A high reliability personalized deployment method and system for maximizing service quality in SMEC environment

Abstract

The present invention discloses a high-reliability personalized deployment method and system for maximizing service quality in an SMEC environment, including: first, generating an estimated optional execution ratio; then, allocating optimal processor resources to tasks; then, optimizing processor resource allocation and task execution strategy based on the cross entropy (CE) method; then, adjusting the execution ratio and generating a new deployment plan according to real-time feedback; then, dynamically adjusting the service quality parameters of the Internet of Things application; then judging whether the performance of the new plan is better than the previous plan, if satisfied, continuing to execute; if not satisfied, rolling back to the previous deployment plan. The present invention can optimize the service quality in real time and maximize the system performance according to dynamically changing application requirements and environmental conditions while ensuring system reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things (IoT) application deployment, and in particular to a high-reliability personalized deployment method and system for maximizing service quality in an SMEC environment. Background Art

[0002] With the rapid growth of the number of Internet of Things (IoT) devices, edge computing (MEC) has become an important technology to support low-latency IoT applications. MEC improves the flexibility and responsiveness of the system by pushing computing tasks from centralized cloud servers to the edge of distributed networks. However, the traditional MEC architecture has problems such as uneven resource allocation and insufficient reliability, making it difficult to meet the dynamically changing needs of IoT applications.

[0003] To address these challenges, researchers proposed the concept of serverless computing (SC) and combined it with MEC to form a server-aware edge computing (SMEC) framework. The SMEC framework greatly improves the scalability and deployment efficiency of the system by organizing applications into directed acyclic graphs (DAGs) or service function chains.

[0004] However, most existing SMEC research focuses on optimizing a single performance metric, such as latency, energy consumption, or cost, but ignores key factors such as system reliability and quality of service (QoS). As processor feature sizes and operating voltages continue to shrink, the incidence of soft errors has increased significantly. In addition, bit errors on communication links are also inevitable, which can lead to system failure and data loss.

[0005] Deng et al. studied a latency-aware combination of function placement and flow mapping, while Angelelli et al. developed a multi-objective function scheduling strategy. However, these studies mainly focused on static function deployment. Pelle et al. proposed a framework for dynamic configuration, and Zhang et al. developed a feedback control scheme to deploy serverless functions. Although these methods optimize system performance to a certain extent, they usually only focus on a single optimization objective and fail to fully consider the dynamic requirements of IoT applications and the overall service quality of the system. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a high-reliability personalized deployment method and system for maximizing service quality in a SMEC environment. The method combines offline optimization and online optimization, and dynamically adjusts the task execution ratio and deployment scheme under the premise of satisfying system energy, delay and reliability constraints, thereby maximizing the overall service quality of IoT applications.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A high reliability personalized deployment method for maximizing service quality in an SMEC environment includes the following steps:

[0009] Step 1: Get the task set of the server system;

[0010] Step 2: generating an estimated optional execution ratio based on the task set of the server system;

[0011] Step 3: Determine the initial deployment plan for processor resource allocation and task execution based on the cross entropy method;

[0012] Step 4: Based on the estimated optional execution ratio and the initial deployment plan, adjust the execution ratio and generate a new deployment plan;

[0013] Step 5: Based on the new deployment plan, the service quality parameters of the IoT application are dynamically adjusted by using proximal policy optimization to obtain an updated execution ratio and deployment plan.

[0014] Preferably, in step 2, the method for generating an estimated optional execution ratio based on the task set of the server system includes:

[0015] Calculate the energy requirement of each task based on the energy consumption coefficient and the assumed execution ratio;

[0016] Based on the energy requirements of each task, calculate the overall energy requirement;

[0017] Calculate total energy supply;

[0018] comparing the total energy demand to the total energy supply;

[0019] If the total energy demand is greater than the total energy supply, the optimal optional execution ratio is calculated.

[0020] Preferably, in step 3, the method for determining processor resource allocation and task execution based on the cross entropy method includes:

[0021] Set iteration counter ;

[0022] Set the initial probability vector ;

[0023] Based on the probability vector generate Possible samples ;

[0024] For each sample Conduct performance evaluations;

[0025] Will The samples are sorted in descending order by performance evaluation value, and selected An elite sample;

[0026] based on The average performance evaluation value of elite samples is used to calculate the new performance threshold ;

[0027] Based on the new performance threshold , update the probability vector ;

[0028] Based on the new probability vector , select the best performing sample;

[0029] Based on the best performing samples, an initial deployment plan for processor resource allocation and task execution is determined.

[0030] Preferably, in step 4, based on the estimated optional execution ratio and the initial deployment plan, the method of adjusting the execution ratio and generating a new deployment plan includes:

[0031] Calculate initial control variables ;

[0032] Sort all apps in descending order and select the ones that meet the preset requirements;

[0033] Based on the initial control variables and the application that meets the preset requirements, set the control variables ;

[0034] If the current control variable ,and ,calculate , get the execution ratio increment of each task;

[0035] Based on the execution ratio increment of each task, update , regenerate the deployment plan .

