A computational task offloading optimization method considering multiple performance indicators

Abstract

The present invention provides a computing task unloading optimization method considering multiple performance indicators, which belongs to the field of edge computing technology. The method adopts a hybrid of orthogonal multiple access technology and non-orthogonal multiple access technology to provide services for multiple users at the same time. The method includes: calculating the time cost of each computing task under a preset computing path; inputting the time cost of each computing task into a pre-established computing task allocation model to generate a preliminary allocation plan; establishing a MINLP model for computing task unloading optimization problems; taking the preliminary allocation plan as a starting point, using Lagrangian relaxation method and dual decomposition method to solve the MINLP model, and obtaining the optimal computing task unloading optimization plan. When using Lagrangian relaxation and dual decomposition methods to solve the MINLP model, the present invention uses a computing task allocation model to provide a preliminary allocation plan, which reduces the search range, reduces the computational complexity, and improves the real-time performance of the system.

Description

Technical Field

[0001] The present invention relates to a computing task offloading optimization method considering multiple performance indicators, and belongs to the field of edge computing technology. Background Art

[0002] In the field of contemporary wireless communications, with the surge in the number of mobile devices and the increasingly complex computing requirements of application services, traditional ground base stations have gradually become unable to meet users' needs for high-performance computing, especially in special scenarios such as remote areas or disaster areas. In order to overcome these limitations, researchers have begun to explore the use of low-orbit satellites (LEO) to build edge cloud (SEC) networks to provide computing resources and services without geographical restrictions. However, when dealing with task offloading, existing satellite communication systems often ignore the dynamic changes in path loss and channel conditions, which may lead to inflexible and inefficient task offloading decisions. In actual network environments, the number of users, task types, and channel quality served by different base stations vary, and simple offloading paths such as simple base station-satellite cannot flexibly adapt to these changes. The simple combination mode of existing technologies in task offloading paths is difficult to meet the diverse user needs and task requirements in complex network environments.

[0003] In solving nonlinear mixed integer programming (MINLP) problems, traditional methods such as branch and bound method and particle swarm optimization algorithm PSO have been widely studied.

[0004] The branch and bound method breaks down the problem into sub-problems and gradually searches for the optimal solution, but the computational complexity increases dramatically when dealing with large-scale problems. It is essentially an exhaustive search strategy that requires traversing all possible solution spaces of the problem to determine the optimal solution. In each branching process, the newly generated sub-problems need to be evaluated and calculated, resulting in a rapid expansion of the amount of calculation as the scale of the problem increases. At the same time, during the search process, due to the lack of an effective guidance mechanism, it is easy to get lost in a large number of branches and cannot quickly focus on the area where the optimal solution is located.

[0005] The particle swarm optimization algorithm (PSO) is prone to falling into local optimal solutions, which makes it impossible to find the global optimal solution, thus affecting the effect of task offloading and resource allocation. The update of particles in the algorithm mainly depends on its own historical optimal position and the historical optimal position of the group. This update mechanism limits the search range and exploration ability of particles to a certain extent. When a particle finds a local optimal position during the search process, due to the characteristics of its update formula, the particle may linger around the local optimal value and find it difficult to jump out of the local optimal area to explore other possible better solutions, resulting in premature convergence to the local optimal value and missing the global optimal solution.

[0006] In addition, the PSO algorithm is relatively difficult to deal with complex constraints, and additional constraint processing strategies may interfere with its original search process, further affecting its ability to find the global optimal solution.

[0007] Therefore, there is an urgent need for an efficient MINLP problem solving method to be applied to the task offloading path decision of satellite communication systems to meet the diverse user needs and task requirements in complex network environments. Summary of the invention

[0008] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a computing task offloading optimization method considering multiple performance indicators, thereby solving the problem that the MINLP problem of task offloading path decision in a satellite communication system is difficult to solve.

[0009] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0010] The present invention provides a computing task offloading optimization method considering multiple performance indicators, adopting orthogonal multiple access OMA technology to divide the entire available broadband spectrum into multiple orthogonal subcarriers; in each of the orthogonal subcarriers, non-orthogonal multiple access NOMA technology is adopted to provide services for multiple corresponding users at the same time, and for the computing tasks generated by users within a preset time period, the following steps are performed:

[0011] Step S1: Calculate the time cost of each computing task under the preset computing path respectively;

[0012] Step S2: inputting the time cost of each computing task under the preset computing path into a pre-established computing task allocation model to generate a preliminary allocation plan; wherein the computing task allocation model is established based on a neural network;

[0013] Step S3: Establishing a MINLP model for the computational task offloading optimization problem;

[0014] Step S4: taking the preliminary allocation scheme as the starting point for solving the MINLP model, solving the MINLP model using the Lagrangian relaxation method and the dual problem decomposition method, and obtaining the optimal computing task offloading optimization scheme.

