Calculation unloading and resource allocation method for edge cooperation dependent task

Abstract

The invention discloses a computing unloading and resource allocation method for an edge cooperative dependent task, and belongs to the field of communication, the method comprises the following steps: constructing an edge network architecture with a task dependency relationship, the edge network architecture comprising a plurality of Internet of Things devices and a plurality of base stations carrying mobile edge computing servers; performing mathematical modeling on task dependency based on an edge network structure, and determining a dependency relationship between sub-tasks in the Internet of Things equipment; determining a subtask unloading decision of the Internet of Things equipment based on the dependency relationship, and determining an execution mode of the task based on the subtask unloading decision; based on the execution mode of the task, calculating communication delay and waiting delay of the task to obtain total delay of the task; on the basis of task total time delay, a depth deterministic strategy gradient algorithm is adopted, and unloading decisions, transmitting power and local computing resource allocation of subtasks are jointly optimized so as to minimize energy consumption and time delay weighted sum of all terminal devices.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and particularly relates to a method for computing offloading and resource allocation for edge collaboration-dependent tasks. Background Art

[0002] The large-scale growth and diversification of various emerging applications in wireless networks, such as intelligent transportation systems, virtual reality, and augmented reality, have triggered a large demand for terminal access and computing resources. To relieve the traffic load on the core network and reduce the service latency of applications, mobile edge computing technology has been widely adopted. In an MEC system, computing tasks on terminals can be transferred to edge servers for execution, where it is crucial to jointly optimize computing offloading and resource management to meet the requirements of performance-demanding tasks, ensuring the user experience while avoiding excessive network overhead and traffic load.

[0003] Many existing works and studies are carried out under the assumption that computing tasks are independent of each other. However, in actual industrial edge networks, the dependency relationships between tasks generated by terminal devices need to be considered, that is, terminal devices often need to execute multiple related tasks, where the input of one task requires the output of another. Since the execution is coupled between tasks, the optimization design is much more difficult than the case where independent tasks can be executed in parallel. Based on the above technical problems, the present invention proposes a method for computing offloading and resource allocation for edge collaboration-dependent tasks. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a method for computing offloading and resource allocation for edge collaboration-dependent tasks to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides a method for computing offloading and resource allocation for edge collaboration-dependent tasks, including:

[0006] Constructing an edge network architecture with task dependency relationships, where the edge network architecture includes several Internet of Things devices and several base stations equipped with mobile edge computing servers;

[0007] Mathematically modeling task dependencies based on the edge network structure to determine the dependency relationships between subtasks in the Internet of Things devices;

[0008] Determining subtask offloading decisions for the Internet of Things devices based on the dependency relationships, and determining the execution mode of tasks based on the subtask offloading decisions;

[0009] Calculating the communication delay and waiting delay of tasks based on the execution mode of tasks to obtain the total task delay;

[0010] Based on the total task latency, the Deep Deterministic Policy Gradient algorithm is adopted to jointly optimize the offloading decision, transmission power, and local computing resource allocation of subtasks to minimize the weighted sum of energy consumption and latency of all terminal devices.

[0011] Optionally, the process of constructing an edge network architecture with task dependencies includes:

[0012] Determine the number and location of Internet of Things (IoT) devices, as well as the number and location of base stations;

[0013] Based on the number and location of the IoT devices and the number and location of the base stations, establish an association relationship between the IoT devices and the base stations; where each IoT device is associated with only one base station;

[0014] Define the set of subtasks of the IoT devices and the dependencies between the subtasks.

[0015] Optionally, the calculation process of the communication latency includes:

[0016] According to the offloading decision of the subtasks of the IoT devices, determine the transmission path of the task data;

[0017] Calculate the communication latency under different transmission paths, where the communication latency includes the uplink transmission link delay and the downlink transmission link delay;

[0018] Among them, the calculation expression of the communication latency is:

[0019]

[0020] In the formula, represents the communication latency of the i-th task of the k IoT devices, x i,k represents whether the i-th task of the k-th IoT device selects local computing or offloads to the edge server for computing, represents the uplink transmission latency between the k-th IoT device and the edge service, represents the downlink transmission latency between the k-th IoT device and the edge service.