[0036] Preferably, in step 5, based on the new deployment scheme, the method of dynamically adjusting the service quality parameters of the Internet of Things application by using proximal policy optimization includes:

[0037] Use the Xavier method to initialize the strategy and evaluation network, and set the initialization control variables And build an adjustable application set for the current time slice and fixed application sets ;

[0038] Calculate the constraint violation degree of the current time slice ;

[0039] Based on the constraint violation level of the current time slice , calculate the reward ;

[0040] Adjustable application set based on current time slice , Fixed application set and constraint violation , update status to ;

[0041] Based on the updated status, determine whether there is a new sliding window;

[0042] If there is a new sliding window, update the policy network parameters and evaluation network parameters;

[0043] Based on updating the physical strategy network parameters and evaluating the network parameters, the service quality parameters of the IoT application are dynamically adjusted to obtain the updated execution ratio and deployment plan.

[0044] The present invention also provides a high-reliability personalized deployment system for maximizing service quality in a SMEC environment, comprising the following steps: a task acquisition module, a ratio generation module, a scheme initialization module, a ratio adjustment module, and a scheme optimization module;

[0045] The task acquisition module is used to acquire the task set of the server system;

[0046] The ratio generation module is used to generate an estimated optional execution ratio based on the task set of the server system;

[0047] The scheme initialization module is used to determine the initial deployment scheme of processor resource allocation and task execution based on the cross entropy method;

[0048] The ratio adjustment module is used to adjust the execution ratio and generate a new deployment plan based on the estimated optional execution ratio and the initial deployment plan;

[0049] The solution optimization module is used to dynamically adjust the service quality parameters of the Internet of Things application based on the new deployment solution and utilize proximal policy optimization to obtain an updated execution ratio and deployment solution.

[0050] Preferably, the ratio generation module includes: a first calculation unit, a second calculation unit, a third calculation unit, a comparison unit and a fourth calculation unit;

[0051] The first calculation unit is used to calculate the energy requirement of each task according to the energy consumption coefficient and the assumed execution ratio;

[0052] The second calculation unit is used to calculate the overall energy demand based on the energy demand of each task;

[0053] The third calculation unit is used to calculate the total energy supply;

[0054] The comparison unit is used to compare the total energy demand with the total energy supply;

[0055] The fourth calculation unit is used to calculate the optimal optional execution ratio if the total energy demand is greater than the total energy supply.

[0056] Preferably, the scheme initialization module comprises: a first setting unit, a second setting unit, a sample generation unit, a performance evaluation unit, a sample sorting unit, a threshold calculation unit, a vector update unit, a sample selection unit and an initialization unit;

[0057] The first setting unit is used to set the iteration counter ;

[0058] The second setting unit is used to set the initial probability vector ;

[0059] The sample generation unit is used to generate a sample based on a probability vector generate Possible samples ;

[0060] The performance evaluation unit is used to evaluate each sample Conduct performance evaluations;

[0061] The sample sorting unit is used to The samples are sorted in descending order by performance evaluation value, and selected An elite sample;

[0062] The threshold calculation unit is used to The average performance evaluation value of elite samples is used to calculate the new performance threshold ;

[0063] The vector updating unit is configured to update the vector based on the new performance threshold , update the probability vector ;

[0064] The sample selection unit is used to select the sample based on the new probability vector , select the best performing sample;

[0065] The initialization unit is used to determine an initial deployment plan for processor resource allocation and task execution based on the best performing sample.

[0066] Preferably, the ratio adjustment module includes: a variable calculation unit, an application sorting unit, a variable updating unit, a ratio increment unit and a scheme updating unit;

[0067] The variable calculation unit is used to calculate the initial control variable ;

[0068] The application sorting unit is used to sort all applications in descending order and select applications that meet preset requirements;

[0069] The variable updating unit is used to update the initial control variable based on the initial control variable. and the application that meets the preset requirements, set the control variables ;

[0070] The proportional increment unit is used to ,and ,calculate , get the execution ratio increment of each task;

[0071] The solution updating unit is used to update the execution ratio increment of each task based on the execution ratio increment of each task. , regenerate the deployment plan .

[0072] Preferably, the solution optimization module includes: an application set construction unit, a violation degree calculation unit, a reward calculation unit, a state update unit, a sliding window update unit, a parameter update unit and a dynamic adjustment unit;

[0073] The application set building unit is used to initialize the strategy and evaluation network using the Xavier method, and to set the initialization control variables And build an adjustable application set for the current time slice and fixed application sets ;

[0074] The violation degree calculation unit is used to calculate the constraint violation degree of the current time slice. ;

[0075] The reward calculation unit is used to calculate the constraint violation degree based on the current time slice. , calculate the reward ;

[0076] The state update unit is used to adjust the application set based on the current time slice. , Fixed application set and constraint violation , the updated status is ;

[0077] The sliding window updating unit is used to determine whether there is a new sliding window based on the updated state;

[0078] The parameter updating unit is used to update the policy network parameters and the evaluation network parameters if there is a new sliding window;

[0079] The dynamic adjustment unit is used to dynamically adjust the service quality parameters of the Internet of Things application based on updating the physical strategy network parameters and evaluating the network parameters to obtain the updated execution ratio and deployment plan.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] The present invention improves the overall service quality of IoT applications in SMEC environments through a hybrid method that combines offline optimization and online optimization. The offline phase adopts a function placement strategy based on convex optimization and cross entropy methods, while the online phase uses a reinforcement learning scheme based on neighbor policy optimization (PPO) to dynamically adjust the task execution ratio and deployment scheme to adapt to the dynamic needs of IoT applications. The present invention maximizes the service quality while improving system reliability, and has important application value for IoT applications and embedded real-time systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0083] Figure 1 A flow chart of a high-reliability personalized deployment method for maximizing service quality in a SMEC environment according to the present invention;

[0084] Figure 2 This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set including 1000 server systems in an offline state;

