A multi-UAV assisted edge computing joint offloading and deployment method

By using clustering and optimization algorithms in the MEC network supported by drones to cluster ground users and optimize drone deployment, the complex task dependence problem is solved, and efficient offloading and deployment of deep learning tasks is achieved, reducing the Maximum completion time improves system performance.

CN118612750BActive Publication Date: 2024-12-03SOUTHWEST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410703913.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-03
Publication Date
2024-12-03
Estimated Expiration
2044-06-03

AI Technical Summary

Technical Problem

Existing drone-supported mobile edge computing (MEC) networks assume that computing tasks can be performed in parallel, ignoring the interdependence between tasks, especially in deep learning and augmented reality networks, where task dependencies are complex, resulting in Unable to effectively optimize the uninstallation and deployment of deep learning tasks.

Method used

The clustering algorithm is used to cluster ground users, each cluster is equipped with a drone aerial edge server. Through joint optimization of sub-task offloading strategies and drone location deployment, mathematical models are used to describe the maximum completion time optimization problem of the task, and alternately The iterative optimization method is decomposed into subproblems for solving, including the subtask offload algorithm based on latency and dependency priorities and the Successive Convex Approximation (SCA) drone deployment algorithm.

Benefits of technology

It effectively reduces the maximum completion time of deep learning tasks, ensures the effective utilization of computing and communication resources, improves system performance, and meets the requirements of dependence of complex tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118612750B_ABST
    Figure CN118612750B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-UAV assisted edge computing joint unloading and deployment method, the method comprising step 1: obtaining the location information of ground users; step 2: clustering the ground users according to the location information of each ground user using a clustering algorithm to generate M clusters; step 3: setting a UAV for each cluster, each UAV is equipped with an air edge server to execute the DL tasks unloaded from the ground users in the cluster; step 4: by jointly optimizing the subtask unloading strategy and UAV position deployment, the maximum completion time of all DL tasks is minimized; step 5: using the optimized subtask unloading strategy and UAV position deployment results, unloading the subtasks and deploying the position of the UAV. The method of the present invention effectively reduces the maximum completion time of the computing task.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a multi-UAV assisted edge computing joint offloading and deployment method. Background Art

[0002] The rapid development of 5G infrastructure and the Internet of Things (IoT) has led to the rapid rise of innovative mobile applications such as augmented reality (AR), autonomous driving, and smart healthcare systems, which urgently require fast and reliable connectivity and low-latency computing capabilities to support these applications. However, many IoT devices face limitations in energy consumption and computing resources. To address these challenges, Mobile Edge Computing (MEC) has become a crucial solution. MEC allows devices to offload tasks to nearby edge servers, which reduces latency and overhead compared to traditional cloud computing. Despite its many advantages, MEC is still limited by the limited coverage and fixed locations of static edge servers. Therefore, MEC supported by Unmanned Aerial Vehicles (UAVs) has emerged. UAVs provide flexible deployment, cost-effectiveness, and high mobility to quickly provide services on demand to areas in need. In addition, UAVs can establish Line-of-Sight (LoS) links with Ground Users (GUs), improving the stability and reliability of connections while meeting Quality of Service (QoS) requirements.

[0003] In drone-supported MEC networks, there have been related works that combine and optimize various strategies to achieve different optimization goals. However, existing studies assume that all computing tasks can be executed in parallel, ignoring the interdependencies between computing tasks. In practical scenarios, modern mobile applications, such as deep learning networks and augmented reality networks, are composed of a large number of interdependent computationally intensive tasks. Having task dependencies means that a task can only start after receiving the results of the previous task. In addition, existing studies do not consider the real dependencies between tasks or oversimplify the intricate task dependency topology in deep learning (DL) tasks. Therefore, a multi-drone-assisted edge computing joint offloading and deployment method for optimizing deep learning tasks is needed to solve the above problems. Summary of the invention

[0004] In view of the problems in the background technology, the present invention provides a multi-UAV assisted edge computing joint offloading and deployment method, which includes the following steps:

[0005] Step 1: Obtain the location information of ground users, where the number of ground users is K and the set of ground users is represented as For any ground user s k The horizontal position coordinates of

[0006] Step 2: Based on the location information of each ground user, a clustering algorithm is used to cluster the ground users to generate M clusters;

[0007] Step 3: Set up a UAV for each cluster. Each UAV is equipped with an air edge server to perform DL tasks offloaded from ground users in the cluster. The channel between each ground user and the associated UAV is an orthogonal communication channel. k Generate N k DL tasks, each DL task contains multiple subtasks, and the UAV set is represented as

[0008] Step 4: Minimize the maximum completion time of all DL tasks by jointly optimizing the subtask offloading strategy and UAV location deployment;

[0009] Step 5: Use the optimized subtask offloading strategy and UAV position deployment results to offload the subtasks and deploy the UAV position.

[0010] The optimization problem in step 4, denoted as P1, is expressed as:

[0011]

[0012] in,

[0013] C1-C5 represent five constraints;

[0014] Indicated by ground user s k The nth DL task generated, represents the set of all DL tasks, that is 1≤n≤N k , 1≤k≤K;

[0015] Represents DL tasks The time of generation;

[0016] Represents DL tasks Tolerable delay;

[0017] Represents DL tasks DAG structure;

[0018] It is a DL task A set of subtasks, each of which corresponds to a layer of a deep neural network. Representing a collection The number of elements in The total number of subtasks is L;

[0019] is an edge set, representing the DL task The dependencies between the subtasks, the data flow edges Represents a subtask It is a subtask The parent task, i.e. the subtask Dependencies on subtasks

[0020] Indicates the deployment of drones, Indicates drone u m Horizontal position;

[0021] represents the subtask offloading strategy, x kmnl ∈{0,1},x kmnl =1 indicates a subtask will be unloaded to the drone m On processing, x kmnl =0 means subtask In users k Local execution; symbols The formula is the definition formula;

[0022] q i and q j Denote any two drones u i and u j Horizontal position;

[0023] d min Indicates the minimum safe distance that should be maintained between any two drones;

[0024]

[0025] Indicates the task The maximum completion time of all subtasks in ;

[0026] Represents a subtask completion time;

[0027] Represents a subtask The start time on device d, which is the ground user s k or drone m , by subtask Whether it needs to be uninstalled is determined;

[0028] Represents a subtask The computation time required.