[0021] Optionally, the expression for calculating the waiting latency is:

[0022]

[0023] In the formula, represents the waiting latency of the i-th task of the k IoT devices for the j-th task of the m IoT devices, represents the dependency relationship of the i-th task of the k IoT devices for the j-th task of the m IoT devices, O j,m represents the output data of the j-th task of the m IoT devices, Denote the time required for m Internet of Things devices to run to task j. Denote the output data O of the j-th task among m Internet of Things devices. j,m The uplink transmission time for transmitting to the associated MEC nm. Denote the time for the associated MEC nk to return the output data O of the j-th task among m Internet of Things devices to k Internet of Things devices. j,m Time. Denote the time for sending O between MEC nm and MEC nk. j,m Time.

[0024] Optionally, the process of joint optimization includes:

[0025] Pre-optimize the offloading decision using the Gibbs sampling algorithm, select the offloading strategy according to the probability distribution, and obtain the optimized offloading strategy.

[0026] Based on the optimized offloading strategy, optimize the transmit power and local computing resources through the DDPG algorithm to obtain the minimum weighted sum of energy consumption and delay of all terminal devices.

[0027] Optionally, in the process of selecting the offloading strategy according to the probability distribution, the calculation expression of the probability distribution is:

[0028]

[0029] In the formula, x k Denote the offloading strategy selected according to the probability distribution. Denote the probability distribution, θ represents the θ-th sampling, and x′ k Denote the sampling set of the subtask offloading decision in the k-th device. Denote the sampling set generated from each Internet of Things device. Denote the optimal value of the objective function under the current given offloading decision.

[0030] Optionally, the process of jointly optimizing the offloading decision, transmit power, and local computing resource allocation of subtasks using the deep deterministic policy gradient algorithm includes:

[0031] Initialize the Critic network, Actor network, and network parameters.

[0032] Determine the state of the current Internet of Things device based on the state space, and based on the state of the current Internet of Things device, select the transmit transmission and computing resources and add noise to obtain the current action.

[0033] Obtain the reward and the next state based on the current action.

[0034] Store the status, current actions, reward, and next state of the current IoT device in the experience pool;

[0035] Extract data chunks from the experience pool for training, update the Critic network and Actor network, and obtain optimized transmission power and local computing resources.

[0036] Optionally, the state space is a set of subtask offloading status, bandwidth allocation, MEC relevance, task dependency matrix, edge server computing resource allocation, and the weighted sum of current latency and energy consumption.

[0037] Compared with the prior art, the present invention has the following advantages and technical effects:

[0038] The present invention proposes a method for computing offloading and resource allocation of edge collaboration-dependent tasks. By constructing an edge network architecture with task dependencies, it mathematically models the subtask dependencies in IoT devices, and then determines the offloading decisions and execution methods of subtasks. On this basis, it accurately calculates the communication latency and waiting latency of tasks to obtain the total task latency. Finally, it uses the deep deterministic policy gradient algorithm to jointly optimize the offloading decisions, transmission power, and local computing resource allocation of subtasks to minimize the weighted sum of energy consumption and latency of all terminal devices. This method can effectively solve the optimization problems faced in dealing with task dependencies in the prior art, significantly reduce the energy consumption and latency of the system, improve the resource utilization efficiency, and enhance the user experience. Through the introduction of deep learning algorithms, it realizes dynamic optimization in complex network environments, enabling the system to adaptively adjust resource allocation strategies to cope with changing network conditions and task requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0040] Figure 1 It is an industrial edge network architecture with task dependencies according to an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of the time of a dependency-based task according to an embodiment of the present invention;

[0042] Figure 3 It is an algorithm for computing offloading and resource allocation of industrial edge network dependency-based tasks according to an embodiment of the present invention;

[0043] Figure 4 It is a simulation model of a dependency-based task according to an embodiment of the present invention;

[0044] Figure 5For the convergence of the cumulative reward of the algorithm in the embodiments of the present invention;

[0045] Figure 6 For the relationship between the weighted sum of the system delay and energy consumption and the server computing resources in the embodiments of the present invention;

[0046] Figure 7 For the relationship between the weighted sum of the system delay and energy consumption and the bandwidth B in the embodiments of the present invention. Detailed implementation manners

[0047] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will detail this application with reference to the drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowchart of the drawings may be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than here.