[0085] Figure 3 This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set including 1500 server systems in an offline state;

[0086] Figure 4 This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set including 2000 server systems in an offline state;

[0087] Figure 5 This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set containing 1000 server systems in an online state;

[0088] Figure 6This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set including 1500 server systems in an online state;

[0089] Figure 7 This is a comparison chart of the algorithm proposed by the present invention and the benchmark algorithm in improving the overall service quality under the condition of a task set containing 2000 server systems in an online state. DETAILED DESCRIPTION

[0090] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0091] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0092] Embodiment 1

[0093] The present invention discloses a high-reliability personalized deployment method for maximizing service quality in a SMEC environment, that is, a method for high-reliability personalized function deployment and approximate computing IoT application in a SMEC environment for maximizing service quality, comprising the following steps:

[0094] Step 1: Read the application set and SMEC system;

[0095] Step 2: Generate an estimated alternative execution ratio;

[0096] Step 3: Determine processor resource allocation and task execution based on the cross entropy (CE) method;

[0097] Step 4: Adjust the execution ratio and generate a new deployment plan;

[0098] Step 5: Dynamically adjust the service quality parameters of IoT applications based on proximal policy optimization (PPO);

[0099] Step 6: Evaluate new options;

[0100] Step 7: Return to the final solution.

[0101] In this embodiment, the optional execution ratio of generating the estimate in step 2 specifically includes:

[0102] Step A1: Calculate overall energy requirements , using the formula .in, is the average energy consumption per instruction, It's a task The number of subtasks, is the instruction cycle that must be executed, is an optional execution ratio, It is an optional instruction cycle;

[0103] Step A2: Calculate total energy supply , the formula is: ;

[0104] Step A3: Compare energy demand with energy supply and determine Is it less than or equal to , if the conditions are met, execute step A4, otherwise execute step A9;

[0105] Step A4: Assign a value of 1, that is: ;

[0106] Step A5: Settings ;

[0107] Step A6: Update , the formula is: ;

[0108] Step A7: Judgment Is it true? If so, go to step A4; otherwise, go to step A8;

[0109] Step A8: Assign a value of 1, that is: ;

[0110] Step A9: Partially execute tasks and calculate the optimal optional execution ratio , the formula is:

[0111] .in, is the priority weight, is the Lagrange multiplier, is the Lagrange multiplier used to represent the adjustment of the energy constraint, is the Lagrange multiplier used to represent the adjustment amount of the priority constraint;

[0112] Step A10: Set to ,Right now ;

[0113] Step A11: ;

[0114] Step A12: Determination Is it true? If so, go to step 10; otherwise, go to step 13;

[0115] Step A13: Return execution ratio ;

[0116] Step A14: Determine the end of the optional execution ratio and exit.

[0117] In this embodiment, step 3 determines processor resource allocation and task execution based on a cross entropy (CE) method, specifically including:

[0118] Step B1: Set the iteration counter ;

[0119] Step B2: Set the initial probability vector ;

[0120] Step B3: Based on the probability vector generate Possible samples ;

[0121] Step B4: Determine whether the stop condition is met. If so, go to step B5; otherwise, go to step B3;

[0122] Step B5: Assign a value of 1, that is: ;

[0123] Step B6: Inference Sample Performance ;

[0124] Step B7: Update , the formula is: ;

[0125] Step B8: Judgment Is it true? If so, go to step B6; otherwise, go to step B9;

[0126] Step B9: Samples by performance Sort in descending order;

[0127] Step B10: Select elite samples and mark their indexes as ;

[0128] Step B11: Derive threshold variables ;

[0129] Step B12: Assign a value of 1, that is ;

[0130] Step B13: Assign a value of 1, that is ;

[0131] Step B14: Assign a value of 1, that is ;

[0132] Step B15: Derive the assignment probability , the formula is: ,in, The task is assigned to the edge server The state variable of IoT applications The first task is assigned to the edge server ,but , otherwise 0. Index of IoT applications, For IoT applications Function index in , is the index of the edge server, is the indicator function, when the sample Performance Less than or equal to threshold hour, , otherwise 0; is the indicator function, when the tasks in the sample are assigned to the server hour, , otherwise 0;

[0133] Step B16: Update , the formula is: ;

[0134] Step B17: Judgment Is it true? If so, go to step B9; otherwise, go to step B12;

[0135] Step B18: Update , the formula is: ;

[0136] Step B19: Judgment Is it true? If so, go to step B7; otherwise, go to step B14;

[0137] Step B20: Update , the formula is: ;

[0138] Step B21: Determination Is it true? If so, go to step B6; otherwise, go to step B16;

[0139] Step B22: Update iteration counter ;

[0140] Step B23: Select the best performing sample and set it as ;

[0141] Step B24: Return to the function deployment plan ;

[0142] Step B25: Determine that processor resource allocation and task execution are complete, and exit.

