Cloud edge-end collaborative task unloading strategy supporting cache

Abstract

The invention relates to the technical field of mobile edge computing, and provides the following technical solution aiming at the problem of cloud edge-end collaborative task unloading under a multi-user multi-server scene: firstly, constructing an edge computing system of a multi-user multi-server scene supporting service caching, introducing a caching mechanism into an MEC server, and modeling the edge computing system; on the basis, a cloud side-end collaborative task unloading framework supporting caching is designed, and a dynamic caching strategy based on task unloading scores is provided, so that the task unloading efficiency and the utilization rate of system resources are improved. Meanwhile, in order to further optimize the total profit of the server, a user demand curve is represented by adopting a linear function, and the load condition of the server and the relationship between the CPU frequency and the income of the server are simulated. Based on this, a task flexible pricing strategy based on a market pricing model is designed, so that the system can dynamically adjust pricing according to market changes, and the flexibility and economy of resource allocation are improved. And finally, solving the target optimization function in combination with an improved SWO algorithm, and providing an optimized cloud edge-end collaborative task unloading strategy supporting cache. According to the invention, on the basis of considering the income and cost of the service provider, the demands of the user for time delay, energy consumption and calculation payment are met, and the use experience of different users is improved while the profit of the service provider is maximized.

Description

Technical Field

[0001] The present invention belongs to the field of mobile edge computing, and specifically relates to a task offloading strategy for cloud-edge-end collaboration that supports caching. Background Art

[0002] With the rapid development of communication technology, people are increasingly starting to use non-local and more powerful servers to process complex applications, but this has also led to a huge challenge for massive communication and computing due to the increasing number of mobile devices and Internet data. MEC can significantly reduce communication and computing overhead and greatly improve user QoE. As the number of users gradually increases and the application complexity increases, the channel between the MEC server and the user becomes congested, and even requires long queues, seriously affecting the transmission and communication efficiency of tasks. Therefore, how to enhance the performance of the MEC network has become crucial.

[0003] Service caching is an effective method to improve the performance of the MEC network. By caching the content or application programs frequently accessed by users in the MEC server, the latency caused by repeated requests and the computing load of the server can be significantly reduced. This can not only shorten the response time of user requests, but also relieve the communication pressure between the MEC server and the core network, thus significantly improving user QoE. Although service caching can effectively improve system performance, the implementation of caching faces many challenges, such as which content needs to be cached and the update strategy in a complex environment. These issues need to comprehensively consider various factors such as server differences, network status, and user behavior. The caching strategy is closely related to the task offloading decision and affects the profit of service providers and user QoE. Therefore, in order to better meet the growing user needs and complex application scenarios, combining the caching strategy with the offloading decision and achieving collaborative optimization is an important research direction at present.

[0004] The research on profit is different from the research on server cost. Profit further considers the income of the server. Although a fixed resource price can meet some needs, in a complex scenario of multiple users and multiple servers, a dynamic pricing strategy can better adapt to the changes in resource supply and demand. Therefore, it is particularly crucial to characterize the relationship between the resources provided by the MEC server and the charging price of the UE. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention proposes a task offloading strategy for cloud-edge-end collaboration that supports caching. The strategy includes:

[0006] S1: Construct an edge computing system in a multi-user and multi-server scenario that supports caching;

[0007] S2: Propose a dynamic caching strategy for task offloading scoring based on the edge computing system;

[0008] S3: Based on the edge computing system, calculate the profit obtained by the computing server provider when processing user tasks and the QoE of users;

[0009] S4: Construct an objective optimization function according to the task elastic pricing strategy of the market pricing model;

[0010] S5: Use the improved SWO algorithm to solve the objective function to obtain the task offloading strategy.

[0011] Preferably, the task caching model is: Use to represent the task set I of users u The task task i in Edge j Whether there is a cache is as shown in the formula specifically.

