An unmanned aerial vehicle assisted mobile edge computing system and a dependency-aware service migration method for the system
Patent Information
- Application Number
- CN202610929429.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-25
- Publication Date
- 2026-10-13
AI Technical Summary
上述方法在无人机辅助MEC场景下仍存在明显不足：例如，就近或局部贪心策略往往忽略微服务依赖耦合与迁移成本的长期累积，容易造成关键路径时延上升或频繁迁移；阈值触发策略难以在不同移动强度与负载水平下自适应权衡；强化学习等方法虽具备一定的自学习能力，但通常需要较长训练周期与大量样本，且对环境变化的适应性与可解释性难以保证，难以直接满足在线实时与稳定性要求
[0037]1、本发明通过依赖DAG关键路径建模，迁移决策直接优化对用户体验最关键的端到端时延组成，避免“迁移单点服务但引入更大跨节点依赖通信”的问题；
Smart Images
Figure CN122909615A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a drone-assisted mobile edge computing system and a dependency-aware service migration method for the system, belonging to the fields of mobile edge computing and drone communication network technology. Background Technology
[0002] With the development of low-latency IoT applications and 6G networks, mobile edge computing reduces service latency and improves service experience by bringing computing and storage capabilities closer to the user. In recent years, unmanned aerial vehicles (UAVs), with their advantages of on-demand deployment, mobility, and dynamically adjustable coverage, have been used as mobile edge servers in disaster relief, field operations, and large-scale events, providing mobile users with temporary or supplementary edge computing and connectivity. Compared to the traditional model of service provided by fixed base stations or fixed edge servers, UAV-assisted MEC can quickly form aerial computing and communication infrastructure, providing resilient support when ground infrastructure is damaged, hotspot areas experience sudden congestion, or coverage blind spots appear.
[0003] However, the operating environment and service model of UAV-assisted MEC systems dictate that "service deployment and migration" must face more complex constraints and coupling relationships than traditional MEC. These are mainly reflected in the following aspects: First, dual mobility and dynamic topology cause link quality and coverage to change rapidly over time, meaning an optimal deployment at one moment may quickly become ineffective in the next. Second, the CPU and memory resources of UAV edge nodes are limited, and the deployment after migration must meet capacity feasibility requirements; otherwise, overload, queuing, and a surge in latency are likely to occur. Third, wireless transmission is affected by distance fading, interference, and bandwidth allocation. When the user-associated UAV and the service host UAV are inconsistent, cross-UAV communication latency will be introduced and will fluctuate significantly with topology changes. Fourth, modern applications often adopt a microservice architecture, where a single user request often... Triggering cascading calls between multiple microservices, service dependencies can be described by a directed acyclic graph (DAG). The end-to-end experience is often determined by the critical path in the dependency chain. If migration only focuses on the entry service or the "nearest deployment" of a single service while ignoring the deployment location of its dependent services, it may lead to amplification of dependency communication across drones, resulting in "service chain bloat," where the benefits of migration are offset by greater dependency communication overhead. Fifth, service migration usually requires state transfer and startup or deployment on the target node, which introduces service interruption and additional overhead. If only short-term latency minimization is pursued, it may lead to frequent migrations causing jitter, which in turn consumes a lot of migration costs, reduces stability, and even affects availability. Therefore, migration costs need to be continuously controlled in the form of long-term average constraints or budgets.
[0004] Existing technologies commonly employ strategies such as proximity / greedy deployment based on geographic location or signal quality, heuristic methods that migrate only when a threshold is triggered, and intelligent decision-making using methods like reinforcement learning. However, these methods still have significant shortcomings in UAV-assisted MEC scenarios: for example, proximity or locally greedy strategies often ignore the long-term accumulation of microservice dependency coupling and migration costs, easily leading to increased critical path latency or frequent migrations; threshold-triggered strategies struggle to adaptively balance different movement intensities and load levels; while reinforcement learning and other methods possess some self-learning capabilities, they typically require long training periods and large datasets, and their adaptability and interpretability to environmental changes are difficult to guarantee, making it difficult to directly meet online real-time and stability requirements. Summary of the Invention
[0005] The purpose of this invention is to address the problems existing in the prior art and to provide a UAV-assisted mobile edge computing (MEC) system and a dependency-aware service migration method for the system. This method is used to achieve low latency and migration stability in dual mobility scenarios involving both UAV movement and user movement. It can make online decisions without relying on future mobility information, and can both dependency-awarely optimize end-to-end latency dominated by the critical path and suppress the accumulation of migration costs through long-term constraint mechanisms. Thus, it achieves a balance between low latency and migration stability in dynamic UAV-assisted MEC systems.