[0143] In this embodiment, step 4 adjusts the execution ratio and generates a new deployment plan, specifically including:

[0144] Step C1: Call step 2 to generate a set of estimated optional execution ratios ;

[0145] Step C2: Call step 3 to generate a deployment plan ;

[0146] Step C3: Calculate initial control variables , the formula is:

[0147] ,in, , represents the control coefficient proportional to the current error, represents the control coefficient proportional to the accumulated error, represents the control coefficient proportional to the error change rate, Indicates The energy error of the iteration is Indicates the length of the time window when calculating the cumulative error, Indicates the length of the time window when calculating the error change rate;

[0148] Step C4: Arrange all applications in descending order;

[0149] Step C5: Assigned to ,Right now ;

[0150] Step C6: Set control variables ;

[0151] Step C7: Determine the current control variable Is it true? If so, go to step C8; otherwise, go to step C15;

[0152] Step C8: If , then execute step C9; otherwise, go to step C11;

[0153] Step C9: Calculation , the formula is: ,in, is the priority weight of application h;

[0154] Step C10: Update , go to step C13;

[0155] Step C11: Calculation , the formula is: ;

[0156] Step C12: Update ;

[0157] Step C13: Regenerate the solution by calling step 3 ;

[0158] Step C14: Update the control variable using the formula , go to step C7;

[0159] Step C15: Update , the formula is: ;

[0160] Step C16: Judgment Is it true? If so, go to step C6; otherwise, go to step C17;

[0161] Step C17: Judgment Is it true? If so, go to step C18; otherwise, go to step C19;

[0162] Step C18: Force rollback of deployment plan and terminate;

[0163] Step C19: Return and deployment plans .

[0164] In this embodiment, step 5 dynamically adjusts the service quality parameters of the IoT application based on proximal policy optimization (PPO), specifically including:

[0165] Step D1: Initialize the policy and critic networks using the Xavier method;

[0166] Step D2: Assign a value of 1, that is ;

[0167] Step D3: Build an adjustable application set for the current time slice and fixed application sets ;

[0168] Step D4: and Execute actions on , using the policy function ;

[0169] Step D5: Calculate the constraint violation degree of the current time slice ;

[0170] Step D6: Calculate the reward , using the formula .in, is the deadline constraint violation degree, is the degree of violation of reliability constraints, is the degree of energy constraint violation;

[0171] Step D7: Update status ;

[0172] Step D8: Determine whether there is a new sliding window. If yes, go to step D9; otherwise, go to step D15;

[0173] Step D9: Define policy network parameters The objective function , the formula is: ,in, For status, For action, is the policy function, is the old policy function, For a specific function, is the clipping function, is the clipping interval;

[0174] Step D10: Define evaluation network parameters The objective function , the formula is: ,in, For reward, is the discount factor, Status The value function under ;

[0175] Step D11: Calculate policy network parameters The gradient is: ;

[0176] Step D12: Update the policy network parameters using the gradient ascent method based on the calculated gradient : ,in, is the learning rate;

[0177] Step D13: Calculate and evaluate network parameters The gradient is:

[0178] ;

[0179] Step D14: Based on the calculated gradient, use the gradient descent method to update the evaluation network parameters : , go to step D16;

[0180] Step D15: Update , the formula is: , go to step D3;

[0181] Step D16: Return the updated optional execution ratio ;

[0182] Step D17: Dynamically adjust the service quality parameters of the IoT application based on proximal policy optimization (PPO) and exit.

[0183] In this embodiment, step 6 of evaluating the new solution specifically includes:

[0184] Step E1: Calculate the performance of the new solution , using the formula ;

[0185] Step E2: Calculate the performance of the previous solution , using the formula ;

[0186] Step E3: Judgment Is it true? If so, return to the new solution; otherwise, return to the previous solution;

[0187] Step E4: Evaluation of new plan completed, exit.

[0188] In this embodiment, step 7 returns the final solution, which specifically includes:

[0189] Step F1: Return the final optional execution ratio and deployment plans ;

[0190] Step F2: Return to the final solution and exit.

[0191] Embodiment 2

[0192] In conjunction with the following specific embodiments and drawings, the personalized function deployment and reliability awareness technology of server-aware edge computing (SMEC) in the Internet of Things (IoT) application in the present invention is further described in detail. The implementation process, conditions, and construction of the simulation platform of the methods and technologies in the present invention, except for the contents specifically mentioned below, are all common knowledge in the field of edge computing and serverless computing.

[0193] In the SMEC system architecture used in the present invention, each edge server is configured with highly scalable hardware resources and supports dynamic virtualization technology. ,in, Indicates The computing power of each server is determined by the number of CPU cores and frequency, and these servers are capable of processing real-time data from IoT devices.

[0194] In the IoT application model adopted in the present invention, the application set is composed of It consists of personalized IoT applications with dependent functions, namely All applications use a directed acyclic graph (DAG) to describe their functions and dependencies. At the same time, these applications are independent of each other and do not share data, ensuring data isolation and security during execution.

[0195] In the reliability model implemented in the present invention, a new evaluation mechanism is introduced for soft errors and bit errors in edge computing environments. The system dynamically adjusts the task deployment strategy by monitoring the status of edge servers and the quality of communication links in real time to prevent these errors from affecting the reliability of task execution. Time Edge Server The reliability of , the calculation formula is:

[0196] ;

[0197] in, Representative Server The failure arrival rate.

[0198] In the system model, the computing power of the edge server Expressed as the maximum workload that the server can support to ensure that its processing capacity is not exceeded. , which defines its demand for resources , which includes computing and communication resources: ;

[0199] here, and Represents tasks The basic and optional execution cycles, It determines the proportion of optional execution parts of the application that are activated during actual operation.

[0200] The goal of this invention is to optimize the overall quality of service (QoS) of IoT applications by intelligently selecting suitable edge servers and task scheduling strategies, and adopting advanced fault-tolerance mechanisms. In this way, it aims to minimize the overall operating cost and energy consumption of the system under the constraints of ensuring task reliability and response time.

