defined as Figure 2 As shown, the requirement chain is the User Requirement Chain (URC), and the number of URs in each URC has a fixed range. For each UR, the CPU cost requirement for deploying its required microservices on the edge ES is 60MHz, and the memory cost requirement is 40MB. The resource consumption of microservices is evenly distributed. Specifically, the CPU consumption is MHz, memory consumption is MB, bandwidth consumption is Mbps. Signal propagation speed between users and ES , and the speed of communication between ES are all known. The delay constraints required by each user The communication delay between any ES and CS is fixed at 30 milliseconds. Since all network resources are considered equally important, the weight setting of formula (12) adopts a uniform value. .

[0041] Three benchmark algorithms were set as control groups to evaluate the performance of the proposed method and its performance in terms of task execution success rate, resource utilization, and resource consumption. The three benchmark algorithms are described as follows: The first is the Random algorithm: This algorithm randomly assigns each UR in the URC to an ES. The ES then decides whether to deploy the corresponding microservice based on the microservice involved in the received user request. If the microservice of the type already exists on the server, the user request is processed directly; otherwise, the required microservice is considered for installation on the current server. The second is the Nearest ES First (NESF) algorithm: The core of this algorithm is that whenever a user request arises and needs to be processed, it is preferentially assigned to the ES closest to the current user for processing. If the nearest ES does not have the corresponding service support, the ES sends a service request to the CS to download and deploy the image file of the specific microservice to meet the service request. The third is the Greedy Matching (MG) algorithm: This algorithm models the user request center and the network structure as two weighted graphs: the physical network graph and the microservice forwarding graph. By calculating the adjacency matrix of these two graphs, a bipartite matching technique is used to optimize the matching.

[0042] Example 1 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with the other three benchmark algorithms (MG, NESF and Random) in terms of the effectiveness of task execution delay guarantee.

[0043] User demand latency performance is a key indicator for evaluating system real-time processing. This Example 1 is used to test the MSPA method of the present invention to deploy microservices to achieve resource optimization while also meeting the constraints of user demand real-time processing. Figure 2The trend of the number of URCs in the network changing over time is shown. It can be seen that the number of URCs first decreases, then increases, and finally decreases again, which shows that the network load is changing dynamically. The change in network load has an important impact on the delay in processing user demands. When the network load is large (the number of URCs is large), it means that under the current system resource status, the number of user demands that need to be processed increases, and the competition for system resources increases, which will cause some user demands to not be processed in time or even packet loss. When the network load is low (the number of URCs is small), it means that under the current system resource status, the number of user demands that need to be processed is small, and the system resources are sufficient to meet the needs of user processing. Figure 3 As shown, no matter how the network load fluctuates, the MSPA method of the present invention can handle more user demands and meet the maximum constraint of user demand delay as much as possible under the current resource conditions of the system. This is because the MSPA method of the present invention not only takes into account the optimal utilization of resources, but also takes into account the service status, that is, it detects in real time whether all services in the network are executing user demands, and releases idle services, which greatly saves ES resource consumption. The MSPA method of the present invention focuses on adjusting service delay from a dynamic aspect, and then optimizes the resource utilization of CPU, memory, link bandwidth and activated ES, avoiding system resource bottlenecks to a great extent.

[0044] Example 2 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with the other three benchmark algorithms (MG, NESF and Random) in terms of the effectiveness of ES resource utilization.

[0045] Resource utilization is the ratio of used resources to total available resources and is also an important indicator for evaluating resource consumption. This Example 2 compares the ES resource utilization of different algorithms to illustrate that the MSPA algorithm of the present invention has better performance.