[0049] Embodiment 1

[0050] As Figure 3 shown, the present invention proposes a task offloading and resource allocation scheme for industrial Internet of Things scenarios with task dependencies, aiming to minimize the weighted sum of the energy consumption and delay of all devices in the industrial Internet of Things scenario where the combination nature of offloading decisions between tasks and the strong coupling with resource allocation exist, so as to meet the requirements of different industrial system benefits. The method includes the following steps:

[0051] Construct an edge network architecture with task dependencies, where the edge network architecture includes several Internet of Things devices and several base stations equipped with mobile edge computing servers; perform mathematical modeling on task dependencies based on the edge network structure to determine the dependency relationships between subtasks in the Internet of Things devices; determine the subtask offloading decisions of the Internet of Things devices based on the dependency relationships, and determine the execution manner of the tasks based on the subtask offloading decisions; calculate the communication delay and waiting delay of the tasks based on the execution manner of the tasks to obtain the total task delay; based on the total task delay, adopt the deep deterministic policy gradient algorithm to jointly optimize the offloading decisions, transmission power, and local computing resource allocation of the subtasks to minimize the weighted sum of the energy consumption and delay of all terminal devices.

[0052] As a specific implementation manner of this embodiment, as Figure 1 shown, the first part provides an industrial edge network architecture with task dependencies, and this architecture includes the following steps:

[0053] Step 1: Based on the traditional edge network architecture, the system consists of K IIoT devices, denoted as TN k, and N base stations equipped with MEC servers, denoted as MEC n, represented by the sets and respectively. Each TN can only be associated with one MEC, indicating that TN k is associated with MEC n, and the subtasks in TN k can only be offloaded to MEC n.

[0054] Furthermore, the process of constructing an edge network architecture with task dependencies includes: determining the number and location of IoT devices, and the number and location of base stations; establishing the association relationship between IoT devices and base stations based on the number and location of IoT devices and the number and location of base stations; where each IoT device is only associated with one base station; defining the set of subtasks of IoT devices and the dependency relationship between subtasks.

[0055] Step 2: As shown in Figure 2 , mathematically model the dependencies between tasks. Assume that TN k has I k subtasks to be executed in sequence, and each subtask requires the output of the previous subtask. When there are dependency relationships between different IIoT devices, one TN requires the output of multiple subtasks from TN. The tasks of each TN can be represented by {L i,k , I i,k , O i,k , C i,k}, where i ∈ {0, 1,..., I k + 1}, L i,k represents the data size of the i-th subtask of TN k, I i,k represents the size of the pre-data of the i-th subtask of TN k, O i,k represents the size of the output data of the i-th subtask of TN k, where L i,k , I i,k , O i,k are all in bits, and C i,k represents the number of CPU cycles required to complete one bit of the task, in CPU cycles / bit.

[0056] Step 3: Each TN has I k subtasks to be executed in sequence. Denote as indicating that the i-th task of TN k has a dependency relationship with the j-th task of TN m, that is, the i-th subtask of TN k not only requires the output of its own previous subtask, but also requires the output of the j-th subtask of TN m.

[0057] Step 4: Each TN can choose local computing or offloading to the edge server for computing. Denote xi,k = 1 indicates that the i-th task of TN k is offloaded to the edge server for execution, x i,k = 0 indicates that the i-th task of TN k is locally computed on TN k.

[0058] As a specific implementation of this embodiment, the second part provides a general method for calculating the delay of dependent tasks in this embodiment, including the following specific steps:

[0059] Step 1: When the i-th subtask of TN k has a dependency on the j-th subtask of TN m, and TN m is associated with MECnm and TN k is associated with MEC nk, the edge network model between TNs can be specifically divided into 4 cases, as follows: (1) The j-th task of TN m and the i-th task of TN k are both executed locally. In this case, x j,m = 0, x i,k = 0, the MEC server acts as a relay node. TN m first uploads the output O of its j-th task j,m to the associated MEC nm, then performs information interaction between servers, and finally MEC nk forwards O j,m to TN k; (2) The j-th task of TN m is executed on the edge server, and the i-th task of TN k is executed locally. In this case, x j,m = 1, x i,k = 0. After the j-th task of TN m is computed on the edge server, it is directly forwarded to MEC nk, and then further forwarded by MEC nk to TN k; (3) The j-th subtask of TN m is executed locally, and the i-th subtask of TN k is executed on the edge server. In this case, x j,m = 0, x i,k = 1. TN m needs to upload the output O of its j-th subtask j,m to MEC nm and perform information interaction between servers, but MEC nk does not need to send data to TN k; (4) The j-th task of TN m and the i-th task of TN k are both executed on the edge server. In this case, x j,m = 1, x i,k = 1, and the task data O j,m is only transmitted between the two servers and does not require uplink or downlink transmission.