[0029] Further, Calculated by the following formula:

[0030]

[0031] in, Represents a subtask The number of FLOPs required, f k and f m Represents ground users s k and droneu m The computing power, or floating point operations per second;

[0032]

[0033] in, Represents a subtask The transmission delay to device d, Represents a subtask The queuing delay in the computation queue of device d.

[0034] Furthermore, the transmission delay Dependencies on subtasks Dependent subtasks Completion time on device d′ Subtasks Waiting time in the transmission queue And subtasks The generated output result data is sent to the subtask The time required for the execution of the device Right now

[0035]

[0036] Among them, the subtask Dependent subtasks Subtask It is a subtask A parent task of Represents a subtask The set of all parent tasks of , device d′ is the ground user s k or dronem , by subtask Determined by whether it is uninstalled;

[0037]

[0038] represents an uninstalled execution with local dependencies, i.e., when , indicating a subtask is offloaded to the UAV, and its dependent subtasks Execute locally, otherwise x kmnl′ Represents a subtask Whether it is offloaded to the UAV;

[0039] Represents a subtask The output data generated from the ground user k Transmit to drone m Uplink transmission duration;

[0040]

[0041] , Indicates that the subtask The size of the generated output data, Represents ground user s k Offload tasks to drones m Uplink rate at 1000 ms;

[0042] represents a local execution with offloaded dependencies, i.e., when , indicating a subtask Executed locally, and its dependent subtasks is offloaded to the UAV, otherwise

[0043] Indicates that the subtask The output data generated from the drone m Transmission to ground users k Downlink transmission duration;

[0044]

[0045] , Represents ground user s k From UAV m Downlink rate when downloading data;

[0046] Represents a subtask The transmit queue on device d′ Get the queuing delay of channel resources in

[0047]

[0048] and Indicated by ground user s k′ The i-th and j-th subtasks in the generated n′-th DL task;

[0049] Represents the transfer subtask Previously, the transmit queue of device d′ The data flow edge set in , where the transmission subtask The previous transmission queue of device d′ Also recorded as

[0050] Represents a subtask Transmission delay to device d′;

[0051] Queuing Delay The formula is:

[0052]

[0053] in, Represents a subtask The completion time on device d, Indicates that device d is computing subtasks Previously, the compute queue of device d The subtask set in which the computation subtask Previously, the compute queue of device d Also recorded as Indicated by ground user s k′ The l′th subtask in the generated n′th DL task.

[0054] Furthermore, the method further includes converting the P1 problem in step 4 into two sub-problems, namely, a sub-task offloading decision problem and a UAV position deployment problem, and performing alternating iterative optimization on the two sub-problems using an alternating iterative optimization method, thereby obtaining an optimal sub-task offloading strategy and UAV position deployment, wherein the alternating iterative optimization process is:

[0055] S11: Initialize UAV position deployment, and set the number of iterations r = 0;

[0056] S12: In the given Under these conditions, the priority-based subtask offloading algorithm is used to determine the subtask offloading Indicates that the drone at the rth iteration u The position of m, represents x at the r+1th iteration kmnl ;

[0057] S14: Let t = 0 and set the rth iteration Assign to The initial value of where r t Indicates the number of internal iterations when using the SCA method;

[0058] S15: In the given When, through the SCA method, with the given Determine the optimal solution Continue to perform internal iterations until the internal iteration termination condition is met;

[0059] S16: The internal iteration is terminated. Assign to And assign r+1 to r;

[0060] S17: Go back to step S12 and repeat steps S12-S16 until the iteration termination condition is met. The subtask unloading and UAV position deployment results at the time of iteration termination are the optimal solution.

[0061] Furthermore, the priority-based subtask offloading problem is to optimize the offloading strategy {x kmnl}, and record this subproblem as P2, which can be expressed as:

[0062]

[0063] st(C1)-(C5),

[0064] For P2, a priority-based subtask offloading algorithm is adopted, including:

[0065] S1: For all DL tasks, according to each DL task Sort the tolerable delays and determine the processing order of DL tasks based on the sorting results;

[0066] S2: For each DL task The subtasks in the probability ascending rank algorithm are used to determine each subtask Priority That is, give higher priorities to subtasks that depend on data transfer for a longer time, and prioritize subtasks according to their priorities;

[0067] S3: According to the sorting results in S1 and S2, for each subtask in all DL tasks Tasks are offloaded in order of priority.

[0068] Furthermore, S2 has a priority The method of determining is:

[0069]

[0070] in, Indicates all dependencies A collection of subtasks, is a binary variable, indicating that Is the output considered? The transmission time; and are the average transmission rate of all links and the average computing power of all devices, respectively.

[0071]

[0072] in, rand() is a function that returns a random value in the interval [0,1), where λ is a constant.