[0012]

[0013] Among them, L j represents the load condition of server j; L0 represents the load threshold of the server. At the same time, the MEC server Edge j needs to satisfy the formula.

[0014]

[0015] Among them, ω i represents the size of the task task i The task task i Whether there is a cache in the edge layer is represented by y i as shown in the formula specifically.

[0016]

[0017] Preferably, the task computing model is: The task Task u generated by UE u can also be processed locally, on the MEC server, and on the cloud server. The binary variables and are used to determine the offloading location of the tasks in Task u satisfy Since service caching is introduced in the MEC server, the task offloading strategy must fully consider the impact of caching on the offloading process. Specifically, if the target task is already cached in the MEC server, the user can directly obtain the calculation result, thereby greatly reducing the latency and energy consumption of task processing. The transmission rate r of each user in the channel u is obtained according to.

[0018] When Task uWhen the task in is executed locally, the computing delay of the task

[0019]

[0020] is as follows, as shown in the formula.

[0021]

[0022] When the task in Task u is offloaded to ES for execution, the network topology relationship D i is expressed as

[0023]

[0024] The transmission delay of the task is as follows,

[0025]

[0026] The transmission energy consumption is as shown in the formula.

[0027]

[0028] Task u The processing delay of the task on the server Edge j is as follows, as shown in the formula.

[0029]

[0030] The processing energy consumption is as shown in the formula.

[0031]

[0032] Task u The total computing delay of the task on the server Edge j is as follows, as shown in the formula.

[0033]

[0034] The computing energy consumption is as shown in the formula.

[0035]

[0036] The total computing delay u and energy consumption of ES processing Task are shown in the following two formulas.

[0037]

[0038] When the task in Task u is offloaded to the cloud server for execution, the task is offloaded to the cloud server through the optical fiber, and the task transmission delay and energy consumption are as shown in the following two equations.

[0039]

[0040] The computing delay of the task in Task u on the cloud server is as shown in the equation,

[0041]

[0042] The computing energy consumption is as shown in the equation.

[0043]

[0044] The total delay of the cloud server in executing Task u is specifically expressed as the following equation,

[0045] [[ID=]43]

[0046] The total energy consumption of executing the task is as shown in the equation.

[0047]

[0048] The total offloading delay T u and energy consumption E u of Task u are as shown in the following two equations.

[0049]

[0050] Preferably, based on the dynamic caching strategy of task offloading scoring, the specific process is as follows:

[0051] Obtain the total number of all tasks in the current system and the tasks that need to be processed currently;

[0052] Initialize the cache list. The initialized list may contain some cached tasks, calculate the free cache space, and ensure whether the remaining space is sufficient to cache new tasks.

[0053] ​Update the scores of the cached tasks. Traverse all tasks. For each task, if the task is already cached, update the task popularity and calculate the new score; if not cached, skip the process.

[0054] Detect whether the current task is already cached. If the task is in the cache, directly return the current task list without update.

[0055] Processing when the cache space is sufficient. When the space is enough, directly cache the task, add the task to the cache list, and return the updated task list.

[0056] Trigger the replacement process when the space is insufficient. Update the cache list using the knapsack problem and return the updated task list.

[0057] Preferably, its dynamic caching policy based on task offloading scores is specifically implemented as follows: Task i in the cache list of server j is represented by a tuple: where represents the offloading score of the cached task i in MEC server Edge j , represents the popularity of the task. The offloading evaluation is a comprehensive score considering service providers and users, specifically as the formula:

[0058]

[0059] where χ1, χ2, and χ3 represent the preference coefficients for profit, user QoE, and popularity respectively. χ1 and χ2 change dynamically according to the load, satisfying χ1 + χ2 = 1. In this chapter, χ3 is fixed at 0.5. To avoid the denominator being zero, a minimum value ∈ = 1×10 -5 is introduced; to alleviate the cold start bias, is initialized to 1, and a decay coefficient h0 is introduced to dynamically adjust the task popularity. In this chapter, h0 = 0.5 is adopted.