[0015] Furthermore, the step S1 includes:

[0016] Step S11: Calculate the time cost of the current user's local calculation as the first time cost, the first time cost includes: local execution time ;

[0017] The local execution time calculation formula is:

[0018] ;

[0019] In the formula, is the CPU cycle frequency of the current user, is the current computing task data volume, The number of CPU cycles required to process each bit of data, in cycles / bit;

[0020] Step S12: Calculate the time cost of the current user offloading the computing task to the corresponding satellite edge computing node SEC through the user-satellite link for computing, and use it as the second time cost. The second time cost includes: the user-satellite two-way propagation delay , User-satellite data transmission delay and satellite execution time ;

[0021] The calculation formula for the user-satellite two-way propagation delay is:

[0022] ;

[0023] In the formula, h is the distance between the current user and the corresponding satellite, and c is the speed of light;

[0024] The user-satellite data transmission delay calculation formula is:

[0025] ;

[0026] In the formula, is the transmission rate between the current user and the corresponding satellite;

[0027] The satellite execution time calculation formula is:

[0028] ;

[0029] In the formula, is the CPU cycle frequency of the corresponding satellite;

[0030] Step S13: Calculate the time cost of the current user offloading the computing task to the corresponding satellite edge computing node SEC through the user-base station-satellite link for computing, and use it as the third time cost. The third time cost includes: base station-satellite two-way propagation delay , User-base station data transmission delay , Base station-satellite data transmission delay and satellite execution time ;

[0031] The base station-satellite two-way propagation delay is The two-way propagation delay between the user and the satellite same;

[0032] The calculation formula for the user-base station data transmission delay is:

[0033] ;

[0034] In the formula, is the transmission rate between the current user and the corresponding base station;

[0035] The base station-satellite data transmission delay calculation formula is:

[0036] ;

[0037] In the formula, The total amount of computing task data transmitted together with the current task through the corresponding base station to the corresponding satellite; is the transmission rate between the corresponding base station and the corresponding satellite, which is the same as the transmission rate between the current user and the corresponding satellite ;

[0038] Step S14: Calculate the time cost of the current user offloading the computing task to the corresponding ground cloud server through the user-satellite-ground cloud link for computing, and use it as the fourth time cost. The fourth time cost includes: the user-satellite two-way propagation delay , User-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time ;

[0039] The satellite-to-ground cloud data transmission delay calculation formula is:

[0040]

[0041] In the formula, is the transmission rate between the satellite and the corresponding ground cloud, which is a preset constant value;

[0042] The ground cloud execution time calculation formula is:

[0043]

[0044] In the formula, is the CPU cycle frequency of the corresponding ground cloud;

[0045] Step S15: Calculate the time cost of the current user offloading the computing task to the corresponding ground cloud server through the user-base station-satellite-ground cloud link for computing, and use it as the fifth time cost. The fifth time cost includes: base station-satellite two-way propagation delay , User-base station data transmission delay , Base station-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time .

[0046] Furthermore, the transmission rate between the current user and the corresponding satellite and the transmission rate between the current user and the corresponding base station These are the transmission rates in the NOMA environment;

[0047] The transmission rate calculation process between the current user and the corresponding satellite includes:

[0048]

[0049]

[0050]

[0051] In the formula, is the path loss between the current user and the corresponding satellite for communication, is the distance between the current user and the corresponding satellite, is the frequency of the communication signal, is the channel gain for the current user to communicate with the corresponding satellite, is the channel width for the current user to communicate with the corresponding satellite, is the transmission power of the current user communicating with the corresponding satellite, is the power spectral density of the noise, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding satellite and whose channel gain is greater than that of the current user;

[0052] The transmission rate calculation process between the current user and the corresponding base station includes:

[0053]

[0054]

[0055]

[0056] In the formula, is the path loss between the current user and the corresponding base station, is the reference path loss, is the path loss index, is the distance from the current user to the corresponding base station, is the reference distance; is the channel gain for the current user to communicate with the corresponding base station, is the channel width for the current user to communicate with the corresponding base station, is the transmission power of the current user communicating with the corresponding base station, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding base station and whose channel gain is greater than that of the current user.

[0057] Further, the MINLP model includes: an objective function, a path selection constraint, a time delay constraint, a power constraint, and a computing resource constraint;

[0058] The MINLP model constructs an objective function by minimizing the time cost of all computing tasks generated within a preset period of time. The objective function is:

[0059]

[0060] In the formula, The total number of computing tasks generated within the preset period; The decision variable for selecting the jth computation path for the i-th computation task, , The time cost of selecting the jth computation path for the i-th computation task;

[0061] The MINLP model uses a path selection constraint that each computing task can only select one preset computing path for computing. The path selection constraint is:

[0062]

[0063] The MINLP model uses the processing time of all computing tasks must not be greater than the maximum tolerable time as the time delay constraint, and the time delay constraint is:

[0064]

[0065] in, is the maximum tolerable time;

[0066] The MINLP model uses the power consumption of each computing task not to exceed the maximum transmission power limit as the power constraint, and the power constraint is:

[0067] ;

[0068] in, is the power consumption of the i-th computing task on the j-th computing path, is the maximum transmission power limit;

[0069] The MINLP model takes the computing resource constraint that the computing resources for executing the computing task do not exceed the total computing resources as the computing resource constraint, and the computing resource constraint is:

[0070]

[0071] in, Is the path On the task The computing resources consumed; is the total computing resources of all computing nodes.

[0072] Furthermore, the Lagrangian relaxation method and the dual problem decomposition method are used to solve the MINLP model, including:

[0073] Step S41: Based on the objective function and constraints of the MINLP model and the preliminary allocation scheme, Lagrangian multipliers are introduced to construct a Lagrangian function for the computing task offloading optimization problem;

[0074] Step S42: Decomposing the computing task offloading optimization problem into independent path optimization sub-problems, power optimization sub-problems and computing resource optimization sub-problems by using a dual problem decomposition method;

[0075] The path optimization subproblem is used to find a computing task offloading solution that minimizes the total delay of all computing tasks under fixed transmission power and computing resources;

[0076] The power optimization subproblem is used to optimize the transmission power under a fixed computing task offloading scheme. , so that it meets the power constraint;

[0077] The computing resource optimization subproblem is used to optimize the computing resource allocation so that it satisfies the computing resource constraints;

[0078] Step S43: Solve the power optimization sub-problem, computing resource optimization sub-problem and path optimization sub-problem respectively to obtain the optimal computing task offloading optimization solution.