[0046] Reference Figure 4 , is the ES resource utilization of the MSPA method of the present invention and the other three algorithms. As the network load increases, that is, the number of URCs increases, the ES resource utilization of the four algorithms also increases. This is mainly because as the number of user demands increases, more ES resources need to be consumed to process user demands. In the process of resource optimization, the MSPA method of the present invention uses network load changes as a guide for real-time microservice deployment, while always paying attention to the status of microservices deployed on ES in the system, and releasing and adding microservices according to the network change status. In contrast, the MG algorithm tends to optimize resource utilization based on the load changes of the entire network. When the network load is high, the number of ESs and network connections required to process user requests will increase, which results in relatively low ES resource utilization under the MG algorithm. From Figure 4 It can be seen that when the network load is small, the ES resource utilization of the MSPA method of the present invention can be maintained at 75%; when the network load increases, the ES resource utilization of the MSPA method of the present invention can also be maintained at a level close to 70%.

[0047] Example 3 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with the other three benchmark algorithms (MG, NESF and Random) in terms of the effectiveness of bandwidth resource utilization.

[0048] Resource utilization is the ratio of used resources to total available resources and is also an important indicator for evaluating resource consumption. This Example 3 illustrates the performance of the MSPA method of the present invention by comparing the bandwidth resource utilization of different algorithms.

[0049] Reference Figure 5 , is the bandwidth utilization of the MSPA method of the present invention and the other three algorithms when the network load changes dynamically. As the network load changes, the bandwidth utilization of each algorithm shows different fluctuation trends. Specifically, when the network load is low, the bandwidth utilization of all algorithms is relatively low; and as the network load increases, the bandwidth utilization of each algorithm also increases. Among them, since the MG algorithm comprehensively considers the matching mapping of the entire network, the bandwidth resource utilization of this algorithm is the lowest. Because the NESF algorithm always gives priority to the nearest edge server (ES) to process the URC, it consumes the least bandwidth resources and the link utilization efficiency is relatively low. The MSPA method of the present invention and the NESF algorithm show similar results in terms of bandwidth resource utilization. The bandwidth resource utilization of the MSPA method of the present invention is maintained at approximately 55%, and that of the NESF algorithm is approximately 52.7%.

[0050] Example 4 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with the other three benchmark algorithms (MG, NESF and Random) in terms of the effectiveness of ES resource consumption.

[0051] like Figure 6 As shown in the figure, during microservice deployment, ES resource consumption primarily consists of two components: the resources consumed by deploying a specific microservice on ES and the processing services provided by the microservices deployed on ES for UR. Because the memory and CPU resource consumption for each microservice type deployed on each ES is the same, ES resource consumption primarily depends on the resources consumed by each microservice on ES to process user requests.

[0052] Reference Figure 6, is a comparison of the ES resource consumption performance of the MSPA method of the present invention and three other traditional algorithms. Figure 6 It can be seen that the ES resource consumption performance of the MSPA method of the present invention is the best compared with the other three algorithms. This is mainly because the MSPA method of the present invention can merge user demands with the same microservice type and then deploy microservices during the process of user demand processing and microservice deployment. At the same time, the MSPA method of the present invention can regularly check the execution status of different microservices according to the network load status, that is, by observing the execution status of microservices, release or increase the number of microservice deployments to provide services for user needs as much as possible, so as to reduce ES resource consumption. In contrast, the existing Random algorithm and NESF algorithm do not consider the sharing of services (that is, merging requests for the same service), they only focus on completing service tasks. In addition, the MSPA method of the present invention can also regularly review and release unnecessary microservices according to changes in network load, thereby further saving resources. In contrast, the existing MG algorithm does not regularly check the execution status of microservices in the network, resulting in low resource consumption performance.

[0053] Example 5 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with the other three benchmark algorithms (MG, NESF and Random) in terms of the effectiveness of bandwidth resource consumption.

[0054] Figure 7 This is the impact of the MSPA method of the present invention and the other three algorithms on bandwidth resource consumption, among which the NESF algorithm and the MG algorithm perform best in terms of network bandwidth utilization efficiency. Since the NESF algorithm always gives priority to the nearest ES to process user needs, each user demand only needs to be transmitted the nearest distance to the ES for user demand outlet, thus achieving the lowest network bandwidth consumption. The MG algorithm uses two weighted graphs to map and match network user needs and network resources. At the same time, in each mapping and matching process, the MG algorithm tends to map multiple user needs in the URC to one ES for processing, which also causes multiple user needs in the URC to be concentrated in one ES, reducing the transmission resource consumption of user needs in the network link. The MSPA method of the present invention comprehensively considers ES resource consumption, bandwidth resource consumption and user demand processing delay, and significantly reduces the transmission bandwidth consumption of user needs between different ESs by merging and mapping the same user needs on the same microservice type.