[0060]

[0061] Preferably, the formula for calculating user QoE is:

[0062]

[0063] θ t , θ e and θ r , respectively represent the user's preference coefficients for latency, energy consumption, and price when executing tasks. Among them, θ t , θ e , θ r ∈[0,1], satisfying θ t + θ e + θ r= 1.

[0064] Preferably, the objective function is the integer optimization problem P.

[0065] P: Max Φ profit

[0066] s.t. C1:

[0067] C2:

[0068] C3:

[0069] C4:

[0070] C4:

[0071] Profit u represents the profit obtained by the server provider for processing the tasks of user u, and QoE u represents the QoE of user u, and QoE th represents the QoE threshold of the user, U represents the number of users, and x u represents the offloading strategy of user u.

[0072] Preferably, the improved algorithm integrates a variety of improvement strategies on the basis of the basic SWO algorithm, as follows:

[0073] (1) Discrete optimization of the SWO algorithm. The basic SWO algorithm is designed for continuous optimization problems and cannot effectively handle discrete problems. The individual update position in the algorithm is a continuous value However, for the task offloading problem of MEC, the solution needs to be discretized, and the current solution is mapped to an integer k ∈ {0, 1,..., J + 1}. The specific rules for discretized update are: First, calculate the normalized distance d of each candidate integer k from the continuous value k , as shown in the following formula.

[0074]

[0075] Then map the distance to a probability p k , as shown in the following formula.

[0076]

[0077] where ∈ = 1 × 10 -5 . Finally, according to the probability distribution {p0, p1,..., p J+1} generate a discrete decision K through polynomial distribution sampling and update the current solution

[0078] In the population update stage, a common problem is that the update amplitude is too small, resulting in the loss of update information after discretizing the current solution, thus falling into a local optimal solution. To solve this problem, the update strategy in this chapter introduces the Cauchy perturbation mutation strategy to achieve the discrete optimization of the SWO algorithm. The Cauchy perturbation mutation formula is shown as follows.

[0079]

[0080] (2) Dynamic adjustment of parameters. In the SWO algorithm, the parameters Tr and Cr directly affect the behavior of spider wasps. The parameter Tr affects the exploration ability during the algorithm search process. When Tr is small, the algorithm tends to perform more local searches, that is, the spider wasp tends to follow the current local optimal solution and narrow the search range; when Tr is large, the algorithm tends to perform more global searches. This means that the spider wasp is more likely to search randomly, jump out of the current local search area, and enhance the diversity of exploration. A reasonable value can balance global search and local search. In this chapter, the value of Tr is dynamically adjusted according to the iteration number by exponential decay, encouraging the population to actively explore in the early stage of iteration and focusing on exploitation in the later stage, as shown in the following formula.

[0081]

[0082] where t and T max represent the current iteration number and the maximum iteration number; Tr0 represents the initial value of Tr; ν1 is a decay coefficient.

[0083] The parameter Cr controls whether the spider wasp generates a new solution by crossover or directly selects the current solution during the mating stage, thus affecting the exploration ability and population diversity of the algorithm. When Cr is large, the algorithm is more inclined to generate diverse new solutions through crossover operations, which helps to introduce new solution spaces and enhance the diversity of the population; when Cr is small, the algorithm relies more on the current solution and local searches, showing stronger local search stability. In this chapter, Cr is dynamically adjusted based on the fitness value: when the optimal solution has not been improved for a long time, the algorithm enhances the exploration ability of the population, and the specific adjustment strategy is shown in the following formula.

[0084]

[0085] where d represents the change value of the fitness after δ iterations; d0 represents the threshold of the fitness change, and in this chapter, d0 = 5; v2 and ν3 represent the growth coefficient and the decay coefficient respectively; Cr min and Cr max represent the minimum and maximum values of the parameter Cr respectively.