[0073] Furthermore, S3 specifically includes, according to the priority order of the subtasks, for each subtask Execute steps S31-S40 until all subtasks in all DL tasks are completed, and then you can get {x kmnl},

[0074] Among them, S31-S40 include:

[0075] S31: Determine whether l=1 or Is it true? If so, execute S32; if not, execute S33;

[0076] S32: Make x kmnl =0, and Placed on the ground k The calculation queue The tail of the team;

[0077] S33: Determine x according to the following formula kmnl ,Right now:

[0078]

[0079] in, Represents a subtask Unload to its associated drone m Time subtask completion time; Represents a subtask The completion time during local execution, then, execute S34;

[0080] S34: Judge x kmnl =1 whether it is true, if it is true, it means the subtask Need to be uninstalled to u m If the condition is not met, execute S35; if not, it means that the subtask If it is executed locally and does not need to be uninstalled, then S36 is executed;

[0081] S35: Add to u m The calculation queue At the end of the team, then, execute S37;

[0082] S36: Will Add to ground users k The calculation queue of At the end of the team, then, execute S37;

[0083] S37: Judgment exist Li Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Execute locally, subtask Need to be unloaded to the drone m , execute S38; otherwise, execute S39;

[0084] S38: Join Ground Users k Transmission queue The tail of the team;

[0085] S39: Judgment exist And Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Unloaded on drone m , and the subtask If local execution is required, execute S40;

[0086] S40: Join UAV m Transmission queue The end of the team.

[0087] Furthermore, the SCA-based UAV deployment problem is to fix the offloading decision {x kmnl}, optimize the UAV position deployment {q m}, transform this subproblem into P3, which is expressed as:

[0088]

[0089]

[0090] in,

[0091] C6-C8 represent constraints;

[0092] and Respectively represent the uplink and downlink rates at the rth iteration;

[0093] Indicates an increase in the uplink rate;

[0094] Indicates an increase in the downlink rate;

[0095] η represents the minimum lifting rate;

[0096] Indicates user s k and droneu m Is there data transmission on the link between them? If y km >0 indicates user s k and droneu m There is data transmission between the links, otherwise y km =0.

[0097] Furthermore, the method further includes converting problem P3 into problem P4, where P4 is expressed as:

[0098]

[0099] in,

[0100] C9-C11 represent constraints;

[0101] Indicates the uplink rate The lower bound of

[0102]

[0103] Indicates the downlink rate The lower bound of

[0104]

[0105]

[0106] H represents the vertical height of the drone from the ground, B is the channel bandwidth, Indicates the received signal-to-noise ratio at 1 meter, P k For k The transmission power, σ 2 is the background noise power; Where P m is u m The transmission power; β0 is the channel power gain at a reference distance of 1 meter; the superscript r is added t The q represents the t The horizontal position of the drone at the first internal iteration; and Respectively expressed in the rth t Two local points at the inner iteration,

[0107] Problem P4 is a standard convex optimization problem. In S15, given When, through the SCA method, with the given Solve problem P4 and get the optimal solution

[0108] The beneficial effects of the above technical solution are:

[0109] The present invention considers a MEC system supporting multiple drones, each drone is equipped with an air edge server to provide computing services for GU, which generates some DL tasks and is modeled as a directed acyclic graph (DAG) structure. Using the task tolerable delay as an indicator of task urgency, a mathematical model is established to characterize the joint optimization problem of offloading decision and drone deployment to minimize the maximum completion time of all tasks; the present invention decomposes the mixed integer non-convex optimization problem derived from the model into two sub-problems, and proposes corresponding solutions, specifically including a sub-task offloading algorithm based on delay and dependency priority and a drone deployment algorithm (Successive Convex Approximation, SCA), and the present invention also proposes an effective iterative algorithm to alternately optimize the two sub-problems, so that the present invention effectively improves the performance in minimizing the maximum completion time of the task and ensures the effective use of computing and communication resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] The accompanying drawings of the present invention are as follows:

[0111] Figure 1 It is a schematic diagram of a multi-UAV-assisted MEC system;

[0112] Figure 2 This is a diagram of the offloading process in the VGG 16 task, which consists of 26 subtasks;

[0113] Figure 3 (a)-3(c) are the simulation results of the relationship between the number of tasks, bandwidth size and computing power and the maximum completion time. DETAILED DESCRIPTION

[0114] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0115] The present invention considers a multi-UAV-assisted MEC system, where each UAV is equipped with an air edge server that can provide computing services for K GUs, i.e., execute the offloaded DL tasks. The set of ground users is represented as For any ground user s k The horizontal position coordinates of The present invention considers that each GU can only be associated with one drone, and one drone can provide computing services for multiple users. By using a clustering algorithm (such as the K-means algorithm), the GU is divided into M clusters, each cluster is served by a drone to alleviate the user's computing power limitations. For safety and simplification, the drone will hover at a height of H above the ground.

[0116] The present invention discloses a multi-UAV assisted edge computing joint unloading and deployment method, which comprises the following steps:

[0117] Step 1: Obtain the location information of ground users, where the number of ground users is K;

[0118] Step 2: Based on the location information of each ground user, a clustering algorithm is used to cluster the ground users to generate M clusters;

[0119] Step 3: Set up a UAV for each cluster. Each UAV is equipped with an air edge server to perform DL tasks offloaded from ground users in the cluster. The UAV set is represented as Indicates drone u m , wherein the channel between each ground user and the associated UAV is an orthogonal communication channel;

[0120] Step 4: Minimize the maximum completion time of all DL tasks by jointly optimizing the subtask offloading strategy and UAV location deployment;

[0121] Step 5: Use the optimized subtask offloading strategy and UAV position deployment results to offload the subtasks and deploy the UAV position.

[0122] The present invention assumes that each s k Generate N k DL reasoning tasks (i.e. DL tasks) and aggregate all tasks into the task queue In represents the set of all DL tasks, that is 1≤n≤N k , 1≤k≤K.