[0201] The objective function of the present invention can be expressed as:

[0202] ;

[0203] ;

[0204] ;

[0205] ;

[0206] ;

[0207] in, It is Quality of Service (QoS) for IoT applications, It is IoT Applications The execution end time of a task, From the function arrive The communication time, It is IoT Applications The execution start time of a task, It is an application The completion time of all exit functions in It is The deadline for IoT applications, It is IoT Applications functions, is a function The deployment location, is a function The parent task collection of is the deployment location of the parent task collection, It is Reliability goals for IoT applications, It is an edge server energy requirements, It is an edge server In time interval Internal energy supply.

[0208] like Figure 1 As shown, in this embodiment, the task set is defined as The system has a total of 100 samples. The basic execution cycles of the four tasks in sample 1 are , , , cycle, the optional execution cycles are , , , The activity factors are 0.5, 0.5, 0.5, and 0.5. The system consists of 50 Dell PowerEdge XE8545 servers, each with a power supply of 50 Joules. The total power supply of the servers is Joules. Energy consumption coefficient , execution ratio .

[0209] Step 1: Generate an estimated optional execution scale:

[0210] According to the energy consumption coefficient and the assumed execution ratio , first calculate the energy requirement of each task. Taking sample 1 as an example, for task , basic execution cycle Cycle, optional execution cycle cycle. The energy requirement is: Similarly, calculate the energy requirements of the other three tasks: The total energy requirement is: Exceeds the total energy supply of the system Total energy supply , needs to be evenly distributed to each task, considering that the energy required for each task is different, the execution ratio The calculation method is: Substituting the values ​​into the formula, we get: The final execution ratio of sample 1 was determined to be 0.833.

[0211] Step 2: Determine processor resource allocation and task execution based on the cross entropy (CE) method:

[0212] Initialize the iteration counter . Set the initial probability vector to:

[0213] ;

[0214] Based on the initial probability vector , the system generates samples. Each sample represents a task allocation scheme. Each task is randomly assigned to one of the 50 servers according to the probability vector. Next, the system uses the overall QoS to allocate each sample. To evaluate performance. Use the formula For sample 1, the task :

[0215]

[0216]

[0217] ;

[0218] Empathy Task :

[0219] ;

[0220] Task :

[0221]

[0222] ;

[0223] Task :

[0224]

[0225] ;

[0226] The overall QoS is the sum of the QoS of all tasks:

[0227] ;

[0228] The overall QoS of sample 1 is determined to be 2.634.

[0229] For each generated sample, the system sorts it based on the performance evaluation value of the overall quality of service (QoS). After sorting, the system selects the top The best performing elite samples are used for further optimization. , that is, from The top 20 samples are selected as elite samples. Based on the selected elite samples, the system calculates the average performance evaluation value of these elite samples as the performance threshold of this iteration. Using the formula Calculate performance thresholds. For elite samples The overall QoS of the elite sample is calculated as , then the performance threshold of this iteration is: Performance evaluation results based on elite samples and calculated performance thresholds , using the formula For the probability vector to update.

[0230] When the maximum number of iterations is reached or the performance evaluation value of the elite sample has not been significantly improved in several consecutive iterations, the iteration is stopped and the sample with the best overall QoS is selected as the final task allocation solution. The deployment case of sample 1 is as follows: Task Deployed on server 1, the quota is adjusted to 0.85, the original quota is 0.8; Task Deployed on server 2, the quota is adjusted to 0.75, the original quota is 0.7; Task Deployed on server 3, the quota remains at 0.85 and is not adjusted; Task Deployed on server 4, the quota is adjusted to 0.80 and the original quota is 0.75.

[0231] Step 3: Adjust the execution ratio and generate a new deployment plan:

[0232] Calculate initial control variables , the formula is: .set up , the current error The errors in the cumulative error window are , we can calculate: . Arrange all applications in descending order and select the application that meets the preset requirements. Based on the initial control variables and applications that meet preset requirements, set control variables If the current control variable and ,in, ,but This condition is met. Calculation , the formula is: .set up , the total priority weight is , calculate the execution ratio increment of each task: Similarly, we can calculate Update based on the execution ratio increment of each task , and regenerate the deployment plan .

[0233] Step 4: Dynamically adjust the quality of service parameters of IoT applications:

[0234] First, use the Xavier method to initialize the strategy and evaluation network and set the initial control variables , and build an adjustable application set for the current time slice and fixed application sets Next, calculate the constraint violation degree of the current time slice , the formula is: . Assume application The actual service quality They are , target service quality Both , so it can be calculated that: . Based on the constraint violation degree of the current time slice , calculate the reward The formula is: . Assume that the degree of constraint violation , and the overall QoS sum , then: Next, the adjustable application set based on the current time slice and fixed application sets , and the degree of constraint violation , the updated status is: . Assume the status is updated to ,and Still Next, based on the updated state, determine whether there is a new sliding window. If there is a new sliding window, the policy network parameters and evaluation network parameters should be updated, and then based on these updated parameters, the service quality parameters of the IoT application are dynamically adjusted to obtain the updated execution ratio and deployment plan. The execution ratio of sample 1 task is updated as follows: , , , , the task deployment scheme is: On Server 1, On Server 2, On server 3, On server 4.

[0235] Step 5: Evaluate the new solution, including determining whether the performance of the new solution is better than the previous solution. If not, roll back the deployment:

[0236] After judgment, the new plan is better than the previous one, so the new plan is selected as the final result.