[0079] Furthermore, the Lagrangian function is:

[0080] ;

[0081] In the formula, is the Lagrangian function, is the first multiplier, is the second multiplier;

[0082] The path optimization sub-problem is:

[0083] ;

[0084] The power optimization sub-problem is:

[0085] ;

[0086] The computing resource optimization sub-problem is:

[0087] .

[0088] Furthermore, the step S43 includes:

[0089] establishing a first rule for updating the first multiplier and a second rule for updating the second multiplier;

[0090] Initialize the first multiplier and the second multiplier;

[0091] Solving the power optimization subproblem by iteratively updating the first multiplier according to the first rule to obtain optimal power allocation;

[0092] Solving the computing resource optimization subproblem by iteratively updating the second multiplier according to the second rule to obtain optimal computing resource allocation;

[0093] Under the conditions of optimal power allocation and optimal computing resource allocation, the linear programming algorithm is used to solve the path optimization subproblem and obtain the optimal computing task offloading optimization solution.

[0094] Furthermore, the first rule is:

[0095]

[0096] In the formula, For the The first multiplier value after the iteration update, For the The first multiplier value after the iteration update, The first multiplier The step size of the iteration update;

[0097] The second rule is:

[0098]

[0099] In the formula, For the The second multiplier value obtained by iterative update is For the The second multiplier value obtained by iterative update is The second multiplier The step size of the iterative update.

[0100] Furthermore, the step length of iterative update of the first multiplier or the second multiplier is a constant;

[0101] Alternatively, the step size of the iterative update of the first multiplier or the second multiplier gradually decreases with the number of iterations.

[0102] Furthermore, the establishment of the computing task allocation model includes: building a neural network structure and setting a loss function;

[0103] The neural network structure includes: an input layer, a first hidden layer, a second hidden layer and an output layer; the input layer is used to generate input features according to input information, and input the input features into the first hidden layer; the input information includes: computing task characteristic information of each user, the time cost of each computing task under a preset computing path and network status information; the computing task characteristic information includes: task size, computing density and maximum tolerable delay; the network status information includes: the location, channel conditions and available computing resources of the satellite edge computing node SEC.

[0104] The first hidden layer uses a fully connected layer to process the input features, maps the input features to the representation space of the first hidden layer, uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a first output feature, and inputs the first output feature into the second hidden layer.

[0105] The second hidden layer uses a fully connected layer to process the first output feature, maps the first output feature to the representation space of the second hidden layer, uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a second output feature, and inputs the second output feature into the output layer.

[0106] The output layer processes the second output feature through three different output channels to obtain the path allocation, power allocation and computing resource allocation of the computing task respectively, and form a preliminary allocation plan; the path allocation channel uses the activation function Softmax to normalize the second output feature to obtain the selection probability of each computing task under each preset computing path; the power allocation channel uses the activation function ReLU to normalize the second output feature to ensure that the power under each preset computing path is non-negative; the computing resource allocation channel uses the activation function ReLU to normalize the second output feature to limit the resource allocation under each preset computing path to be non-negative.

[0107] The loss function includes: path allocation loss, power allocation loss, computing resource allocation loss and total loss;

[0108] The path allocation loss uses cross entropy loss to characterize the gap between the preliminary path allocation result and the actual path allocation result. The path allocation loss is:

[0109]

[0110] In the formula, Assign a loss to the path, For the The actual path of a computing task, For the The predicted path of a computing task;

[0111] The power allocation loss uses the mean square error MSE to characterize the gap between the predicted power allocation and the actual power allocation. The power allocation loss is:

[0112]

[0113] In the formula, is the power distribution loss, is the predicted power allocation, is the actual power allocation;

[0114] The computing resource allocation loss uses the mean square error MSE to characterize the gap between the predicted computing resource allocation and the actual resource allocation. The computing resource allocation loss is:

[0115]

[0116] In the formula, Allocate loss for computing resources, For the predicted computing resource allocation, For actual computing resource allocation;

[0117] The total loss is a combination of the path allocation loss, the power allocation loss and the computing resource allocation loss, and the total loss is:

[0118]

[0119] In the formula, is the total loss, , and is the adjustment factor.

[0120] Compared with the prior art, the present invention has the following beneficial effects:

[0121] (1) The computing task offloading optimization method considering multiple performance indicators provided by the present invention constructs a computing task allocation model by introducing a neural network, intelligently learns and predicts the computing task allocation scheme suitable for the current computing task, and uses Lagrangian functions and dual methods to handle multiple complex MINLP problems in the system. When solving the MINLP problem, the computing task allocation model provides a high-quality preliminary allocation scheme for the Lagrangian relaxation method. Compared with the traditional branch and bound method, this method reduces the search range, reduces the computational complexity, speeds up the optimization speed, improves the real-time performance of the system, meets the time-sensitive application requirements, and is particularly suitable for complex scenarios with multiple users, multiple paths, and multiple resource allocations. In addition, the present invention combines Lagrangian relaxation and dual decomposition methods to effectively balance global search and local exploration, and avoid the algorithm from falling into a local optimal solution.

[0122] (2) The computing task offloading optimization method considering multiple performance indicators provided by the present invention divides the channel into multiple subcarriers through the OMA method, and then uses the NOMA transmission method in each subcarrier to offload computing tasks, so that multiple users can share the same wireless resources. NOMA allows multiple users to transmit under the same spectrum resources and uses the power domain to distinguish users. NOMA technology can effectively manage the reception order of multi-user signals by using serial interference cancellation technology at the receiving end, which helps to reduce the waiting time for task processing and significantly improves spectrum utilization and system transmission efficiency, thereby improving system performance in an environment where spectrum resources are scarce and reducing task processing time.