[0086] (3) Introduce chaotic mapping to initialize the population. In the traditional SWO algorithm, a simple random method is usually used to generate the initial population. This method is prone to uneven population distribution, which in turn affects the search efficiency and solution accuracy of the algorithm. To solve this problem, chaotic mapping is introduced in this chapter to initialize the population. The Tent mapping generates a chaotic sequence through iteration, as shown in Equation [eq:3-36], which can produce an initial population with uniform distribution and rich diversity. This initialization method not only avoids the limitations of the traditional random method but also significantly improves the global search ability of the algorithm.

[0087]

[0088] Among them, num i represents the i-th chaotic sequence number. Based on Equation [eq:3-36], the population initialization is shown as follows

[0089]

[0090] (4) Introduce the elite opposition-based learning strategy. This strategy effectively avoids the algorithm from falling into local optima by introducing more solutions with potential optimization value in the population, thus enhancing the global search ability. At the same time, the construction of opposition solutions can expand the solution space and promote population diversity, which is of great significance for solving complex discrete optimization problems. Specifically, it is shown as follows.

[0091]

[0092] Among them, SW t op represents the opposition solution of the elite individual; represents the elite individual, i.e., the optimal solution, in the t-th iteration; rand is a random number in the range (0, 1].

[0093] The above is the improved spider wasp optimization algorithm. By using the improved SWO algorithm, we can obtain the final task offloading strategy.

[0094] Advantages and beneficial effects of the present invention:

[0095] A service caching mechanism is introduced in the MEC system of the present invention, and a detailed model is established for the multi-user multi-server scenario. The present invention proposes a steady-state caching strategy based on task offloading scoring, which can effectively improve the profit of service providers in comparison with the LFU, LRU, and CE strategies. In addition, the present invention designs a flexible pricing scheme for tasks based on the market pricing model, and proposes a task offloading strategy based on cloud-edge-end collaboration in combination with the above scheme. The offloading strategy proposed by the present invention can not only effectively improve the revenue of the server but also meet the QoE requirements of more users. This strategy shows more excellent performance, demonstrating its advantages in complex scenarios. Brief Description of the Drawings

[0096] Figure 1 is a flowchart of the task offloading strategy for cloud-edge-end collaboration with cache support according to the present invention;

[0097] Figure 2 is an architecture diagram of the cloud-edge-end collaboration system with cache support according to the present invention Detailed Embodiments

[0098] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0099] The system architecture diagram of the present invention is as shown in Figure 2 described. This architecture consists of multiple user devices (UEs), multiple edge servers (ESs) with different configurations, and a cloud server. Among them, the user devices constitute the user layer, the ESs constitute the edge layer, and the cloud server constitutes the cloud layer. Each user device generates certain computing tasks (such as application execution tasks), and the edge layer and the cloud layer provide computing services for these tasks, especially for UEs with limited resources. In this architecture, the ESs not only have computing capabilities but also are equipped with a certain amount of storage space, which can cache the processed task data, or download and store application programs from the central cloud through wired links to accelerate the task processing speed. This three-layer collaborative architecture makes full use of the computing and processing capabilities of each layer: at the user layer, the UEs can process some local tasks; at the edge layer, the ESs provide local offloading services for tasks with higher resource requirements; at the cloud layer, the cloud server provides powerful computing capabilities, which are suitable for processing complex and computationally intensive tasks.

[0100] The present invention proposes a task offloading strategy for cloud-edge-end collaboration with cache support, as shown in Figure 1 . This strategy includes:

[0101] S1: Construct an edge computing system in a multi-user multi-server scenario with cache support.

[0102] In this system, the task set I is represented as I = {1, 2,..., I}, and task i is represented by the tuple: task i = {Id i , S i , C i}. Among them, Id i represents the unique number of the task; S i represents the size of the task, in kb; C iRepresents the computing resources required by the user, i.e., the number of CPU cycles required, with the unit of cycle / kb.