[0123] Indicated by ground user s k The nth DL task generated, Represents DL tasks The time of generation; Represents DL tasks Tolerable delay; Represents DL tasks DAG structure. Among them, It is a DL task A set of subtasks, each of which corresponds to a layer of a deep neural network. Representing a collection The number of elements in The total number of subtasks is L. Represents a subtask The number of floating point operations (FLOPs). Indicates that the subtask The size of the generated output result data. The edge set in the DAG graph represents the dependencies between subtasks. For any data flow edge It represents a subtask It is a subtask The parent task, i.e. the subtask Dependencies on subtasks Due to the dependencies between subtasks, a subtask can only be executed after receiving the processing results of all its parent tasks. and Respectively The parent task collection and child task collection of .

[0124] The present invention adopts a binary offloading strategy, that is, a subtask is either fully calculated locally or fully offloaded to a drone for calculation. The advantage of using a binary offloading strategy is that it can meet the requirements of subtasks for data integrity and indivisibility, simplify the division and scheduling of subtasks, eliminate the complex overhead of subtask segmentation and merging in the context of DL tasks, and ensure the integrity and security of the entire task, effectively reducing the risk of data leakage and tampering that is common in partial offloading.

[0125] For users k Uninstall subtask To UAV m In the case of kmnl ∈{0,1}, where x kmnl =1 indicates a subtask will be unloaded to the drone m Upper processing; x kmnl =0 means the subtask is in user s k Local execution. Since each GU is associated with at most one UAV, the constraint It is important to note that the first and last subtasks of a DL task must be executed locally on the GU that generates the task, so there is a constraint For subtasks The number of FLOPs required is Users k and droneu m The computing power (number of floating-point operations per second) is expressed as f k and f m To express.

[0126] Therefore, the subtask Required computation time It can be expressed as

[0127]

[0128] Since drones are deployed in the air, the air-to-ground communication link can be considered a line-of-sight (LoS) link. m and users k The channel gain between Where α≥2 represents the path loss exponent, β0 is the channel power gain at a reference distance of 1 meter, It's a drone m and users k The present invention considers orthogonal communication channels so that the data transmission between multiple GUs and their associated UAVs does not interfere with each other.

[0129] The following describes three cases of offloading computation of subtasks with dependencies.

[0130] 1) Executed on the same device: subtasks and its dependent subtasks When executed on the same device (either locally or on the drone), there is no need to transfer dependent data between devices.

[0131] 2) Uninstall execution with local dependencies: define Specifically, when , indicating a subtask Offloaded to the UAV, and the subtasks it depends on Execute locally, otherwise When users k Offload tasks to drones m When calculating the uplink rate Expressed as

[0132]

[0133] Where B is the channel bandwidth, represents the received signal-to-noise ratio (SNR) at 1 meter, where P k For k The transmission power, σ 2 is the background noise power. In the above case, the subtask The output data generated from the ground user k Transmit to drone m Uplink transmission time It can be expressed as

[0134]

[0135] Indicates that the subtask The size of the generated output data.

[0136] 3) Local execution with uninstalled dependencies: This case is represented by Specifically, when , indicating a subtask Executed locally, and its dependent subtasks is offloaded to the UAV; otherwise When users k From UAV m Download dependent data downlink rate It can be expressed as

[0137]

[0138] In the formula Where P m is u m The transmission power of The output data generated from the drone m Transmission to ground users k The downlink transmission duration is:

[0139]

[0140] According to the above three situations, the subtasks can be The output result data is sent to the subtask The transmission time is uniformly expressed as:

[0141]

[0142] The present invention also establishes a task queuing model. In the present invention,

[0143] Each GU and UAV has two First Input First Output (FIFO) queues: a computation queue and a transmit queue The subscript Represents the network entity set, that is, the set of all devices (including ground users and drones). The present invention considers that the device processor is a single-core processor, that is, it can only process a single subtask at the same time. Therefore, when the device is occupied, any new subtask added to the execution will be In addition, due to the link bandwidth limitation, the link can only send a single data stream at a time. When the link is occupied, the subtasks ready for transmission Generated output data Will be in This mechanism ensures in-order execution of devices and reliable data transmission, and complies with the inherent limitations of device processing power and network bandwidth.

[0144] Subtask on device d The start time depends on the calculation queue Queuing delay and transmission delay in Specifically, the subtask Start time It is determined by the moment when the child task has all the required input data (that is, all its parent tasks must complete and receive the data generated by all parent tasks) and obtains the computing resources of the device. Therefore, the start time can be expressed as

[0145]

[0146] in, It is a computational subtask Compute queue on device d The queuing delay in the computation queue is determined by the queuing delay of the subtask in the computation queue. The completion time of the previous subtask is determined, that is:

[0147]

[0148] Here, Represents a subtask The completion time on device d, Indicates that in calculation Previously, the compute queue of device d The subtask set in which the computation subtask Previously, the compute queue of device d Also recorded as Indicated by ground user s k′ The l′th subtask in the generated n′th DL task.

[0149] Transmission delay Dependencies on subtasks Dependent subtasks Completion time on device d′ Subtasks Waiting time in the transmission queue And subtasks The generated output result data is sent to the subtask The time required for the execution of the device Right now

[0150] Among them, the subtask Dependent subtasks Subtask It is a subtask A parent task of Represents a subtask The set of all parent tasks of , device d′ is the ground user s k or drone m , by subtask Determines whether it is uninstalled. Represents a subtask Need to occupy the channel for data transmission, Represents a subtask Get the queueing delay of channel resources in the transmit queue:

[0151]

[0152] and Indicated by ground user s k′ The i-th and j-th subtasks in the generated n′-th DL task;

[0153] Represents the transfer subtask Previously, the transmit queue of device d′ The data flow edge set in , where the transmission subtask The previous transmission queue of device d′ Also recorded as

[0154] Represents a subtask Transmission delay to device d′.