[0123] (3) The computing task offloading optimization method considering multiple performance indicators provided by the present invention performs computing task offloading in complex terrain and harsh environments through low-orbit satellites, base stations, and ground cloud-assisted edge computing technology MEC, solving the problem that ground base stations cannot effectively cover. The present invention supports users to select local computing, user-satellite link to satellite edge computing node, base station-satellite link to satellite edge computing node, user-satellite link to ground cloud or base station-satellite link to ground cloud according to real-time conditions, helping users to select the best computing task offloading strategy according to their own needs and conditions, improve the utilization of computing resources, and effectively reduce the propagation delay of task offloading of satellite-borne edge clouds in complex scenarios, providing a solution with delay optimization as the goal for low-orbit satellites to select offloading methods and perform offloading processing.

[0124] (4) The computational task allocation model based on the neural network in the present invention takes into account multiple tasks, network status, power and other influencing factors, and can quickly obtain the optimal initial allocation plan. The application of neural networks enables the present invention to avoid the problems of high computational complexity and long solution time in traditional algorithms when processing a large number of complex tasks, thereby accelerating the optimization solution speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0125] Figure 1 This is a flowchart of the computing task offloading optimization method considering multiple performance indicators provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0126] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0127] Example 1

[0128] The present invention provides a computing task offloading optimization method considering multiple performance indicators. The method uses orthogonal multiple access OMA technology to divide the entire available broadband spectrum into multiple orthogonal subcarriers; in each of the orthogonal subcarriers, non-orthogonal multiple access NOMA technology is used to provide services to multiple corresponding users at the same time. Figure 1 As shown, for the computing tasks generated by the user within the preset time period, the following steps are performed:

[0129] Step 1: Calculate the time cost of each computing task under the preset computing path.

[0130] Satellite edge cloud SEC is introduced to solve the problem that users in remote areas or disaster areas have difficulty connecting to the ground edge cloud network infrastructure. It aims to provide computing offloading services for these users, enabling them to transfer computing-intensive tasks to more powerful computing resources for processing. In this scenario, two ways of user access to SEC are considered. One is that the user directly establishes communication with the LEO satellite through the C-band user-satellite radio link to achieve task offloading; the other is that the user indirectly accesses SEC through the ground base station. In this indirect access method, the user first connects to the base station through the C-band link, and then the base station transmits the task to the SEC through the Ka-band link. In addition, the LEO satellite can also connect to the cloud server of the ground network through the Ka-band backhaul link to provide cloud services to users.

[0131] In a specific embodiment, the user has multiple options for processing computing tasks. The user can perform the computing locally or offload the task to the SEC for computing. The specific methods include: directly offloading to the SEC via a user-satellite link or offloading the task to the SEC via a base station. At the same time, the user can also choose to offload the task to the ground cloud, including directly offloading to the ground cloud via a user-satellite link or offloading the task to the ground cloud via a base station-satellite link.

[0132] Step 2: Input the time cost of each computing task under the preset computing path into a pre-established computing task allocation model to generate a preliminary allocation plan.

[0133] Specifically, the computing task allocation model is established based on a neural network, including: building a neural network structure and setting a loss function; the neural network structure includes: an input layer, a first hidden layer, a second hidden layer and an output layer.

[0134] The input layer is used to generate input features according to input information, and input the input features into the first hidden layer; the input information includes: computing task characteristic information of each user, time cost of each computing task under a preset computing path and network status information; the computing task characteristic information includes: task size, computing density and maximum tolerable delay; the network status information includes: location, channel conditions and available computing resources of the satellite edge computing node SEC.

[0135] The input dimensions are set to ,in, is the number of users, is the number of features, and the input feature matrix is ​​expressed as .

[0136] The first hidden layer uses a fully connected layer to process the input features, maps the input features to the representation space of the first hidden layer, uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a first output feature, and inputs the first output feature into the second hidden layer.

[0137] Specifically, the first output feature of the first hidden layer output is for:

[0138]

[0139] in, is the weight matrix from the input layer to the first hidden layer, is the first bias.

[0140] The second hidden layer uses a fully connected layer to process the first output feature, maps the first output feature to the representation space of the second hidden layer, uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a second output feature, and inputs the second output feature into the output layer.

[0141] Specifically, the second output feature of the second hidden layer output is for:

[0142]

[0143] in, is the weight matrix from the first hidden layer to the second hidden layer, is the second bias.

[0144] The output layer processes the second output feature through three different output channels to obtain path allocation, power allocation and computing resource allocation of the computing task respectively, and form a preliminary allocation plan.

[0145] Specifically, the path allocation channel uses the activation function Softmax to normalize the second output feature to obtain the decision variable for the i-th computing task to select the j-th computing path ;

[0146]

[0147] in, The weight matrix assigned to the path, is the third bias.

[0148] The power allocation channel uses the activation function ReLU to normalize the second output feature to ensure that the power consumption of the i-th computing task on the j-th computing path is a non-negative number;

[0149]

[0150] in, is the weight matrix for power allocation, is the fourth bias.

[0151] The computing resource allocation channel uses the activation function ReLU to normalize the second output feature to limit the path On the task Computing resources consumed is a non-negative number;

[0152]

[0153] in, is the weight matrix for computing resource allocation, is the fifth bias.