[0060] Step 2: Through the offloading decision x i,k represents the general communication delay (1) When x i-1,k = x i,k , the two subtasks are processed on the same device, and there is no communication delay; (2) When x i-1,k = 0, xi,k When x = 1, there is an uplink transmission link delay; (3) When x i-1,k = 1, x i,k = 0, there is a downlink transmission link delay.

[0061] The calculation process of the communication delay includes: determining the transmission path of the task data according to the sub-task offloading decision of the Internet of Things device; calculating the communication delay under different transmission paths, and the communication delay includes the uplink transmission link delay and the downlink transmission link delay. In summary, for the communication delay of the i-th task of TN k The calculation formula is:

[0062]

[0063] Step 3: Implementation of the data collection module: After determining the data reporting period of the Internet of Things device, the Internet of Things device regularly reports information to the data collection module of the control platform. After the data collection module collects the data information, it stores it in the information database for the data analysis module to call. For an industrial edge network with task dependencies between terminals, the waiting delay of the i-th task of TN k for the task it depends on also needs to be considered. Taking the example that the i-th task of TN k has a dependency relationship with the j-th task of TN m, when is the case, the waiting delay of the i-th task of TN k for the j-th task of TN m is denoted as It can be divided into four parts: (1) The time required for TN m to run to task j; (2) The uplink transmission time for the j-th task of TN m to output data O j,m transmitted to the associated MEC nm, if the j-th task of TN m selects offloading calculation, this item is 0; (3) The time for sending O j,m between MEC nm and MEC nk, if the MEC associated with TN k and TN m is the same, this item is 0; (4) The downlink transmission time for the MEC nk associated with TN k to transmit O j,m to it, if the i-th task of TN k selects offloading calculation, this item is 0. Then the waiting delay of the i-th task of TN k for the j-th task of TN m is calculated as follows:

[0064]

[0065] Among them, represents the dependency relationship between the i-th sub-task of TN k and the j-th sub-task of TN m; represents the time required for TN m to run to task j; represents the output data O of the j-th task of TN m j,m transmitted to the associated MEC n mUplink transmission time; Indicates the associated MEC n k Return the j-th task output data O of TN m to TN k j,m Time; Indicates MEC n m Between MEC n k Send O between j,m Time.

[0066] Step 4: Combining Step 2 and Step 3, for the i-th task delay T of TN k i,k Can be obtained by comparing two parts. The first part is the sum of the computing delay of the i-th task of TN k and the communication transmission delay between the i-th task of TN k and the previous subtask. The second part is the waiting delay of the i-th subtask of TN k for its dependent tasks. The value of the i-th task delay of TN k is the larger of the two, then T i,k The calculation formula is as follows:

[0067]

[0068] Among them, Represents the computing delay of the i-th task of TN k; Represents the communication transmission delay between the i-th task of TN k and the previous subtask; Is the waiting delay of the i-th subtask of TN k for its dependent tasks.

[0069] As a specific implementation manner of this embodiment, the third part provides an algorithm for computing offloading and resource allocation of dependent tasks in an industrial edge network, including the following specific steps:

[0070] Step 1: Based on the above edge network architecture and general delay calculation method, for industrial edge networks with different subtask dependencies between TNs, under the constraints of device transmission power, computing and communication resources, considering task dependencies and cross-regional MEC collaboration issues, jointly optimize the subtask offloading decisions, transmission power, and local CPU computing frequencies of each terminal device to minimize the weighted sum of terminal device computing task delays and energy consumption.

[0071] Step 2: When solving the problem of minimizing the weighted sum of terminal device computing task delays and energy consumption, handle integer variables and continuous variables separately. First, handle integer variables, that is, offloading decision variables. To reduce the computational complexity, it can be observed that when At this time, when each TN performs at most one offloading, the optimal value of the objective function can be obtained.