[0103] The user layer consists of multiple UEs, and the set of UEs is represented as: UEs = {1, 2,..., U}. The tasks generated by each UE are defined as a tuple type and represented as where I u represents the set of tasks that user u needs to process, satisfying I u ∈ I; represents the maximum tolerable delay of the task; represents the maximum tolerable energy consumption of the task; represents the maximum tolerable payment of the task; ES u corresponds to a server, serving as the local server of user u. The UE has a certain computing ability, and the UE u can be represented as where represents the local computing ability, i.e., the number of cycles processed by the CPU per second; represents the number of CPU cores, indicating the number of tasks that can be processed concurrently.

[0104] The edge layer contains multiple ESs with different configurations, represented as the set Edges = {1, 2,..., J}. Each server is represented by a tuple where represents the server number; represents the computing ability of server Edge j ; represents the number of CPU cores of server Edge j ; Ω h represents the cache size of server Edge j ; List j represents the cache list of server Edge j ; represents the distance between the cloud server and server Edge j ;

[0105] The cloud server Cloud = {f c} has stronger computing ability f c , but is deployed far from the user, requiring a certain transmission delay and a higher cost-economic coefficient.

[0106] S2: Propose a dynamic caching strategy for task offloading scoring based on the edge computing system;

[0107] Task caching model design: Use to represent the task task in the user's task set I u ; iIs there a cache on Edge j as shown in the formula below.

[0108]

[0109] where L j represents the load condition of server j; L0 represents the load threshold of the server. At the same time, the MEC server Edge j needs to satisfy the formula.

[0110]

[0111] where ω i represents the size of task task i . Whether there is a cache of task task i in the edge layer is indicated by y i as shown in the formula below.

[0112]

[0113] Task calculation model: The task Task u generated by UE u can also be processed locally, on the MEC server, and on the cloud server. The binary variables and are used to determine the offloading location of the task in Task u satisfying Since service caching is introduced in the MEC server, the task offloading strategy must fully consider the impact of caching on the offloading process. Specifically, if the target task is already cached in the MEC server, the user can directly obtain the calculation result, thereby significantly reducing the latency and energy consumption of task processing. The transmission rate r u of each user in the channel is obtained according to.

[0114] When the task in Task u is executed locally, the computational latency of the task is as shown in the formula below,

[0115]

[0116] The computational energy consumption of the task is as shown in the formula.

[0117]

[0118] When the task in Task u is offloaded to the ES for execution, the network topology relationship D i is expressed as,

[0119]

[0120] Transmission delay of the task Specifically as shown in the formula,

[0121]

[0122] Transmission energy consumption As shown in the formula.

[0123]

[0124] Task u The task in j Processing delay on the server Edge As shown in the formula,

[0125]

[0126] Processing energy consumption As shown in the formula.

[0127]

[0128] Task u The task in j Total computing delay on the server Edge As shown in the formula,

[0129]

[0130] Computing energy consumption As shown in the formula.

[0131]

[0132] ES processes Task u Total computing delay and energy consumption Are shown in the following two formulas.

[0133]

[0134] When Task u The task in is offloaded to the cloud server for execution. The task is offloaded to the cloud server through optical fiber. The task transmission delay and energy consumption Are shown in the following two formulas.

[0135]

[0136] Task u [[ID=�4]]Computing delay of the task in the cloud server As shown in the formula,

[0137]

[0138] Calculate the energy consumption As shown in the formula.

[0139]

[0140] The cloud server executes Task u The total latency Specifically expressed as the following formula,[[]]

[0141]

[0142] The total energy consumption of executing the task As shown in the formula.

[0143]

[0144] Task u The total offloading latency T u And the energy consumption E u Are shown in the following two formulas.

[0145]

[0146] S3: Based on the edge computing system, calculate the profit obtained by the server provider when processing user tasks and the user's QoE;

[0147] In this system, the task Task generated by the UE i Can be executed locally or offloaded to the MEC server or the cloud server for execution. Each Task i Is determined by three binary variables, i∈I respectively represents that the task is offloaded to the local UE for execution, the task is offloaded to the server Edge j For execution or execution on the cloud server, and satisfies the following formula.