[0154] The loss function includes: path allocation loss, power allocation loss, computing resource allocation loss and total loss; the path allocation loss uses cross entropy loss to characterize the gap between the preliminary path allocation result and the actual path allocation result, and the path allocation loss is:

[0155]

[0156] In the formula, Assign a loss to the path, is the total number of computing tasks generated within the preset period, For the The actual path of a computing task, For the The predicted path of a computing task;

[0157] The power allocation loss uses the mean square error MSE to characterize the gap between the predicted power allocation and the actual power allocation. The power allocation loss is:

[0158]

[0159] In the formula, is the power distribution loss, is the predicted power allocation, is the actual power allocation;

[0160] The computing resource allocation loss uses the mean square error MSE to characterize the gap between the predicted computing resource allocation and the actual resource allocation. The computing resource allocation loss is:

[0161]

[0162] In the formula, Allocate loss for computing resources, For the predicted computing resource allocation, For actual computing resource allocation;

[0163] The total loss is a combination of the path allocation loss, the power allocation loss and the computing resource allocation loss, and the total loss is:

[0164]

[0165] In the formula, is the total loss, , and is the adjustment factor.

[0166] Step 3: Establish a MINLP model for the computational task offloading optimization problem.

[0167] Specifically, the MINLP model includes: an objective function, a path selection constraint, a time delay constraint, a power constraint, and a computing resource constraint.

[0168] The MINLP model constructs an objective function by minimizing the time cost of all computing tasks generated within a preset period of time. The objective function is:

[0169]

[0170] In the formula, The total number of computing tasks generated within the preset period; The decision variable for selecting the jth computation path for the i-th computation task, , The time cost of selecting the jth computation path for the i-th computation task.

[0171] The MINLP model uses a path selection constraint that each computing task can only select one preset computing path for computing. The path selection constraint is:

[0172] ;

[0173] The MINLP model uses the processing time of all computing tasks must not be greater than the maximum tolerable time as the time delay constraint, and the time delay constraint is:

[0174]

[0175] in, is the maximum tolerable time.

[0176] The MINLP model uses the power consumption of each computing task The power constraint is not greater than the maximum transmission power limit, and the power constraint is:

[0177] ;

[0178] in, is the power consumption of the i-th computing task on the j-th computing path, is the maximum transmission power limit;

[0179] The MINLP model takes the computing resource constraint that the computing resources for executing the computing task do not exceed the total computing resources as the computing resource constraint, and the computing resource constraint is:

[0180]

[0181] in, Is the path On the task The computing resources consumed; is the total computing resources of all computing nodes.

[0182] Step 4: Using the preliminary allocation plan as the starting point for solving the MINLP model, the Lagrangian relaxation method and the dual problem decomposition method are used to solve the MINLP model to obtain the optimal computing task offloading optimization plan.

[0183] The preliminary allocation scheme obtained by the computing task allocation model is used as the starting point for solving the MINLP. The path allocation, power allocation, and computing resource allocation output by the neural network of the computing task allocation model are substituted into the Lagrangian function for constructing the computing task offloading optimization problem. Then, the dual problem decomposition method is used to decompose the global computing task offloading optimization problem into independent path optimization sub-problems, power optimization sub-problems, and computing resource optimization sub-problems. Finally, the power optimization sub-problem and computing resource optimization sub-problem are solved. According to the changes in the power and computing resource penalty terms, the path optimization strategy will be readjusted based on the new multiplier until the change in the objective function value becomes very small in multiple iterations. It can be considered that it is close to the optimal solution, and finally the best path allocation is obtained. The best path allocation is the optimal computing task offloading optimization scheme.

[0184] Example 2

[0185] Based on Example 1, this example provides five computing task calculation paths for the user to choose from and a time cost calculation method under each path.

[0186] In a specific embodiment, the time cost of the user performing calculations locally is calculated and used as the first time cost, the time cost of the user unloading the computing task to the corresponding satellite edge computing node SEC for calculation via the user-satellite link is calculated and used as the second time cost, the time cost of the user unloading the computing task to the corresponding satellite edge computing node SEC for calculation via the user-base station-satellite link is calculated and used as the third time cost, the time cost of the user unloading the computing task to the corresponding ground cloud server for calculation via the user-satellite-ground cloud link is calculated and used as the fourth time cost, and the time cost of the user unloading the computing task to the corresponding ground cloud server for calculation via the user-base station-satellite-ground cloud link is calculated and used as the fifth time cost.

[0187] In a specific embodiment, the first time cost includes: local execution time ;

[0188]

[0189] In the formula, is the CPU cycle frequency of the current user, is the current computing task data volume, The number of CPU cycles required to process each bit of data, measured in cycles / bit.

[0190] In a specific embodiment, the second time cost includes: user-satellite two-way propagation delay , User-satellite data transmission delay and satellite execution time ;in,

[0191]

[0192]

[0193]

[0194] In the formula, h is the distance between the current user and the corresponding satellite, c is the speed of light, is the transmission rate between the current user and the corresponding satellite, is the CPU cycle frequency of the corresponding satellite.

[0195] In a specific embodiment, the third time cost includes: base station-satellite two-way propagation delay (Same as user-satellite two-way propagation delay )、User-base station data transmission delay , Base station-satellite data transmission delay and satellite execution time ;in,

[0196]

[0197]

[0198] In the formula, is the transmission rate between the current user and the corresponding base station; The total amount of computing task data transmitted together with the current task through the corresponding base station to the corresponding satellite; is the transmission rate between the corresponding base station and the corresponding satellite, which is the same as the transmission rate between the current user and the corresponding satellite .

[0199] It should be noted that since the distance between the user and the base station is almost negligible compared to the distance between the user or the base station and the satellite, the two-way propagation delay between the base station and the satellite is The two-way propagation delay between the user and the satellite Similarly, the transmission rate between the base station and the corresponding satellite The value of is related to the transmission rate between the user and the corresponding satellite same.