[0237] The following experiments are conducted to verify the effect of the present invention in improving the overall service quality. In order to make the experimental data more sufficient, we use the infrastructure data set provided by Shanghai Telecom to establish the SMEC system.

[0238] During the implementation, three SMEC systems were built, each containing 1000, 1500, and 2000 edge servers, respectively. In addition, a series of commercial edge servers were built, and the parameter settings are shown in Table 1. The constructed IoT application set contains 20,365 different actual DAG applications, and the parameter settings are shown in Table 2.

[0239] Table 1

[0240]

[0241] Table 2

[0242]

[0243] The infrastructure dataset includes a total of 3233 base stations, which are scattered in different geographical locations in Shanghai, China. Since each edge server should be co-located with a single base station, we randomly assign a certain number of edge servers to a subset of the 3233 base stations. The communication bandwidth between two edge servers is The failure arrival rates of each edge server and communication link are respectively and Select from.

[0244] In terms of offline, the improvement of the standardized overall service quality of the present invention is compared with that of NEMO, HiveMind, and REPFS algorithms; in terms of online, the improvement of the standardized overall service quality of the present invention is compared with that of OARC, EneX, and HAOL algorithms.

[0245] On the offline side, the NEMO algorithm uses the idle cycles of network functions to accelerate the execution of serverless functions to optimize latency. HiveMind is an integrated hardware and software stack designed for edge clusters containing autonomous devices. REPFS uses a function scheduling technology that ensures reliability to jointly optimize the profits of service providers and the quality of experience of end users.

[0246] On the online side, the OARC algorithm is an online rate adaptive reservation and control method for optimizing latency, and the EneX algorithm is an energy-aware scheduler for minimizing the energy consumption of latency-constrained function execution. Since reliability is not a design constraint, the HAOL algorithm is an integrated algorithm for solving cost- and latency-aware application deployment problems.

[0247] Figure 2 The experimental results of using the present invention, NEMO, HiveMind, and REPFS in standardizing the overall service quality for a task set distributed in 1000 server systems in the offline phase are given. Figure 2 As shown, the standardized overall service quality of the present invention is much higher than that of NEMO, HiveMind, and REPFS. For example, on average, the overall QoS achieved by the offline solution of the present invention is improved by 17.9%.

[0248] Figure 3 The experimental results of using the present invention, NEMO, HiveMind, and REPFS in standardizing the overall service quality for a task set distributed in 1500 server systems in the offline phase are given. Figure 3 As shown, the standardized overall service quality of the present invention is much higher than that of NEMO, HiveMind, and REPFS. For example, on average, the overall QoS achieved by the offline solution of the present invention is improved by 15.3%.

[0249] Figure 4 The experimental results of using the present invention, NEMO, HiveMind, and REPFS in standardizing the overall service quality for a task set distributed in 2000 server systems in the offline phase are given. Figure 4 As shown, the standardized overall service quality of the present invention is much higher than that of NEMO, HiveMind, and REPFS. For example, on average, the overall QoS achieved by the offline solution of the present invention is improved by 20.8%.

[0250] Figure 5 The experimental results of the present invention and OARC, EneX and HAOL in standardizing the overall service quality for a task set distributed in 1000 server systems in the offline phase are given. Figure 5 As shown, the standardized overall service quality of the present invention is much higher than that of OARC, EneX and HAOL. For example, under average conditions, the overall QoS achieved by the online solution of the present invention is improved by 31.2%.

[0251] Figure 6The experimental results of the present invention and OARC, EneX and HAOL in standardizing the overall service quality for a task set distributed in 1500 server systems in the offline phase are given. Figure 6 As shown, the standardized overall service quality using the present invention is much higher than that using OARC, EneX, and HAOL. For example, under average conditions, the overall QoS achieved by the online solution of the present invention is improved by 34.2%.

[0252] Figure 7 The experimental results of the present invention and OARC, EneX and HAOL in standardizing the overall service quality for a task set distributed in 2000 server systems in the offline phase are given. Figure 7 As shown, the standardized overall service quality using the present invention is much higher than that using OARC, EneX, and HAOL. For example, under average conditions, the overall QoS achieved by the online solution of the present invention is improved by 48.8%.

[0253] From the above experimental data, it can be clearly seen that the present invention has good performance in improving the overall service quality.

[0254] Embodiment 3

[0255] The present invention also provides a high-reliability personalized deployment system for maximizing service quality in a SMEC environment, comprising the following steps: a task acquisition module, a ratio generation module, a scheme initialization module, a ratio adjustment module, and a scheme optimization module;

[0256] The task acquisition module is used to obtain the task set of the server system;

[0257] The ratio generation module is used to generate an estimated optional execution ratio based on the task set of the server system;

[0258] The scheme initialization module is used to determine the initial deployment scheme of processor resource allocation and task execution based on the cross entropy method;

[0259] The ratio adjustment module is used to adjust the execution ratio and generate a new deployment plan based on the estimated optional execution ratio and the initial deployment plan;

[0260] The solution optimization module is used to dynamically adjust the service quality parameters of the Internet of Things application based on the new deployment solution and utilize proximal policy optimization to obtain an updated execution ratio and deployment solution.

[0261] In this embodiment, the ratio generation module includes: a first calculation unit, a second calculation unit, a third calculation unit, a comparison unit and a fourth calculation unit;

[0262] The first calculation unit is used to calculate the energy requirement of each task according to the energy consumption coefficient and the assumed execution ratio;

[0263] The second calculation unit is used to calculate the overall energy requirement based on the energy requirements of each task;

[0264] The third calculation unit is used to calculate the total energy supply;

[0265] A comparison unit is used to compare the total energy demand with the total energy supply;

[0266] The fourth calculation unit is used to calculate the optimal optional execution ratio if the total energy demand is greater than the total energy supply.