[0148]

[0149] (1) Local execution. When When, Task i Is executed on the local device UE, and the computing latency of the task Is shown in the following formula.

[0150]

[0151] Among them, Represents the computing power of the UE i The energy consumption generated by local execution Is shown in the following formula.

[0152]

[0153] Among them, k represents the energy coefficient, which is determined by the CPU architecture of the UE.

[0154] (2) Execution by the edge server. When Task i is offloaded to the ES for execution, the offloading delay includes the transmission delay and the processing delay of the task. Task i offloaded to Edge j The sum of the transmission delays of all the links passed by is shown in the following formula.

[0155]

[0156] In formula [eq:3-7], is a binary variable, indicating whether the link (a, b) between server Edge a and server Edge b acts on Task i , and the shortest path of the link (a, b) is obtained by the dijkstra algorithm; r a,b is the transmission rate of the link (a, b). The energy consumption generated during the entire transmission process is shown in the following formula.

[0157]

[0158] Among them, P a,b represents the channel transmission power of the link (a, b). The delay for Edge j to process Task i is shown in formula [eq:3-9], and the energy consumption is shown in the following formula.

[0159]

[0160] Among them, represents the CPU processing capacity of Edge j . The total delay i for Task j offloaded to Edge for processing is shown in formula [eq:3-11], and the total energy consumption is shown in the following formula:

[0161]

[0162] According to the above formulas, the total calculation delay i for the ES to process Task is shown in the following formula

[0163]

[0164] , the total computing energy consumption is shown as follows:

[0165]

[0166] (3) Executed by the cloud server. When the ES is overloaded, or for compute-intensive applications where the ES cannot provide high-quality services to users, the cloud server will assist in offloading. At this time Different from offloading to the ES, the task is offloaded to the cloud server through the cloud link for execution, and this process is transmitted through optical fiber. Task i The transmission delay of the task offloaded to the cloud server is shown as follows,

[0167]

[0168] The transmission energy consumption is shown as follows:

[0169]

[0170] where r c represents the transmission rate of the cloud link; r0 represents the speed of light; D represents the distance from the user to the cloud server. Task i The delay in processing in the cloud server is shown as follows,

[0171]

[0172] The energy consumption is expressed as:

[0173]

[0174] where f c represents the CPU processing capacity of the cloud server. Task i The total delay of the task offloaded to the cloud server for execution and the energy consumption are shown as follows respectively.

[0175]

[0176] In summary, Task i The total offloading delay T i is expressed as,

[0177]

[0178] The total energy consumption Ei As shown below.

[0179]

[0180] If the user's task is offloaded to the ES and the cache is used, and the user needs to pay for the cache service, the calculation of the payment for user u is specifically as shown in the formula.

[0181]

[0182] Among them, ξ2 represents the price of the cloud server for calculating each Mbits of data; ξ3 represents the price of the cache on the ES. The calculation cost of the server for processing Task u and the cache cost are as shown below. As shown below.

[0183]

[0184] Among them, ξ t and ξ e respectively represent the cost economic coefficients of delay and energy consumption; ξ c is the cost economic coefficient regarding the cache on the ES, and its value needs to be set according to specific application scenarios and requirements.

[0185] The profit u obtained by the server provider from Task is as shown below.

[0186]

[0187] According to the above formula, the total profit Φ profit of the service provider is as shown below.

[0188]

[0189] S4: Construct the target optimization function according to the task elastic pricing strategy of the market pricing model;

[0190] By ensuring the user's QoE while maximizing the server's profit, it is modeled as a linear integer optimization problem P.

[0191] P: Max Φ profit

[0192] s.t. C1:

[0193] C2:

[0194] C3:

[0195] C4:

[0196] C4:

[0197] Profit u represents the profit obtained by the server provider for processing the tasks of user u, and QoE u represents the QoE of user u, and QoE th represents the user QoE threshold, U represents the number of users, and x u represents the offloading strategy of user u.