[0200] In a specific embodiment, the fourth time cost includes: user-satellite two-way propagation delay , User-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time ;in,

[0201]

[0202]

[0203] In the formula, is the transmission rate between the satellite and the corresponding ground cloud, which is a preset constant value; is the CPU cycle frequency of the corresponding ground cloud.

[0204] In a specific embodiment, the fifth time cost includes: base station-satellite two-way propagation delay (Same as user-satellite two-way propagation delay )、User-base station data transmission delay , Base station-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time .

[0205] It should be noted that the present invention uses the orthogonal multiple access OMA technology to divide the entire available broadband spectrum into multiple orthogonal subcarriers; in each of the orthogonal subcarriers, the non-orthogonal multiple access NOMA technology is used to provide services to multiple corresponding users at the same time. Therefore, the transmission rate between the user and the corresponding satellite And the transmission rate between the user and the corresponding base station The transmission rate in the NOMA environment is used.

[0206] Specifically, the transmission rate between the user and the corresponding satellite The calculation process includes:

[0207]

[0208]

[0209]

[0210] In the formula, is the path loss between the current user and the corresponding satellite for communication, is the distance between the current user and the corresponding satellite, is the frequency of the communication signal, is the channel gain for the current user to communicate with the corresponding satellite, is the channel width for the current user to communicate with the corresponding satellite, is the transmission power of the current user communicating with the corresponding satellite, is the power spectral density of the noise, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding satellite and whose channel gain is greater than that of the current user.

[0211] The transmission rate between the user and the corresponding base station The calculation process includes:

[0212]

[0213]

[0214]

[0215] In the formula, is the path loss between the current user and the corresponding base station, is the reference path loss, is the path loss index, is the distance from the current user to the corresponding base station, is the reference distance; is the channel gain for the current user to communicate with the corresponding base station, is the channel width for the current user to communicate with the corresponding base station, is the transmission power of the current user communicating with the corresponding base station, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding base station and whose channel gain is greater than that of the current user.

[0216] Example 3

[0217] Based on Example 1, this example provides a specific process of solving the MINLP model using the Lagrangian relaxation method and the dual problem decomposition method.

[0218] The MINLP model is solved by using the Lagrangian relaxation method and the dual problem decomposition method, including the following steps:

[0219] Step S41: Based on the objective function and constraints of the MINLP model and the preliminary allocation scheme, Lagrangian multipliers are introduced to construct a Lagrangian function for the computational task offloading optimization problem.

[0220] In a specific embodiment, the Lagrangian function is:

[0221] ;

[0222] In the formula, is the Lagrangian function, is the first multiplier, is the second multiplier.

[0223] Step S42: using a dual problem decomposition method to decompose the computing task offloading optimization problem into independent path optimization sub-problems, power optimization sub-problems and computing resource optimization sub-problems.

[0224] In a specific embodiment, the path optimization subproblem is used to find a computing task offloading solution that minimizes the total delay of all computing tasks under fixed transmission power and computing resources; the path optimization subproblem is:

[0225] .

[0226] In a specific embodiment, the power optimization subproblem is used to optimize the transmission power under a fixed computing task offloading scheme. , so that it satisfies the power constraint; the power optimization sub-problem is:

[0227] .

[0228] In a specific embodiment, the computing resource optimization sub-problem is used to optimize the computing resource allocation so that it satisfies the computing resource constraints; the computing resource optimization sub-problem is:

[0229] .

[0230] Step S43: Solve the power optimization sub-problem, computing resource optimization sub-problem and path optimization sub-problem respectively, and then obtain the optimal computing task offloading optimization solution, which specifically includes:

[0231] Step S431: establishing a first rule for updating the first multiplier and a second rule for updating the second multiplier.

[0232] In a specific embodiment, the first rule is:

[0233]

[0234] In the formula, For the The first multiplier value after the iteration update, For the The first multiplier value after the iteration update, The first multiplier The step size of the iterative update.

[0235] The second rule is:

[0236]

[0237] In the formula, For the The second multiplier value obtained by iterative update is For the The second multiplier value obtained by iterative update is The second multiplier The step size of the iterative update.

[0238] Step S432: Initialize the first multiplier and the second multiplier.

[0239] In a specific embodiment, the initial values ​​of the first multiplier and the second multiplier are both 0.

[0240] Step S433: solving the power optimization sub-problem by iteratively updating the first multiplier according to the first rule to obtain optimal power allocation.

[0241] In a specific embodiment, the first multiplier of the current iteration is the product of the step size and the degree of power constraint violation according to the first multiplier value of the previous round, where the degree of power constraint violation is:

[0242]

[0243] When the power exceeds the maximum value, the first multiplier increases, thus driving the optimization algorithm to reduce power consumption.

[0244] Step S434: solving the computing resource optimization sub-problem by iteratively updating the second multiplier according to the second rule to obtain optimal computing resource allocation.

[0245] In a specific embodiment, the second multiplier of the current iteration is calculated based on the second multiplier value of the previous round plus the product of the step length and the degree of violation of the computing resource constraint, where the degree of violation of the computing resource constraint is:

[0246]

[0247] If the allocated computational resources exceed the maximum available resources, the second multiplier will increase, pushing the algorithm to reduce the computational resource allocation for the task.

[0248] Step S435: Under the conditions of optimal power allocation and optimal computing resource allocation, a linear programming algorithm is used to solve the path optimization sub-problem to obtain an optimal computing task offloading optimization solution.