[0155] Each subtask Completion time Determined by the execution order, the generation time of the task and the execution time of each subtask can be calculated:

[0156]

[0157] Here The calculation is from formula (1).

[0158] Finally, the total completion time required for a DL task can be obtained, also known as the makespan, which is expressed as Indicates that its representative mission The maximum completion time of all subtasks in:

[0159]

[0160] Therefore, the optimization problem in step 4, denoted as P1, can be expressed as:

[0161]

[0162]

[0163] in,

[0164] C1-C5 represent five constraints; d min represents the minimum safe distance that should be maintained between any two UAVs. The tolerable delay constraint of each DL task is satisfied in (13a). Due to the complex interdependence between subtasks and the combination of binary variables in (13b) and non-convex constraints in (13e), it can be seen that (P1) is a complex mixed integer non-convex optimization problem.

[0165] In this paper, the original problem (P1) is decomposed into two sub-problems, and corresponding solutions are proposed, namely, a priority-based subtask offloading algorithm and an SCA-based UAV deployment algorithm. Then, an effective iterative algorithm is proposed to alternately optimize the two sub-problems until the algorithm converges.

[0166] Referring to Algorithm 2, the process of alternating iterative optimization is:

[0167] S11: Initialize UAV position deployment, and set the number of iterations r = 0;

[0168] S12: In the given Under these conditions, the priority-based subtask offloading algorithm is used to determine the subtask offloading Indicates the number of drones u at the rth iteration m location, represents x at the r+1th iteration kmnl ;

[0169] S14: Let t = 0 and set the rth iteration Assign to The initial value of where r t Indicates the number of internal iterations when using the SCA method;

[0170] S15: In the given When, through the SCA method, with the given Determine the optimal solution Continuously perform internal iterations until the internal iteration termination condition is met (for example, the improvement of the minimum rate is less than the specified accuracy);

[0171] S16: The internal iteration is terminated. Assign to And assign r+1 to r;

[0172] S17: Return to step S12, and repeat steps S12-S16 until the iteration termination condition is met (for example, the target value reduction rate of problem P1 is less than the specified accuracy). The subtask unloading and UAV position deployment results at the time of iteration termination are recorded as the optimal solution.

[0173] Subtask offloading decision is to optimize the offloading strategy {x kmnl}, and record this subproblem as P2, which can be expressed as:

[0174]

[0175] st(C1)-(C5),

[0176] In order to better meet the delay requirements of DL tasks, the present invention determines the order of DL task processing by sorting them according to their tolerant delay (the tolerant delay reflects the urgency of the task); then the subtasks in each DL task are prioritized, while ensuring that the execution order of the subtasks meets their dependency requirements. Therefore, a priority-based subtask offloading algorithm is proposed, which includes task queuing (i.e., S1 and S2) and task offloading (i.e., S3).

[0177] Specifically, for problem P2, a priority-based subtask offloading algorithm is used to solve it, which includes:

[0178] S1: For all DL tasks, according to each DL task Sort the tolerable delays and determine the processing order of DL tasks based on the sorting results;

[0179] After determining the processing order of DL tasks, the DAG structure of each DL task is considered to determine the priority of each subtask. Specifically, when different subtasks have the same parent task dependency, a higher priority will be set for subtasks with longer dependent data transmission time so that they can be executed earlier and reduce completion time.

[0180] In S2: For each DL task The subtasks in the probability ascending rank algorithm are used to determine each subtask Priority And prioritize the subtasks according to their priority;

[0181] S3: According to the sorting results in S1 and S2, that is, after obtaining the priority list, each subtask in all DL tasks Tasks are offloaded in order of priority.

[0182] Specifically, S2 has a priority The method of determining is:

[0183]

[0184] in, Indicates all dependencies A collection of subtasks, is a binary variable, indicating that Is the output considered? The transmission time; and are the average transmission rate of all links and the average computing power of all devices, respectively.

[0185]

[0186] in, rand() is a function that returns a random value in the interval [0,1). λ is a large constant. As can be seen from formula (15), a small reflects higher transmission costs and will increase The possibility of being 0 satisfies the requirement mentioned above that a higher priority is set for a longer transmission time under the same parent dependency.

[0187] After determining the priority, the system needs to decide whether to offload the subtask to its related drone for execution. Unload to its associated drone m , the completion time is That is, the execution device is specified as d = m; if executed locally, the completion time is That is, the execution device is designated as d = k. Subsequently, the offloading decision will select the device that can complete the subtask earliest, calculated as follows

[0188]

[0189] S3 specifically includes, according to the priority order of the subtasks, for each subtask Execute steps S31-S40 until all subtasks in all DL tasks are completed, and then you can get and {x kmnl},

[0190] Among them, S31-S40 include:

[0191] S31: Determine whether l=1 or Is it true? If so, execute S32; if not, execute S33;

[0192] S32: Make x kmnl =0, and Placed on the ground k The calculation queue The tail of the team;

[0193] S33: Determine x according to the following formula kmnl ,Right now:

[0194]

[0195] in, Represents a subtask Unload to its associated drone m Time subtask completion time; Represents a subtask The completion time during local execution, then, execute S34;

[0196] S34: Judge x kmnl =1 whether it is true, if it is true, it means the subtask Need to be uninstalled to u m If the condition is not met, execute S35; if not, it means that the subtask If it is executed locally and does not need to be uninstalled, then S36 is executed;

[0197] S35: Add to u m The calculation queue At the end of the team, then, execute S37;

[0198] S36: Will Add to ground users k The calculation queue of At the end of the team, then, execute S37;

[0199] S37: Judgment exist Li Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Execute locally, subtask Need to be unloaded to the drone m , execute S38; otherwise, execute S39;

[0200] S38: Join Ground Users k Transmission queue The tail of the team;

[0201] S39: Judgment exist And Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Unloaded on drone m , and the subtask If local execution is required, execute S40;

[0202] S40: Join UAV m Transmission queue The end of the team.