[0072] Step 3: Based on Step 2, assuming that the continuous variables, i.e., the transmission power and local computing resources, are known at this time, the Gibbs sampling algorithm is used to further reduce the computational complexity while obtaining the optimal offloading decision. The offloading decision x of each TN k In the θ-th sampling, according to the probability distribution is updated to Λ k The expression is as follows:

[0073]

[0074] where x -k =(x1,...,x k-1 ,x k+1 ,...,x K ), represents the optimal objective value given the offloading decision, T is a parameter greater than zero, θ represents the θ-th sampling, and x′ k represents the sampling set of the sub-task offloading decision in the k-th device, represents the sampling set generated from each Internet of Things device, represents the optimal value of the objective function under the current given offloading decision. This step can select the offloading decision x corresponding to the smaller objective function value with a higher probability k .

[0075] Step 4: Taking the offloading decision x k obtained in Step 3 as a known condition, optimize the transmission power and local computing frequency. In the dependent task scenario after the offloading decision is given, there are still multiple TNs and MEC servers involved, and the power and computing frequency of each device still need to be jointly optimized. Moreover, there are temporal and logical coupling relationships among the dependent tasks, resulting in dynamic changes in the system state. This high-dimensional state and action space make it difficult for traditional optimization methods, such as convex optimization or heuristic algorithms, to handle efficiently. Therefore, the DDPG algorithm is introduced. Through the function approximation ability of the deep neural network, it can effectively handle high-dimensional inputs and outputs, adapt to complex system models, and explore the global optimal solution in the non-convex space through policy gradients and value function approximation. At the same time, the training process is stabilized through experience replay and target networks.

[0076] Furthermore, the process of joint optimization includes: using the Gibbs sampling algorithm to pre-optimize the offloading decision, selecting the offloading strategy according to the probability distribution, and obtaining the optimized offloading strategy; based on the optimized offloading strategy, optimizing the transmission power and local computing resources through the DDPG algorithm to obtain the minimum weighted sum of energy consumption and delay of all terminal devices.

[0077] Step 5: State s: The state s is used to characterize the specific situation of the entire system at the current moment. For the problem proposed in, given the offloading decision xk Under the dependent task model, the state space is designed as follows: where x k represents the set of offloading situations of subtasks observed by the current TN, B k represents the set of transmission bandwidths allocated to each subtask of the current TN; a k represents the correlation between the current TN and the MEC server, α k represents the dependency relationship matrix of each subtask of the current TN; represents the set of computing resources allocated by the edge server to each subtask; η k represents the weighted sum of the delay and energy consumption of the current TN. When implementing the DDPG algorithm, neural networks are used to approximate the value function and the policy function. To ensure the stability and convergence of neural network training, all state variables need to be normalized.

[0078] Step 6: Action a: Action a is the set of subtask allocation strategies of the current TN. The definition of action a is as follows: a = {p k , f k}, where p k represents the set of transmission powers of each subtask in the current TN; f k represents the set of local computing resources allocated to each subtask by the current TN. Since the transmission power and computing resources of the TN are limited, the range of the control input action size needs to satisfy the constraint 0 ≤ p i,k ≤ P max . To be able to explore the environment as much as possible and also make the algorithm converge stably in the end, a gradually decreasing interference noise is added to action a, and this noise follows a normal distribution. Similar to the state space variables, action a after adding the noise interference should also be normalized.

[0079] Step 7: Reward r: The reward is used as an index to evaluate the quality of actions and greatly affects the performance of the model training network. After the action is executed, what changes will occur in the environment? To solve the joint optimization problem of power p and computing resource f k under the given offloading decision situation x in k , a reward function combining positive incentives and negative penalties needs to be constructed. The specific reward definition is as follows: r = λ1I k + λ2η k . Among them, both λ1 and λ2 are constants, λ1 is positive, encouraging the agent to select p k and f kAfter that, guide the agent to explore effective strategies and avoid falling into local optima; λ2 is negative to minimize the weighted sum of the delay and energy consumption of the current TN.