[0267] In this embodiment, the scheme initialization module includes: a first setting unit, a second setting unit, a sample generation unit, a performance evaluation unit, a sample sorting unit, a threshold calculation unit, a vector update unit, a sample selection unit and an initialization unit;

[0268] The first setting unit is used to set the iteration counter ;

[0269] The second setting unit is used to set the initial probability vector ;

[0270] The sample generation unit is used to generate a sample based on the probability vector generate Possible samples ;

[0271] The performance evaluation unit is used to evaluate each sample Conduct performance evaluations;

[0272] The sample sorting unit is used to The samples are sorted in descending order by performance evaluation value, and selected An elite sample;

[0273] The threshold calculation unit is used based on The average performance evaluation value of elite samples is used to calculate the new performance threshold ;

[0274] The vector updating unit is used to update the performance threshold based on the new performance threshold. , update the probability vector ;

[0275] The sample selection unit is used to select the sample based on the new probability vector , select the best performing sample;

[0276] The initialization unit is used to determine an initial deployment plan for processor resource allocation and task execution based on the best performing sample.

[0277] In this embodiment, the ratio adjustment module includes: a variable calculation unit, an application sorting unit, a variable updating unit, a ratio increment unit and a scheme updating unit;

[0278] The variable calculation unit is used to calculate the initial control variables ;

[0279] The application sorting unit is used to sort all applications in descending order and select applications that meet preset requirements;

[0280] The variable updating unit is used to update the initial control variable based on the initial control variable. and the application that meets the preset requirements, set the control variables ;

[0281] The proportional increment unit is used if the current control variable ,and ,calculate , get the execution ratio increment of each task;

[0282] The solution updating unit is used to update the execution ratio increment of each task based on the execution ratio increment of each task. , regenerate the deployment plan .

[0283] In this embodiment, the solution optimization module includes: an application set construction unit, a violation degree calculation unit, a reward calculation unit, a state update unit, a sliding window update unit, a parameter update unit and a dynamic adjustment unit;

[0284] The application set construction unit is used to initialize the strategy and evaluation network using the Xavier method and set the initialization control variables And build an adjustable application set for the current time slice and fixed application sets ;

[0285] The violation degree calculation unit is used to calculate the constraint violation degree of the current time slice. ;

[0286] The reward calculation unit is used to calculate the constraint violation degree based on the current time slice. , calculate the reward ;

[0287] The state update unit is used to adjust the application set based on the current time slice. , Fixed application set and constraint violation , update status to ;

[0288] The sliding window updating unit is used to determine whether there is a new sliding window based on the updated state;

[0289] The parameter updating unit is used to update the policy network parameters and evaluate the network parameters if there is a new sliding window;

[0290] The dynamic adjustment unit is used to dynamically adjust the service quality parameters of the Internet of Things application based on updating the physical strategy network parameters and evaluating the network parameters, and obtain the updated execution ratio and deployment plan.

[0291] The present invention aims to solve the above problems by improving the overall service quality of IoT applications in SMEC environments through a hybrid method that combines offline optimization and online optimization. The offline phase adopts a function placement strategy based on convex optimization and cross entropy methods, while the online phase uses a reinforcement learning scheme based on neighbor policy optimization (PPO) to dynamically adjust the task execution ratio and deployment scheme to adapt to the dynamic needs of IoT applications. The present invention maximizes the service quality while improving system reliability, and has important application value for IoT applications and embedded real-time systems.

[0292] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A high reliability personalized deployment method for maximizing service quality in a SMEC environment, characterized in that: The following steps are involved: Step 1: Get the task set of the server system; Step 2: generating an estimated optional execution ratio based on the task set of the server system; Step 3: Determine the initial deployment plan for processor resource allocation and task execution based on the cross entropy method; Step 4: Based on the estimated optional execution ratio and the initial deployment plan, adjust the execution ratio and generate a new deployment plan; Step 5: Based on the new deployment plan, using proximal policy optimization, dynamically adjust the service quality parameters of the IoT application to obtain an updated execution ratio and deployment plan; In step 3, the method for determining processor resource allocation and task execution based on the cross entropy method includes: Set iteration counter ; Set the initial probability vector ; Based on the probability vector generate Possible samples ; For each sample Conduct performance evaluations; Will The samples are sorted in descending order by performance evaluation value, and selected An elite sample; based on The average performance evaluation value of elite samples is used to calculate the new performance threshold ; Based on the new performance threshold , update the probability vector ; Based on the new probability vector , select the best performing sample; Based on the best performing samples, an initial deployment plan for processor resource allocation and task execution is determined.

2. The high reliability personalized deployment method for maximizing service quality in a SMEC environment according to claim 1, characterized in that: In step 2, the method for generating an estimated optional execution ratio based on the task set of the server system includes: Calculate the energy requirement of each task based on the energy consumption coefficient and the assumed execution ratio; Based on the energy requirements of each task, calculate the overall energy requirement; Calculate total energy supply; comparing the total energy demand to the total energy supply; If the total energy demand is greater than the total energy supply, the optimal optional execution ratio is calculated.