[0055] Example 6 Using the parameter settings verified by the aforementioned experiments, the MSPA method of the present invention is compared with three other traditional algorithms (MG, NESF, and Random) in terms of the effectiveness of resource consumption in activating the number of ESs.

[0056] In order to verify the performance of the MSPA method of the present invention and the other three algorithms, the number of ES activated by different algorithms was analyzed, and the results are shown in the figure. Figure 8 As shown, it can be concluded that compared with the other three algorithms, the MSPA method of the present invention activates the lowest number of ES. This is because the MSPA method of the present invention utilizes the characteristics of service sharing to merge user demands with the same type of microservice demands and map them to a specific microservice type of the same ES, which can reduce the number of ES activations and greatly improve network resource utilization. In addition, the MSPA method of the present invention can also release microservices that are idle and have not processed user demands by detecting the network load status and the microservice execution status to reduce the number of activated ES. Compared with the MSPA method of the present invention, the other three traditional algorithms did not take into account the impact of network load or the microservice execution status, so more ES need to be activated when processing user demands.

Claims

1. An adaptive microservice deployment resource optimization method for cloud-edge-end collaboration, characterized by: The steps are: Step 1: Model the system model and complete the definition of user requirements and microservice types; Step 2: Model the communication model and obtain the transmission rate of the user transmission task; Step 3: Model the user demand processing delay module to obtain the total delay expression consumed by user demand processing in the network; Step 4: Based on the resource consumption of the microservice types required for user demand processing, model the resource consumption model of microservice deployment and user demand processing; Step 5: Formulate and define the objective function for optimizing microservice deployment resources to achieve high-quality characterization of heterogeneous constraints. Step 6: According to the different dimensions of the multi-objective optimization function, complete the normalization processing of different dimensions; Step 7: Use the tabu search method to solve the multi-objective optimization problem and achieve effective resource utilization during the deployment process.

2. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 1, the specific process is: Define the parameters of the system model: use Represents the topology of a heterogeneous IoT network, where On behalf of the user, M stands for ES, stands for CS, Represents the physical link between ESs; at any given point in time, each ES may receive requests from multiple users; definition It is the set of URCs received by ES at a certain moment, and defines the five-tuple Represents all URCs related to any ES received in a certain time window, where: and Identify the channels through which users need to access the network ES, which is the network entrance and exit; Represents all user requirements UR included in the URC at a certain moment, and each UR is expressed as , Indicates what type of microservice is required for each user's needs; represents the set of links between URs; It refers to the maximum tolerable delay; It is stipulated that when users in each URC access the network, their network access entry and exit must be the same ES.

3. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 2, the specific process is: Communication modeling: Non-orthogonal multiple access technology is used to establish communication between users and ESs. The set of users is represented as , the set of ES is represented as , the set of subcarriers is expressed as ; Let ES and users use a single antenna device, for each ES, the users it serves can share subchannel, and and ; The entire network bandwidth resource is , which is evenly divided into orthogonal multiple access sub-channels and are used by all ESs; any ES and its first The maximum number of users that a subchannel can serve is and ; Any ES The maximum transmission power is ; Divide system running time into different time slots ,but Time Slot ES The set of users served by all sub-channels of , express Time slot access ES the number of users; Any ES in a time slot and its first Users served on a subchannel The channel coefficient between is defined as: (1) in, is the small-scale Rayleigh fading, It is a large-scale channel fading caused by path loss and shadow fading; exist Time slot, ES In its ES The superimposed signal transmitted by the user served on each sub-channel is expressed as: (2) in, yes Time Slot ES Assigned to Subchannels to users Provide the signals required for service transmission The transmission power; is a binary variable reflecting Time Slot ES and users Inter-channel allocation decision; express Time Slot ES Allocate Subchannels for users Provide service support, otherwise ; Inter-signal interference is defined as two types: interval interference and intra-interval interference. From ES No. The signal received by each sub-channel is formulated as: (3) in, is a user In ES No. The Gaussian white noise received by each sub-channel satisfies , is the noise power; Is the ES of the time slot No. The power allocated to each sub-channel; Is the ES of the time slot No. transmission signals allocated to sub-channels; Time slot users In ES No. The channel gain of the sub-channels needs to satisfy ,in, Indicates that the first decoding user needs to decode the previous user information first, and then decode his own information, then there is Time slot users In ES No. The signal to interference plus noise ratio of each subchannel is: (4) Accordingly, Time slot users In ES No. The transmission rate of the sub-channel transmission task is formulated as: (5) Among them, the entire network bandwidth resource is , which is evenly divided into orthogonal multiple access sub-channels and are used by all ESs.

4. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 3, the specific process is: The user demand processing delay module is divided into four parts: user demand transmission delay, user demand waiting delay, user demand execution delay, and user demand propagation delay; define the triple is the user demand feature, where Indicates the size of user demand, Indicates the amount of computation required by the user, This reflects the maximum latency constraints required by users. The details of these four parts are as follows: a) User demand transmission delay: Time slots, user needs The time cost of transmitting data to the corresponding ES over the network, and the time cost of returning the result to the corresponding user over the network after the ES processes the user's request, are collectively expressed as: , among which Indicates the size of the result after ES executes and processes the user's request; b) User demand waiting delay: Time slot, when the user demand arrives at the corresponding ES, in order to simulate the delay of user demand waiting for processing, the queuing theory is used for analysis; let the arrival rate of user demand Satisfying the Poisson distribution, the service rate of ES satisfies the parameter The exponential distribution of ES service capacity is ,in represents the number of ES computing resources, and the user demand waiting time is expressed as ; c) User demand execution delay: Time slot, the time cost of user demand processing by ES defines the ES computing capacity as , then the user demand execution delay is: ; d) User demand propagation delay: Time slots, user needs The time cost of the execution result propagating through the communication medium in the network is expressed as ,in Indicates the link length between the user and the corresponding ES, is the propagation speed of the signal in the communication medium between the user and the ES, Indicates the link length for transferring user demand between ESs, is the propagation speed of the signal in the communication medium between ESs, In summary, the overall expression for defining the user demand processing delay is: 。 5. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 4, the specific process is: Each user requirement can only be handled by one type of microservice, so define express The time slot user demand is processed by the type of microservice on that ES; at the same time, set The time slot user needs to access the network at any time through the same ES. ; Under normal circumstances, The resources consumed by deploying a type of microservice on a certain ES will not only include the basic memory resources for deploying a certain type of microservice on ES. and basic CPU resources , also includes the ES memory resources consumed by starting a microservice to execute a user requirement and CPU resources ; Defining variables express Does the time slot user demand need to be of type If the microservice The time slot user demand needs to be of type Microservices, then ;otherwise ; Defining variables express Is an ES in an idle state in the time slot activated to serve the user's needs? The slot idle state ES is activated, then ;otherwise ; Defining variables express Whether a certain type of microservice is deployed on a certain ES in the time slot. If a specific type of microservice is deployed on the ES, then ;otherwise ; For any user demand, there will be a corresponding physical link from the generation to the result return. The link mapping relationship generated by each user demand in the time slot is defined as If user needs are met by ES If the user demand is successfully received and can be processed, the link mapping relationship is ;otherwise ; Every user requirement in any URC When all user requirements in the URC are mapped to a physical link, there is also a corresponding physical link between them. Define the variable express Physical link mapping between different user requirements in time slots, if different user requirements and Mapping on physical link On, then ;otherwise ; User requirements contained in a URC in a time slot There may be a need for the same microservice or some user needs of the microservice type is deployed on the same ES, define express Some user demands of a certain URC in a time slot are served by the same ES, that is, these user demands are restricted to the same ES; in order to better distinguish Whether the time slot is connected to different or the same ES in the same URC is defined as follows: (7) At the same time, in order to ensure Each user demand in a time slot can be mapped to the corresponding physical link, that is, each user demand can be processed in a timely manner. The following constraints are defined: Based on the above analysis of microservice deployment resource consumption, the CPU resource consumption and memory consumption of deploying any microservice in ES and serving user needs can be expressed as follows: (8) At the same time, the expression of link resource consumption required by each user is: (9) in, Defines the URC and Link bandwidth consumption; It represents the bandwidth consumption caused by physical link mapping between users and ES; When competition for network resources is intense, it is necessary to start the idle ES. Therefore, the expression defining the consumption of ES activation is: (10) in, Is a constant related to system settings.

6. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 5, the specific process is: Deployment resource optimization problem modeling, the microservice deployment resource optimization problem is defined as the following multi-objective optimization problem: (11) The above formula shows that there are four optimization objective functions, namely CPU, memory, link bandwidth, and total resource consumption of activated ES; represents the solution space of the optimization problem, is a feasible solution in the solution space of the optimization problem; Reflects the mapping relationship between user needs, microservices, and ES; C1 indicates that a user need can only be processed by one microservice on ES; C2 reflects user needs Is the type required? Microservices; C3 indicates whether the idle ES is activated to serve user needs; C4 indicates whether a certain type of service is deployed on a certain ES. Microservices; C5 indicates that the bandwidth resource consumption caused by mapping user needs on the physical link must not exceed the available bandwidth resource limit of the current link. C6 states that the delay cost caused by deploying services to process user needs must not exceed the maximum delay constraint limit of the user needs. ; C7 represents the link resource consumption caused by user demand and the corresponding ES link shall not exceed the maximum bandwidth resource limit between each other ; C8 means that the total CPU resource consumption caused by microservice deployment must not exceed the maximum CPU resource limit available to ES itself ; C9 indicates that the total memory resource consumption caused by microservice deployment must not exceed the maximum memory resource limit available to ES itself .

7. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 6, the specific process is: Introducing the normalization method for scalarization processing, transforming formula (11) into formula (12), the expression is: (12) in, It is the parameter of the weighted summation of the normalization method, and also represents the weight of different optimization objectives; Represents the four optimization objectives in the optimization function of formula (11); The maximum-minimum normalization method is used to scalarize the solution value of each objective in formula (11), that is, the solution value of each objective is normalized to between (0, 1), so the four optimization objectives are re-expressed as: ， ， ， ， in, , Reflects the i The solution value of the objective function is minimized; similarly, ,in Representative i The solution value that maximizes the objective function.

8. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 1 is characterized in that: In step 7, the taboo search method, namely the MSPA algorithm, is used. The specific process is: First, generate an initial solution , and then the initial solution Set as current solution And the most well-known solution ,Notice, Indicates all Deployment and scheduling solutions for all URs in the cloud; Next, we iterate and generate List of neighbor solutions to perform the search process, and then make an action among these neighbor solutions Improvement, if there is no non-improved solution, the best neighbor solution in the list is still accepted to avoid the search process falling into the local optimum; In addition, in each iteration, the action corresponding to the accepted solution is marked as Tabu and stored in the taboo list Tabu to prevent the search process from looping back to the previously accepted solution and falling into the local optimum; However, if a solution whose action is in the Tabu list and the solution is better than Better, it can still be acceptable, a situation known as the inhalation standard; Subsequently, in each iteration, the accepted solution becomes the next iteration's , and update ;use represents the solution accepted in each iteration; Finally, the search process is repeated until There is no improvement in a given number of consecutive iterations.

9. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 8 is characterized in that: During the execution of the MSPA algorithm, there are two key resource optimization scheduling schemes: rescheduling scheme and new microservice deployment scheme. The details are as follows: A) Rescheduling plan: First, from the current Find the shortest path among the servers of URs; Then, find the best microservice deployment instance from the list of all deployed microservice types. If the best microservice deployment instance results in the lowest network link traffic scheduling resource consumption and the new microservice deployment instance has better performance than the old microservice deployment instance, define the new microservice deployment instance as the best. Since the rescheduling of microservice deployment will affect the deployment scheduling of other microservices, all microservices are rescheduled in a new round. If the new microservice deployment instance meets the UR maximum delay constraint, the UR is received and processed; Otherwise, reschedule a new round of microservice deployment instances; If all microservice deployments have been rescheduled and the UR maximum latency constraint is not met, proceed to the next step of deploying a new microservice. B) New microservice deployment solution: First, all URs are sorted in descending order based on their waiting time in the link. For each UR, a new microservice is deployed. To find a new server for the new microservice deployment instance, the shortest path between the various servers for the microservice deployment is found based on the UR communication time and waiting time. Then, the server with the most available resources on the shortest path is selected as the best choice for microservice deployment. If the server has enough resources to deploy the microservice, the microservice is deployed and the URs of the required microservice of this type are checked to see if they meet the maximum delay constraint. If so, the corresponding URs are received and the algorithm stops. Otherwise, the worst server is selected for microservice deployment. Finally, the above process is repeated until the microservices on the server can meet the maximum latency constraint of URs.

10. The adaptive microservice deployment resource optimization method for cloud-edge-end collaboration according to claim 8 is characterized in that: The MSPA algorithm consists of five parts, which are described as follows: 1) Initial solution: Generate the initial solution based on the greedy heuristic algorithm , which aims to minimize the UR completion time without considering new microservice deployments; 2) Neighborhood Structure: In this MSPA algorithm, two actions are defined to generate neighboring solutions for each iteration. The first is the UR order exchange action, which aims to exchange the scheduling order of URs in S, because the scheduling order between URs has a significant impact on their completion time. The second is the new microservice deployment action, which aims to deploy the corresponding microservice instances required by a specific UR. Relying solely on the deployed microservice instances cannot meet the UR maximum latency constraint. The following explains the process of generating neighboring solutions based on these two actions: One is UR sequence exchange: given the current solution set of the current user demand chain URC , first randomly select The two URs in The order of the two selected URs is swapped in [15]; finally, two adjacent solutions are generated for this order swapping action; The second is the new microservice deployment: given the current solution First, randomly select URs that do not meet the maximum delay constraint; next, select the microservice with the longest virtual link traffic scheduling total agent waiting time, and select the server node with the most available resources as the microservice host from the shortest path link corresponding to the currently scheduled microservice. This shortest path is based on the sum of communication and waiting time; finally, similar to the previous step, two adjacent solutions are generated for this step. For the first solution, all URs are rescheduled without considering the new microservice deployment plan, while for the second solution, all URs are rescheduled without considering the rescheduling plan and the new microservice deployment plan; 3) Taboo list: In each iteration, the accepted neighbor solutions corresponding to an action can be marked and stored in the taboo list for a given number of Iteration to avoid the search process repeatedly visiting the same solution and avoid falling into local optimality; define a tuple represents the UR sequence exchange action stored in the tabu list, Indicates that two URs exchange order; a tuple is also defined Represents a new microservice deployment action, where and Indicates that it is the first UR deployed a type of Microservices; 4) Neighborhood evaluation and selection: In each iteration, the neighborhood solution is evaluated according to the objective function, defined as Equation (12). In the next iteration, the neighbor solution with the best benefit and whose action is not in the taboo list Tabu is always selected. As the new current solution ; However, if the neighbor solution has the best benefit and the action is in the tabu list Tabu Outperforms the best-known solutions , then the neighbor solution is still selected , so as to meet the expected standards; 5) Termination criteria: If the best known solution The MSPA algorithm terminates if there is no improvement for a given number of consecutive iterations.

Images