[0080] Step 8: The joint optimization algorithm of power and computing resources based on DDPG includes the following steps:

[0081] Step 9: Initialization: The Critic network Q(s,a;θ Q ) and the Actor network μ(s;θ u ), as well as the parameters of the Critic and Actor networks, the target networks Q' and μ', the replay experience pool Buffer, set the soft update learning rate τ and the discount factor γ

[0082] Step 10: Determine the state s of the current TN, Select the transmission power p k and the computing resource f k and add noise to obtain the current action a of the TN. Obtain the reward r and the next state s' according to the current action a, store (s,a,r,s') in the experience pool Buffer, extract a batch of data from the experience pool for training, calculate y = r + γQ'(s',μ'(s',θ u' );θ Q' ) according to the target network, minimize the Loss value to update the Critic network, update the policy with the sampled policy gradient, and update the target network. Repeat this step until the result converges, and obtain the optimal transmission power and local computing resources under the current offloading decision.

[0083] Furthermore, the process of jointly optimizing the offloading decision, transmission power, and local computing resource allocation of subtasks using the deep deterministic policy gradient algorithm includes: initializing the Critic network, Actor network, and network parameters; determining the state of the current Internet of Things device based on the state space, and based on the state of the current Internet of Things device, selecting the transmission and computing resources and adding noise to obtain the current action; obtaining the reward and the next state based on the current action; storing the state, current action, reward, and next state of the current Internet of Things device in the experience pool; extracting data blocks from the experience pool for training, updating the Critic network and Actor network, and obtaining the optimized transmission power and local computing resources.

[0084] Step 11: Take the optimal transmission power and local computing resources calculated in Step 10 as known conditions, and repeat Step 2 until the weighted sum of the energy consumption and delay of each device converges. The current offloading strategy and action space are the optimal computing offloading and resource allocation scheme for industrial edge network-dependent tasks.

[0085] As a specific implementation of this embodiment, consider two base stations that are 1000 m apart. Each TN is randomly distributed within two circles with a radius of 500 m centered on the base stations. There are task dependencies between 2 TNs (TN 1 and TN 2), as Figure 4 shown, and the 2 TNs are respectively associated with different base stations, and the distances from the base stations are d1 = d2 = 15 m.

[0086] To more comprehensively evaluate the performance of the proposed algorithm, it is compared with the following 3 benchmark schemes:

[0087] (1) Full offloading scheme: All tasks are offloaded for calculation and not calculated locally;

[0088] (2) Only local calculation scheme: Tasks are calculated locally without task offloading;

[0089] (3) Independent optimization scheme: Ignoring task dependencies, each TN only optimizes the weighted sum of its own energy consumption and delay;

[0090] (4) DQN scheme: Use DQN to optimize the system, where the transmit power and local computing resources need to be discretized.

[0091] In the cases of K = 2 and K = 5, the convergence of the joint power and computing resource optimization algorithm based on DDPG is as Figure 5 shown. As the training process continues, the cumulative reward gradually increases and then converges, indicating that through continuous interaction with the dynamic environment, storage of experience, and training and update of network parameters, the agent successfully learns the resource allocation optimization strategy, thus verifying the feasibility of the algorithm.

[0092] In the case of K = 10, the influence of the computing resources of the edge server on the weighted sum of the system energy consumption and delay is as Figure 6 shown. Since more powerful computing capabilities shorten the computing time, as the computing resources F MEC of the edge server increase, the weighted sum of the system energy consumption and delay shows a downward trend, and the proposed scheme in this paper performs the best.

[0093] In the case of K = 10, the influence of the bandwidth on the weighted sum of the system energy consumption and delay is as Figure 7 shown. As the total available bandwidth of the system increases, the weighted sum of the system energy consumption and delay gradually decreases. Since a larger bandwidth brings a higher data rate, shortening the communication time for transmitting task data and the corresponding energy consumption, thus reducing the weighted sum of the system energy consumption and delay, and it can be observed from the figure that the proposed scheme in this paper performs the best.

[0094] The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for computing offloading and resource allocation of edge collaboration-dependent tasks, characterized in that Including the following steps: Construct an edge network architecture with task dependencies, where the edge network architecture includes a number of Internet of Things devices and a number of base stations equipped with mobile edge computing servers; Based on the edge network structure, conduct mathematical modeling of task dependencies to determine the dependency relationships between subtasks in the Internet of Things devices; Based on the dependency relationships, determine the subtask offloading decisions of the Internet of Things devices, and based on the subtask offloading decisions, determine the execution methods of the tasks; Based on the execution methods of the tasks, calculate the communication delay and waiting delay of the tasks to obtain the total task delay; Based on the total task delay, adopt the deep deterministic policy gradient algorithm to jointly optimize the subtask offloading decisions, transmit power, and local computing resource allocation to minimize the weighted sum of energy consumption and delay of all terminal devices.