3. The high reliability personalized deployment method for maximizing service quality in a SMEC environment according to claim 1, characterized in that: In step 4, based on the estimated optional execution ratio and the initial deployment plan, the method of adjusting the execution ratio and generating a new deployment plan includes: Calculate initial control variables ; Sort all apps in descending order and select the ones that meet the preset requirements; Based on the initial control variables and the application that meets the preset requirements, set the control variables ; If the current control variable ,and ,calculate , get the execution ratio increment of each task; Based on the execution ratio increment of each task, update , regenerate the deployment plan .

4. The high reliability personalized deployment method for maximizing service quality in a SMEC environment according to claim 1, characterized in that: In step 5, based on the new deployment scheme, the method of dynamically adjusting the service quality parameters of the Internet of Things application by using proximal policy optimization includes: Use the Xavier method to initialize the strategy and evaluation network, and set the initialization control variables And build an adjustable application set for the current time slice and fixed application sets ; Calculate the constraint violation degree of the current time slice ; Based on the constraint violation level of the current time slice , calculate the reward ; Adjustable application set based on current time slice , Fixed application set and constraint violation , the updated status is ; Based on the updated status, determine whether there is a new sliding window; If there is a new sliding window, update the policy network parameters and evaluation network parameters; Based on updating the physical strategy network parameters and evaluating the network parameters, the service quality parameters of the IoT application are dynamically adjusted to obtain the updated execution ratio and deployment plan.

5. A high reliability personalized deployment system for maximizing service quality in a SMEC environment, characterized in that: The following steps are involved: Task acquisition module, ratio generation module, solution initialization module, ratio adjustment module and solution optimization module; The task acquisition module is used to acquire the task set of the server system; The ratio generation module is used to generate an estimated optional execution ratio based on the task set of the server system; The scheme initialization module is used to determine the initial deployment scheme of processor resource allocation and task execution based on the cross entropy method; The ratio adjustment module is used to adjust the execution ratio and generate a new deployment plan based on the estimated optional execution ratio and the initial deployment plan; The solution optimization module is used to dynamically adjust the service quality parameters of the Internet of Things application based on the new deployment solution by using proximal policy optimization to obtain an updated execution ratio and deployment solution; The scheme initialization module includes: a first setting unit, a second setting unit, a sample generation unit, a performance evaluation unit, a sample sorting unit, a threshold calculation unit, a vector update unit, a sample selection unit and an initialization unit; The first setting unit is used to set the iteration counter ; The second setting unit is used to set the initial probability vector ; The sample generation unit is used to generate a sample based on a probability vector generate Possible samples ; The performance evaluation unit is used to evaluate each sample Conduct performance evaluations; The sample sorting unit is used to The samples are sorted in descending order by performance evaluation value, and selected An elite sample; The threshold calculation unit is used to The average performance evaluation value of elite samples is used to calculate the new performance threshold ; The vector updating unit is configured to update the vector based on the new performance threshold , update the probability vector ; The sample selection unit is used to select the sample based on the new probability vector , select the best performing sample; The initialization unit is used to determine an initial deployment plan for processor resource allocation and task execution based on the best performing sample.

6. The high reliability personalized deployment system for maximizing service quality in a SMEC environment according to claim 5, characterized in that: The ratio generation module includes: a first calculation unit, a second calculation unit, a third calculation unit, a comparison unit and a fourth calculation unit; The first calculation unit is used to calculate the energy requirement of each task according to the energy consumption coefficient and the assumed execution ratio; The second calculation unit is used to calculate the overall energy demand based on the energy demand of each task; The third calculation unit is used to calculate the total energy supply; The comparison unit is used to compare the total energy demand with the total energy supply; The fourth calculation unit is used to calculate the optimal optional execution ratio if the total energy demand is greater than the total energy supply.

7. The high reliability personalized deployment system for maximizing service quality in a SMEC environment according to claim 5, characterized in that: The ratio adjustment module includes: a variable calculation unit, an application sorting unit, a variable updating unit, a ratio increment unit and a scheme updating unit; The variable calculation unit is used to calculate the initial control variable ; The application sorting unit is used to sort all applications in descending order and select applications that meet preset requirements; The variable updating unit is used to update the initial control variable based on the initial control variable. and the application that meets the preset requirements, set the control variables ; The proportional increment unit is used to ,and ,calculate , get the execution ratio increment of each task; The solution updating unit is used to update the execution ratio increment of each task based on the execution ratio increment of each task. , regenerate the deployment plan .

8. The high reliability personalized deployment system for maximizing service quality in a SMEC environment according to claim 5, characterized in that: The solution optimization module includes: an application set construction unit, a violation degree calculation unit, a reward calculation unit, a state update unit, a sliding window update unit, a parameter update unit and a dynamic adjustment unit; The application set building unit is used to initialize the strategy and evaluation network using the Xavier method, and to set the initialization control variables And build an adjustable application set for the current time slice and fixed application sets ; The violation degree calculation unit is used to calculate the constraint violation degree of the current time slice. ; The reward calculation unit is used to calculate the constraint violation degree based on the current time slice. , calculate the reward ; The state update unit is used to adjust the application set based on the current time slice. , Fixed application set and constraint violation , the updated status is ; The sliding window updating unit is used to determine whether there is a new sliding window based on the updated state; The parameter updating unit is used to update the policy network parameters and the evaluation network parameters if there is a new sliding window; The dynamic adjustment unit is used to dynamically adjust the service quality parameters of the Internet of Things application based on updating the physical strategy network parameters and evaluating the network parameters to obtain the updated execution ratio and deployment plan.

Images