[0249] In a specific embodiment, according to the changes in the degree of power and computing resource default, the path optimization strategy will be readjusted based on the new multiplier until the change in the objective function value is less than the preset threshold in multiple iterations. It is then considered that the optimal solution has been approached, and the optimal computing task offloading optimization solution has been obtained.

[0250] In a specific embodiment, the step size of iterative update of the first multiplier or the second multiplier is a constant.

[0251] In a specific embodiment, the step size of iterative update of the first multiplier or the second multiplier gradually decreases with the number of iterations.

[0252] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A computing task offloading optimization method considering multiple performance indicators, characterized in that: The entire available broadband spectrum is divided into multiple orthogonal subcarriers using orthogonal multiple access OMA technology; in each of the orthogonal subcarriers, non-orthogonal multiple access NOMA technology is used to provide services to multiple corresponding users at the same time, and the following steps are performed for the computing tasks generated by the users within the preset time period: Step S1: Calculate the time cost of each computing task under the preset computing path respectively; Step S2: inputting the time cost of each computing task under the preset computing path into a pre-established computing task allocation model to generate a preliminary allocation plan; wherein the computing task allocation model is established based on a neural network; Step S3: Establishing a MINLP model for the computational task offloading optimization problem; Step S4: using the preliminary allocation scheme as the starting point for solving the MINLP model, solving the MINLP model using the Lagrangian relaxation method and the dual problem decomposition method, and obtaining the optimal computing task offloading optimization scheme; Wherein, the step S1 comprises: Step S11: Calculate the time cost of the current user's local calculation as the first time cost, the first time cost includes: local execution time ; The local execution time calculation formula is: ; In the formula, is the CPU cycle frequency of the current user, is the current computing task data volume, The number of CPU cycles required to process each bit of data, in cycles / bit; Step S12: Calculate the time cost of the current user offloading the computing task to the corresponding satellite edge computing node SEC through the user-satellite link for computing, and use it as the second time cost. The second time cost includes: the user-satellite two-way propagation delay , User-satellite data transmission delay and satellite execution time ; The calculation formula for the user-satellite two-way propagation delay is: ; In the formula, h is the distance between the current user and the corresponding satellite, and c is the speed of light; The user-satellite data transmission delay calculation formula is: ; In the formula, is the transmission rate between the current user and the corresponding satellite; The satellite execution time calculation formula is: ; In the formula, is the CPU cycle frequency of the corresponding satellite; Step S13: Calculate the time cost of the current user offloading the computing task to the corresponding satellite edge computing node SEC through the user-base station-satellite link for computing, and use it as the third time cost. The third time cost includes: base station-satellite two-way propagation delay , User-base station data transmission delay , Base station-satellite data transmission delay and satellite execution time ; Wherein, the base station-satellite two-way propagation delay The two-way propagation delay between the user and the satellite same; The calculation formula for the user-base station data transmission delay is: ; In the formula, is the transmission rate between the current user and the corresponding base station; The base station-satellite data transmission delay calculation formula is: ; In the formula, The total amount of computing task data transmitted to the corresponding satellite through the base station corresponding to the current task; is the transmission rate between the corresponding base station and the corresponding satellite, which is the same as the transmission rate between the current user and the corresponding satellite ; Step S14: Calculate the time cost of the current user offloading the computing task to the corresponding ground cloud server through the user-satellite-ground cloud link for computing, and use it as the fourth time cost. The fourth time cost includes: the user-satellite two-way propagation delay , User-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time ; The satellite-to-ground cloud data transmission delay calculation formula is: ; In the formula, is the transmission rate between the satellite and the corresponding ground cloud, which is a preset constant value; The ground cloud execution time calculation formula is: ; In the formula, is the CPU cycle frequency of the corresponding ground cloud; Step S15: Calculate the time cost of the current user offloading the computing task to the corresponding ground cloud server through the user-base station-satellite-ground cloud link for computing, and use it as the fifth time cost. The fifth time cost includes: base station-satellite two-way propagation delay , User-base station data transmission delay , Base station-satellite data transmission delay , Satellite-ground cloud data transmission delay and ground cloud execution time ; The MINLP model includes: an objective function, a path selection constraint, a time delay constraint, a power constraint, and a computing resource constraint; The MINLP model constructs an objective function by minimizing the time cost of all computing tasks generated within a preset period of time. The objective function is: ; In the formula, The total number of computing tasks generated within the preset period; The decision variable for selecting the jth computation path for the i-th computation task, , The time cost of selecting the jth computation path for the i-th computation task; The MINLP model uses a path selection constraint that each computing task can only select one preset computing path for computing. The path selection constraint is: ; The MINLP model uses the processing time of all computing tasks must not be greater than the maximum tolerable time as the time delay constraint, and the time delay constraint is: ; in, is the maximum tolerable time; The MINLP model uses the power consumption of each computing task not to exceed the maximum transmission power limit as the power constraint, and the power constraint is: ; in, is the power consumption of the i-th computing task on the j-th computing path, is the maximum transmission power limit; The MINLP model takes the computing resource constraint that the computing resources for executing the computing task do not exceed the total computing resources as the computing resource constraint, and the computing resource constraint is: ; in, Is the path On the task The computing resources consumed; is the total computing resources of all computing nodes; The Lagrangian relaxation method and the dual problem decomposition method are used to solve the MINLP model, including: Step S41: Based on the objective function and constraints of the MINLP model and the preliminary allocation scheme, Lagrangian multipliers are introduced to construct a Lagrangian function for the computing task offloading optimization problem; The Lagrangian function is: ; In the formula, is the Lagrangian function, is the first multiplier, is the second multiplier; Step S42: Decomposing the computing task offloading optimization problem into independent path optimization sub-problems, power optimization sub-problems and computing resource optimization sub-problems by using a dual problem decomposition method; The path optimization subproblem is used to find a computing task offloading solution that minimizes the total delay of all computing tasks under fixed transmission power and computing resources; The path optimization sub-problem is: ; The power optimization subproblem is used to optimize the transmission power under a fixed computing task offloading scheme. , so that it meets the power constraint; The power optimization sub-problem is: ; The computing resource optimization subproblem is used to optimize the computing resource allocation so that it satisfies the computing resource constraints; The computing resource optimization sub-problem is: ; Step S43: Solve the power optimization sub-problem, computing resource optimization sub-problem and path optimization sub-problem respectively to obtain the optimal computing task offloading optimization solution, including: establishing a first rule for updating the first multiplier and a second rule for updating the second multiplier; The first rule is: ; In the formula, For the The first multiplier value after the iteration update, For the The first multiplier value after the iteration update, The first multiplier The step size of the iteration update; The second rule is: ; In the formula, For the The second multiplier value obtained by iterative update is For the The second multiplier value obtained by iterative update is The second multiplier The step size of the iteration update; Initialize the first multiplier and the second multiplier; Solving the power optimization subproblem by iteratively updating the first multiplier according to the first rule to obtain optimal power allocation; Solving the computing resource optimization subproblem by iteratively updating the second multiplier according to the second rule to obtain optimal computing resource allocation; Under the conditions of optimal power allocation and optimal computing resource allocation, the linear programming algorithm is used to solve the path optimization subproblem and obtain the optimal computing task offloading optimization solution.