2. The method for computing offloading and resource allocation of edge collaboration-dependent tasks according to claim 1, wherein, The process of constructing an edge network architecture with task dependencies includes: Determine the number and locations of the Internet of Things devices, as well as the number and locations of the base stations; Based on the number and locations of the Internet of Things devices, the number and locations of the base stations, establish an association relationship between the Internet of Things devices and the base stations; where each Internet of Things device is only associated with one base station; Define the subtask set of the Internet of Things devices and the dependency relationships between the subtasks.

3. The method for computing offloading and resource allocation of edge collaboration-dependent tasks according to claim 1, characterized in that The calculation process of the communication delay includes: According to the subtask offloading decisions of the Internet of Things devices, determine the transmission paths of the task data; Calculate the communication delays under different transmission paths, and the communication delay includes the uplink transmission link delay and the downlink transmission link delay; Among them, the calculation expression of the communication delay is: In the formula, represents the communication delay of the i-th task of the k IoT devices, x i,k represents that the i-th task of the k-th IoT device selects local computing or offloads to the edge server for computing, represents the uplink transmission delay between the k-th IoT device and the edge service, represents the downlink transmission delay between the k-th IoT device and the edge service.

4. The computing offloading and resource allocation method for edge collaboration-dependent tasks according to claim 3, wherein The expression for calculating the waiting delay is: Wherein, represents the waiting delay of the i-th task among the k Internet of Things devices for the j-th task among the m Internet of Things devices; represents the dependency relationship of the i-th task among the k Internet of Things devices for the j-th task among the m Internet of Things devices, O j,m represents the output data of the j-th task among the m Internet of Things devices; represents the time required for the m Internet of Things devices to run to task j; represents the output data O of the j-th task among the m Internet of Things devices j,m The uplink transmission time for transmission to the associated MEC nm; represents the time for the associated MEC nk to return the output data O of the j-th task among the m Internet of Things devices to the k Internet of Things devices j,m ; represents the time for sending O between MEC nm and MEC nk j,m ; 5. The method for computing offloading and resource allocation of edge collaboration-dependent tasks according to claim 1, wherein The process of joint optimization includes: Adopt the Gibbs sampling algorithm to pre-optimize the offloading decisions, select the offloading strategy according to the probability distribution, and obtain the optimized offloading strategy; Based on the optimized offloading strategy, optimize the transmit power and local computing resources through the DDPG algorithm to obtain the minimum value of the weighted sum of energy consumption and delay of all terminal devices.

6. The method for computing offloading and resource allocation of edge collaboration-dependent tasks according to claim 5, wherein During the process of selecting the offloading strategy according to the probability distribution, the calculation expression of the probability distribution is: where x k represents the offloading strategy selected according to the probability distribution, represents the probability distribution, θ represents the θ-th sampling, and x′ k represents the sampling set of sub-task offloading decisions in the k-th device, represents the sampling set generated from each IoT device, represents the optimal value of the objective function under the current given offloading decision.

7. The computing offloading and resource allocation method for edge collaboration-dependent tasks according to claim 1, characterized in that The process of adopting the deep deterministic policy gradient algorithm to jointly optimize the subtask offloading decisions, transmit power, and local computing resource allocation includes: Initialize the Critic network, Actor network, and network parameters; Based on the state space, determine the state of the current Internet of Things device, and based on the state of the current Internet of Things device, select the transmit transmission and computing resources and add noise to obtain the current action; Based on the current action, obtain the reward and the next state; Store the state, current action, reward, and next state of the current Internet of Things device in the experience pool; Extract data blocks from the experience pool for training, update the Critic network and Actor network, and obtain the optimized transmit power and local computing resources.

8. The method for computing offloading and resource allocation of edge collaboration-dependent tasks according to claim 7, characterized in that, The state space is a set of subtask offloading states, bandwidth allocation, MEC association, task dependency matrix, edge server computing resource allocation, and the current weighted sum of delay and energy consumption.

Images