[0203] Algorithm 1 gives the detailed steps of subtask offloading.

[0204]

[0205] For UAV deployment the problem is to fix the offloading decision {x kmnl}, optimize the UAV position deployment {q m}.

[0206] Since the performance indicators related to equipment calculation have been determined, the deployment of drones can be optimized to improve the communication rate of all transmission data links, thereby shortening the task completion time. This sub-problem is transformed into P3, which is the maximum minimum rate improvement. P3 can be expressed as:

[0207]

[0208] in,

[0209] C6-C8 represent constraints;

[0210] and Respectively represent the uplink and downlink rates at the rth iteration;

[0211] Indicates an increase in the uplink rate;

[0212] Indicates an increase in the downlink rate;

[0213] η represents the minimum lifting rate;

[0214] Indicates user s k and droneu m Is there data transmission on the link between them? If y km >0 indicates user s k and droneu m There is data transmission between the links, otherwise y km =0.

[0215] Due to the nonconvex constraints in (17a), (17b), and (17c), (P3) is a nonconvex optimization problem. However, and Relative to the term ‖q m -q k ‖ 2 is convex. t In the iterations, by Applying the first-order Taylor expansion above, we can get the uplink data rate The lower bound of is expressed as follows

[0216]

[0217] in Similarly, the downlink data rate The lower bound of

[0218]

[0219] in

[0220] Convex term ‖q i -q j ‖ 2 Linearization is performed using the t The local points obtained in the iteration and We get the following inequality

[0221]

[0222] In (P3), we use the lower bounds derived in (18), (19), and (20) to find ‖q i -q j ‖ 2 , and By substitution, it can be restated as the following approximate problem, denoted as P4, and P4 is expressed as:

[0223]

[0224] in,

[0225] C9-C11 represent constraints;

[0226] Added superscript r t The q represents the t The horizontal position of the drone at the time of the internal iteration.

[0227] Thus, problem (P4) is a standard convex optimization problem that can be solved using CVX. It should be noted that the constraints of (P4) are stricter than those of (P3) due to the substitution with the lower bound. Therefore, any feasible solution of (P4) is a valid solution of (P3), and thus the objective value of (P4) is the lower bound of (P3). Based on the above discussion, the suboptimal solution of (P1) can be obtained by the alternating optimization method, i.e., iteratively alternating optimization of (P2) and (P4).

[0228] The details of the proposed iterative algorithm are summarized in Algorithm 2. For Algorithm 2, the result value of (P2) does not increase during the iteration process. Since the target value of (P2) is a finite lower bound, the convergence of Algorithm 2 is also guaranteed.

[0229]

[0230] The advantages of the present invention are further illustrated by simulation experiments below.

[0231] For the multi-UAV-assisted MEC system, we set K = 10, i.e., 10 GUs and M = 2, i.e., 2 UAVs. Figure 1 As shown, GUs are uniformly and randomly distributed in the 1.6×1.6 km 2 area. The initial horizontal position of the UAV is obtained through the K-means clustering algorithm and fixed at an altitude of 100 m. In addition, each UAV is equipped with an air edge server "Tesla K20" (3520 GFLOPS), and the GU is randomly equipped with "CPU i7 5930K" (289 GFLOPS) or "CPU i7 4770K" (206GFLOPS). In order to better verify the universality of the solution, four DL models widely used in computer vision applications: Resnet 34, Alexnet V2, VGG 16 and Inception V4 are used for experiments, with the number of subtasks being 59, 16, 26 and 127 respectively. Each GU randomly creates multiple DL inference tasks, and they have different DL network structures

[20] . The relevant channel parameters are set as follows: B = 1 MHz, σ 2 =-110dBm, β0=-50dB, α=2.2, P k =0.1W, P m =0.1W,

[0232] In order to show the offloading details of subtasks in the multi-layer task structure, we Figure 2 The offloading process in the VGG 16 task, which consists of 26 subtasks, is shown in Figure 2. Figure 2 It is known that the offloading decision of the first and last subtask is 0, which means that the two subtasks are executed locally. When the offloading decision is equal to 1, the current subtask is offloaded to the drone, which meets the offloading requirements. Considering the task dependency, the subtask is offloaded to the device that is expected to be completed earliest.

[0233] To further investigate the algorithm performance gain, the proposed algorithm is compared with the following benchmarks: 1) Random offloading benchmark: subtasks are randomly offloaded to associated drones while taking into account task dependencies; 2) Local computing benchmark: all subtasks are executed only on local devices; 3) Air-ground equal computing benchmark: half of the subtasks of each DL task are executed locally, and the other half are offloaded to drones. Figure 3 (a) It can be seen that the completion time of all solutions increases with the task queue. increases with the increase of length, and The larger it is, the more significant the performance improvement is. This is because as the number of tasks increases, it takes more time to obtain computing and communication resources. Therefore, our algorithm takes into account the resource occupancy status of the device and makes overall arrangements to select a more suitable device for each subtask, ensuring the effective use of computing and communication resources. The air-ground equal computing performs better than the local computing scheme because it is able to assign half of the tasks to drones with more powerful computing capabilities. Although the random offloading scheme allows tasks to be offloaded to drones, it ignores the waiting time caused by the random allocation of resources, resulting in an increase in time cost, especially when the number of tasks increases.