2. The computing task offloading optimization method considering multiple performance indicators according to claim 1 is characterized in that: The transmission rate between the current user and the corresponding satellite and the transmission rate between the current user and the corresponding base station These are the transmission rates in the NOMA environment; The transmission rate calculation process between the current user and the corresponding satellite includes: ; ; ; In the formula, is the path loss between the current user and the corresponding satellite for communication, is the distance between the current user and the corresponding satellite, is the frequency of the communication signal, is the channel gain for the current user to communicate with the corresponding satellite, is the channel width for the current user to communicate with the corresponding satellite, is the transmission power of the current user communicating with the corresponding satellite, is the power spectral density of the noise, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding satellite and whose channel gain is greater than that of the current user; The transmission rate calculation process between the current user and the corresponding base station includes: ; ; ; In the formula, is the path loss between the current user and the corresponding base station, is the reference path loss, is the path loss index, is the distance from the current user to the corresponding base station, is the reference distance; is the channel gain for the current user to communicate with the corresponding base station, is the channel width for the current user to communicate with the corresponding base station, is the transmission power of the current user communicating with the corresponding base station, is the channel interference term, which is the sum of the products of the transmission power and the channel gain of all users connected to the corresponding base station and whose channel gain is greater than that of the current user.

3. The computing task offloading optimization method considering multiple performance indicators according to claim 2 is characterized in that: The step length of iterative update of the first multiplier or the second multiplier is a constant; Alternatively, the step size of the iterative update of the first multiplier or the second multiplier gradually decreases with the number of iterations.

4. The computing task offloading optimization method considering multiple performance indicators according to claim 1 is characterized in that: The establishment of the computing task allocation model includes: building a neural network structure and setting a loss function; The neural network structure includes: an input layer, a first hidden layer, a second hidden layer and an output layer; The input layer is used to generate input features according to input information, and input the input features into the first hidden layer; The input information includes: computing task characteristic information of each user, time cost of each computing task under a preset computing path, and network status information; The computing task characteristic information includes: task size, computing density and maximum tolerable delay; The network status information includes: the location, channel conditions and available computing resources of the satellite edge computing node SEC; The first hidden layer uses a fully connected layer to process the input feature, maps the input feature to the representation space of the first hidden layer, and uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a first output feature, and inputs the first output feature into the second hidden layer; The second hidden layer uses a fully connected layer to process the first output feature, maps the first output feature to the representation space of the second hidden layer, and uses an activation function ReLU to perform nonlinear processing on the output of the fully connected layer to obtain a second output feature, and inputs the second output feature into the output layer; The output layer processes the second output feature through three different output channels to obtain path allocation, power allocation and computing resource allocation of computing tasks respectively, and forms a preliminary allocation plan; The path allocation channel uses the activation function Softmax to normalize the second output feature to obtain the selection probability of each computing task under each preset computing path; The power allocation channel uses the activation function ReLU to normalize the second output feature to ensure that the power under each preset calculation path is non-negative; The computing resource allocation channel uses the activation function ReLU to normalize the second output feature to limit the resource allocation under each preset computing path to be a non-negative number; The loss function includes: path allocation loss, power allocation loss, computing resource allocation loss and total loss; The path allocation loss uses cross entropy loss to characterize the gap between the preliminary path allocation result and the actual path allocation result. The path allocation loss is: ; In the formula, Assign a loss to the path, For the The actual path of a computing task, For the The predicted path of a computing task; The total number of computing tasks generated within the preset period; The power allocation loss uses the mean square error MSE to characterize the gap between the predicted power allocation and the actual power allocation. The power allocation loss is: ; In the formula, is the power distribution loss, is the predicted power allocation, is the actual power allocation; The computing resource allocation loss uses the mean square error MSE to characterize the gap between the predicted computing resource allocation and the actual resource allocation. The computing resource allocation loss is: ; In the formula, Allocate loss for computing resources, For the predicted computing resource allocation, For actual computing resource allocation; The total loss is a combination of the path allocation loss, the power allocation loss and the computing resource allocation loss, and the total loss is: ; In the formula, is the total loss, , and is the adjustment factor.

Images