[0198] S5: Solve the objective function using the improved SWO algorithm to obtain the task offloading strategy.

[0199] The improved SWO algorithm is based on the basic SWO algorithm and incorporates a variety of improvement strategies, which are specifically as follows:

[0200] (1) Discrete optimization of the SWO algorithm. The basic SWO algorithm is designed for continuous optimization problems and cannot effectively handle discrete problems. The individual update position in the algorithm is a continuous value However, for the task offloading problem in MEC, the solution needs to be discretized, mapping the current solution to an integer K ∈ {0, 1,..., j + 1}. The specific rules for discrete update are as follows: First, calculate the normalized distance d of each candidate integer K from the continuous value k , as shown in the following formula.

[0201]

[0202] Then map the distance to a probability p k , as shown in the following formula.

[0203]

[0204] where, ∈ = 1 × 10 -5 . Finally, according to the probability distribution {p0, p1,..., p J+1} generate a discrete decision K through polynomial distribution sampling and update the current solution

[0205] In the population update stage, a common problem is that the update amplitude is too small, resulting in the loss of update information after discretizing the current solution, thus falling into a local optimal solution. To solve this problem, the update strategy in this chapter introduces the Cauchy perturbation mutation strategy to achieve the discrete optimization of the SWO algorithm. The Cauchy perturbation mutation formula is as shown in the following formula.

[0206]

[0207] (2) Dynamic adjustment of parameters. In the SWO algorithm, parameters Tr and Cr directly affect the behavior of spider wasps. Parameter Tr affects the exploration ability during the algorithm's search process. When Tr is small, the algorithm tends to perform more local searches, that is, the spider wasp tends to follow the current local optimal solution and narrow the search range; when Tr is large, the algorithm tends to perform more global searches. This means that the spider wasp is more likely to search randomly, jump out of the current local search area, and enhance the diversity of exploration. A reasonable value can balance global search and local search. In this chapter, the value of Tr is dynamically adjusted according to the number of iterations by exponential decay, encouraging the population to actively explore in the early stage of iteration and focusing on exploitation in the later stage, as shown in the following formula.

[0208]

[0209] where t and T max represent the current iteration number and the maximum iteration number; Tr0 represents the initial value of Tr; v1 is a decay coefficient.

[0210] Parameter Cr controls whether the spider wasp generates a new solution by crossover or directly selects the current solution during the mating stage, thus affecting the exploration ability and population diversity of the algorithm. When Cr is large, the algorithm is more inclined to generate diverse new solutions through crossover operations, which helps to introduce new solution spaces and enhance the diversity of the population; when Cr is small, the algorithm relies more on the current solution and local searches, showing stronger local search stability. In this chapter, Cr is dynamically adjusted based on the fitness value: when the optimal solution has not been improved for a long time, the algorithm enhances the exploration ability of the population, and the specific adjustment strategy is shown in the following formula.

[0211]

[0212] where d represents the change value of the fitness after δ iterations; d0 represents the threshold of the fitness change, and in this chapter, d0 = 5; ν2 and ν3 represent the growth coefficient and decay coefficient respectively; Cr min and Cr max represent the minimum and maximum values of parameter Cr respectively.

[0213] (3) Introducing chaotic mapping to initialize the population. The traditional SWO algorithm usually uses a simple random method to generate the initial population. This method is prone to uneven population distribution, which in turn affects the search efficiency and solution accuracy of the algorithm. To solve this problem, this chapter introduces chaotic mapping to initialize the population. The Tent mapping generates a chaotic sequence through iteration, as shown in Equation [eq:3-36], and can produce an initial population with uniform distribution and rich diversity. This initialization method not only avoids the limitations of the traditional random method but also significantly improves the global search ability of the algorithm.