[0234] exist Figure 3 In (b), it can be seen that as the bandwidth B increases, the maximum completion time of the random offloading scheme in the initial stage is significantly reduced. When B is large enough, the limiting factor becomes the computing capacity. In contrast, local computing and air-ground computing schemes are insensitive to changes in B. For the random offloading scheme, the increase in B significantly reduces the transmission cost, while GUs fully utilizes the high computing power of drones, thereby reducing the maximum completion time. Figure 3 (c) describes the maximum completion time as the UAV computing capacity The algorithm shows that the maximum completion time changes with Similarly, communication cost, as a more critical bottleneck factor, decreases with the increase of The improvement of resource efficiency is more likely to lead to slower performance improvement. In air-ground computing schemes, half of the subtasks will be executed on drones, resulting in increased drone occupancy and longer resource waiting time. In addition, although the random offloading scheme also utilizes drones with higher computing power, the performance of this scheme is no different from that of the local computing scheme because it does not consider the impact of resource waiting time, so it is difficult to achieve significant improvement. The comprehensive strategy of jointly optimizing task offloading and drone deployment that we adopted effectively combines drone deployment with offloading decisions, achieving significant performance improvements and improving the efficiency of the system.

[0235] In summary, the present invention considers the problem of minimizing the maximum completion time of deep learning tasks with dependencies in a multi-UAV-assisted MEC system. This problem is a joint optimization problem based on task offloading decisions and UAV deployment. By decomposing the original problem into two sub-problems, the present invention designs a priority-based iterative optimization algorithm to determine the offloading decision of dependent tasks, and a UAV deployment strategy based on the SCA method to improve the communication efficiency during the offloading process. Based on this, the present invention proposes an effective iterative algorithm to alternately optimize the two sub-problems, and proves its convergence. A large number of simulation results show that under different environmental settings, compared with other benchmark schemes, the joint iterative algorithm proposed in the present invention effectively reduces the maximum completion time of the computing task.

[0236] Finally, it should be noted that the embodiments described in this specification are merely illustrative of the present invention, and those skilled in the art should understand that various changes may be made in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A multi-UAV assisted edge computing joint offloading and deployment method, characterized in that: include: Step 1: Obtain the location information of ground users, where the number of ground users is K and the set of ground users is represented as For any ground user s k The horizontal position coordinates of Step 2: Based on the location information of each ground user, a clustering algorithm is used to cluster the ground users to generate M clusters; Step 3: Set up a UAV for each cluster. Each UAV is equipped with an air edge server to perform DL tasks offloaded from ground users in the cluster. The channel between each ground user and the associated UAV is an orthogonal communication channel. k Generate N k DL tasks, each DL task contains multiple subtasks, and the UAV set is represented as Step 4: Minimize the maximum completion time of all DL tasks by jointly optimizing the subtask offloading strategy and UAV location deployment; Step 5: Use the optimized subtask offloading strategy and UAV position deployment results to offload the subtasks and deploy the UAV position. The optimization problem in step 4, denoted as P1, is expressed as: in, C1-C5 represent five constraints; Indicated by ground user s k The nth DL task generated, represents the set of all DL tasks, that is Represents DL tasks The time of generation; Represents DL tasks Tolerable delay; Represents DL tasks DAG structure; It is a DL task A set of subtasks, each of which corresponds to a layer of a deep neural network. Representing a collection The number of elements in The total number of subtasks is L; is an edge set, representing the DL task The dependencies between the subtasks, the data flow edges Represents a subtask It is a subtask The parent task, i.e. the subtask Dependencies on subtasks Indicates the deployment of drones, Indicates drone u m Horizontal position; represents the subtask offloading strategy, x kmnl ∈{0,1},x kmnl =1 indicates a subtask will be unloaded to the drone m On processing, x kmnl =0 means subtask In users k Local execution; symbols The formula is the definition formula; q i and q j Denote any two drones u i and u j Horizontal position; d min Indicates the minimum safe distance that should be maintained between any two drones; Indicates the task The maximum completion time of all subtasks in ; Represents a subtask completion time; Represents a subtask The start time on device d, which is the ground user s k or drone m , by subtask Whether it needs to be uninstalled is determined; Represents a subtask The computation time required; The DAG structure is a directed acyclic graph structure, and the DL task is a deep learning task.

2. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 1, characterized in that: Calculated by the following formula: in, Represents a subtask The number of FLOPs required, where FLOPs is the number of floating point operations, f k and f m Represents ground users s k and droneu m The computing power, or floating point operations per second; , in, Represents a subtask The transmission delay to device d, Represents a subtask The queuing delay in the computation queue of device d.

3. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 2, characterized in that: Transmission delay Dependencies on subtasks Dependent subtasks Completion time on device d′ Subtasks Waiting time in the transmission queue And subtasks The generated output result data is sent to the subtask The time required for the execution of the device Right now Among them, the subtask Dependent subtasks Subtask It is a subtask A parent task of Represents a subtask The set of all parent tasks of , device d′ is the ground user s k or drone m , by subtask Determined by whether it is uninstalled; represents an uninstalled execution with local dependencies, i.e., when , indicating a subtask is offloaded to the UAV, and its dependent subtasks Execute locally, otherwise x kmnl′ Represents a subtask Whether it is offloaded to the UAV; Represents a subtask The output data generated from the ground user k Transmit to drone m Uplink transmission duration; , Indicates that the subtask The size of the generated output data, Represents ground user s k Offload tasks to drones m Uplink rate at 1000ms; represents a local execution with offloaded dependencies, i.e., when , indicating a subtask Executed locally, and its dependent subtasks is offloaded to the UAV, otherwise Indicates that the subtask The output data generated from the drone m Transmission to ground users k Downlink transmission duration; , Represents ground user s k From UAV m Downlink rate when downloading data; Represents a subtask The transmit queue on device d′ Get the queuing delay of channel resources in and Indicated by ground user s k′ The i-th and j-th subtasks in the generated n′-th DL task; Represents the transfer subtask Previously, the transmit queue of device d′ The data flow edge set in , where the transmission subtask The previous transmission queue of device d′ Also recorded as Represents a subtask Transmission delay to device d′; Queuing Delay The formula is: in, Represents a subtask The completion time on device d, Indicates that device d is computing subtasks Previously, the compute queue of device d The subtask set in which the computation subtask Previously, the compute queue of device d Also recorded as Indicated by ground user s k′ The l′th subtask in the generated n′th DL task.

4. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 3, characterized in that: The method further includes converting the P1 problem in step 4 into two sub-problems, namely, a sub-task offloading decision problem and a UAV position deployment problem, and performing alternating iterative optimization on the two sub-problems using an alternating iterative optimization method, thereby obtaining an optimal sub-task offloading strategy and UAV position deployment, wherein the alternating iterative optimization process is: S11: Initialize UAV position deployment, and set the number of iterations r = 0; S12: In the given Under these conditions, the priority-based subtask offloading algorithm is used to determine the subtask offloading Indicates the number of drones u at the rth iteration m location, represents x at the r+1th iteration kmnl ; S14: Let t = 0 and set the rth iteration Assign to The initial value of where r t represents the number of internal iterations when the SCA method is used, wherein the SCA method is a successive convex approximation method; S15: In the given When, through the SCA method, with the given Determine the optimal solution Continue to perform internal iterations until the internal iteration termination condition is met; S16: The internal iteration is terminated. Assign to And assign r+1 to r; S17: Go back to step S12 and repeat steps S12-S16 until the iteration termination condition is met. The subtask unloading and UAV position deployment results at the time of iteration termination are the optimal solution.

5. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 4, characterized in that: The subtask offloading decision problem is to optimize the offloading strategy {x kmnl }, and record this subproblem as P2, which can be expressed as: st(C1)-(C5), For P2, a priority-based subtask offloading algorithm is adopted, including: S1: For all DL tasks, according to each DL task Sort the tolerable delays and determine the processing order of DL tasks based on the sorting results; S2: For each DL task The subtasks in the probability ascending rank algorithm are used to determine each subtask Priority That is, give higher priorities to subtasks that depend on data transfer for a longer time, and prioritize subtasks according to their priorities; S3: According to the sorting results in S1 and S2, for each subtask in all DL tasks Tasks are offloaded in order of priority.

6. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 5, characterized in that: S2 Medium Priority The method of determining is: in, Indicates all dependencies A collection of subtasks, is a binary variable, indicating that Is the output considered? The transmission time; and are the average transmission rate of all links and the average computing power of all devices, respectively. in, rand() is a function that returns a random value in the interval [0,1), where λ is a constant.

7. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 5, characterized in that: S3 specifically includes, according to the priority order of the subtasks, for each subtask Execute steps S31-S40 until all subtasks in all DL tasks are completed, and then you can get {x kmnl }, Among them, S31-S40 include: S31: Determine whether l=1 or Is it true? If so, execute S32; if not, execute S33; S32: Make x kmnl =0, and Placed on the ground k The calculation queue The tail of the team; S33: Determine x according to the following formula kmnl ,Right now: in, Represents a subtask Unload to its associated drone m Time subtask completion time; Represents a subtask The completion time during local execution, then, execute S34; S34: Judge x kmnl =1 whether it is true, if it is true, it means the subtask Need to be uninstalled to u m If the condition is not met, execute S35; if not, it means that the subtask If it is executed locally and does not need to be uninstalled, then S36 is executed; S35: Add to u m The calculation queue At the end of the team, then, execute S37; S36: Will Add to ground users k The calculation queue of At the end of the team, then, execute S37; S37: Judgment exist Li Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Execute locally, subtask Need to be unloaded to the drone m , execute S38; otherwise, execute S39; S38: Join Ground Users k Transmission queue The tail of the team; S39: Judgment exist And Is it true? If so, what are the subtasks? Dependencies on subtasks And, the subtask Unloaded on drone m , and the subtask If local execution is required, execute S40; S40: Join UAV m Transmission queue The end of the team.

8. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 4, characterized in that: The UAV location deployment problem is to fix the unloading decision {x kmnl }, optimize the UAV position deployment {q m }, transform this subproblem into P3, which is expressed as: in, C6-C8 represent constraints; and Respectively represent the uplink and downlink rates at the rth iteration; Indicates an increase in the uplink rate; Indicates an increase in the downlink rate; η represents the minimum lifting rate; Indicates user s k and droneu m Is there data transmission on the link between them? If y km >0 indicates user s k and droneu m There is data transmission between the links, otherwise y km =0.

9. A multi-UAV assisted edge computing joint offloading and deployment method according to claim 8, characterized in that: The method further includes converting problem P3 into problem P4, where P4 is expressed as: in, C9-C11 represent constraints; Indicates the uplink rate The lower bound of Indicates the downlink rate The lower bound of H represents the vertical height of the drone from the ground, B is the channel bandwidth, Indicates the received signal-to-noise ratio at 1 meter, P k For k The transmission power, σ 2 is the background noise power; Where P m is u m The transmission power; β0 is the channel power gain at a reference distance of 1 meter; the superscript r is added t The q represents the t The horizontal position of the drone at the first internal iteration; and Respectively expressed in the rth t Two local points at the inner iteration, Problem P4 is a standard convex optimization problem. In S15, given When, through the SCA method, with the given Solve problem P4 and get the optimal solution