[0214]

[0215] Among them, num i represents the i-th chaotic sequence number. Based on Equation [eq:3-36], the population initialization is shown as follows

[0216]

[0217] (4) Introduce the elite opposition-based learning strategy. This strategy effectively avoids the algorithm from falling into local optima by introducing more solutions with potential optimization value in the population, thereby enhancing the global search ability. At the same time, the construction of opposition solutions can expand the solution space and promote population diversity, which is of great significance for solving complex discrete optimization problems. It is specifically shown as follows

[0218]

[0219] Among them, SW t op represents the opposition solution of the elite individual; represents the elite individual, i.e., the optimal solution, in the t-th iteration; rand is a random number in the range of (0, 1].

[0220] The above is the improved spider wasp optimization algorithm. By using the improved SWO algorithm, we can obtain the final task offloading strategy

[0221] The above examples further elaborate on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above examples are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included in the protection scope of the present invention

Claims

1. A task offloading strategy for cloud-edge-end collaboration with caching support, characterized in that The method includes: S1: Construct an edge computing system in a multi-user and multi-server scenario with cache support; S2: Propose a dynamic caching policy for task offloading scoring based on the edge computing system; S3: Based on the edge computing system, calculate the profit obtained by the server provider when processing user tasks and the QoE of the users; S4: Construct an objective optimization function according to the task elastic pricing strategy of the market pricing model; S5: Use the improved SWO algorithm to solve the objective function to obtain the task offloading strategy.

2. The task offloading strategy for cloud-edge-end collaboration with cache support according to claim 1, characterized in that Its dynamic caching policy based on task offloading scoring, and its specific process is as follows: Obtain the total number of all tasks in the current system and the tasks that need to be processed currently; Initialize the cache list. The initialized list may contain some cached tasks, calculate the free cache space, and ensure whether the remaining space is sufficient to cache new tasks. Update the scores of the cached tasks. Traverse all tasks. For each task, if the task is already cached, update the task popularity and calculate the new score. If it is not cached, skip the process; Detect whether the current task is already cached. If the task is in the cache, directly return the current task list without update; Processing when the cache space is sufficient. When the space is sufficient, directly cache, add the task to the cache list, and return the updated task list; Trigger the replacement process when the space is insufficient. Update the cache list using the knapsack problem and return the updated task list.

3. The task offloading strategy for cloud-edge-end collaboration supporting caching according to claim 1, characterized in that, Its dynamic caching policy based on task offloading scores is specifically implemented as follows: Task i in the cache list of server j is represented by a tuple: where represents the offloading score of the cached task i in the MEC server Edge j , and represents the popularity of the task. The offloading evaluation is a comprehensive score considering both service providers and users, specifically as follows: Among them, χ1, χ2, and χ3 represent the preference coefficients for profit, user QoE, and popularity respectively. χ1 and χ2 change dynamically according to the load and satisfy χ1 + χ2 = 1. In this chapter, χ3 is fixed at 0.

5. To avoid the denominator being zero, a minimum value ∈ = 1×10 -5 is introduced; to alleviate the cold start bias, it is initialized to 1, and a decay coefficient h0 is introduced to dynamically adjust the task popularity. In this chapter, h0 = 0.5 is adopted.

4. A task offloading strategy for cloud-edge-end collaboration with cache support according to claim 1, characterized in that The formula for calculating the user QoE is: θ t , θ e and θ r , respectively represent the user's preference coefficients for latency, energy consumption, and price when performing tasks. Among them, θ t , θ e , θ r ∈[0, 1], and satisfy θ t + θ e + θ r = 1.

5. A task offloading strategy for cloud-edge-end collaboration with cache support according to claim 1, characterized in that, The constructed objective function is: By ensuring the user QoE while maximizing the profit of the server, it is modeled as a linear integer optimization problem P. Profit u represents the profit obtained by the server provider for processing the tasks of user u, and QoE u represents the QoE of user u, and QoE th represents the QoE threshold of the user, U represents the number of users, and x u represents the offloading strategy of user u.

Images