[0006] The core idea of this invention is to unify the short-term performance goal of "minimizing dependency-aware end-to-end latency" with the long-term stability goal of "long-term budget constraints on migration costs" into the same online optimization framework. On the one hand, it uses a directed acyclic graph of microservice dependencies to characterize inter-service calls and data interactions, and uses the critical path to characterize the dominant latency of the end-to-end experience. On the other hand, it introduces a virtual queue to track the degree of overrun of migration costs relative to the budget, and uses a drift plus penalty approach in each time slot to transform the long-term constraint problem into an online decision sub-problem that depends only on the current observable state. Furthermore, to address the combinatorial complexity of the sub-problems in each time slot, it uses an iterative greedy search to quickly generate a migration plan that can simultaneously include migration operations for multiple service instances, thereby achieving a balance between low latency and migration stability while ensuring feasibility (CPU / memory capacity, deployment uniqueness, and migration logic constraints).
[0007] The objective of this invention is achieved through the following technical solution:
[0008] A drone-assisted mobile edge computing system includes: multiple drone edge nodes, one or more user terminals, microservice instances, and a central controller;
[0009] The drone edge node provides limited CPU and memory resources to host service instances and provide edge computing services to mobile users.
[0010] The user terminal is used by mobile users within the ground area. It accesses the edge nodes of the UAV within the coverage area via a wireless link, sends service requests to the UAV edge nodes, and receives response results processed by edge computing.
[0011] Microservice instances are deployed on the edge nodes of drones to receive and process service requests from user terminals. Each microservice instance cascades calls through the call dependency relationship described by the directed acyclic graph to jointly complete end-to-end service processing.
[0012] The central controller is used to collect system status information of each UAV edge node through the uplink at the beginning of each discrete time slot, run the dependent perception service migration algorithm to generate a migration plan, and distribute the migration instructions to the relevant UAV edge nodes for execution through the downlink, while maintaining the global service deployment status and virtual queues.
[0013] A dependency-aware service migration method for a UAV-assisted mobile edge computing system, which operates within discrete time slots, includes the following steps:
[0014] Step S1, System State Modeling: At the beginning of each time slot, the central controller collects the currently observable system state information;
[0015] Step S2: Construct the optimization problem: Construct an optimization problem with the objective of minimizing the long-term average end-to-end service latency and constrained by the long-term average migration cost. Based on the system state information obtained in step S1, calculate the end-to-end service latency and migration cost of each user request in the current time slot.
[0016] Step S3: Introduce a virtual queue: Construct a virtual queue for the long-term average migration cost budget constraint in step S2 to track the cumulative overrun of migration costs relative to the budget; update the virtual queue at the end of each time slot according to preset rules and keep it non-negative.
[0017] Step S4, Drift plus penalty transformation: Based on the virtual queue introduced in step S3, the Lyapunov drift plus penalty method is used to transform the long-term optimization problem constrained by the long-term average migration cost budget in step S2 into a per-slot objective function that is solved only by the current observable state in each slot. Minimizing the per-slot objective function is equivalent to approximately minimizing the upper bound of the drift plus penalty under the current observable state.
[0018] Step S5: Greedy solution and execution of the migration plan for each time slot: Using the empty migration plan as the initial solution, under the premise of satisfying the CPU and memory resource capacity constraints of each drone, the uniqueness constraint of microservice instance deployment, and the migration logic constraint, iteratively traverse the migration operations of candidate single microservice instances and target drones, evaluate the improvement of the objective function of each time slot in step S4 after adding the candidate migration to the current migration plan, select the candidate migration with the largest improvement to add to the current plan and update the resource projection status, repeat the above process until the objective function can no longer be reduced or the preset maximum number of iterations is reached; execute the obtained migration plan in the current time slot, update the deployment status of each microservice instance, and update the virtual queue according to the preset rules in step S3 based on the actual migration cost of this time slot.
[0019] Preferably, the system status information in step S1 includes: the current location of each user terminal, the current location of each UAV edge node, the association between each user terminal and the UAV edge node, the bandwidth allocated to each user terminal, the transmission power of the user terminal, the current deployment status and average request data size of each microservice instance and its status data size, the CPU, memory, and bandwidth resources of each UAV edge node, and the directed acyclic graph of microservice dependencies.
[0020] In the dependent directed acyclic graph, nodes represent service instances, edges represent service call dependencies, and edge weights represent the amount of data transmitted during dependency calls. The critical path dependency latency is obtained by recursively calculating the dependent directed acyclic graph, and the maximum cumulative dependency access latency starting from the target service instance is taken as the critical path dependency latency.
[0021] Preferably, the migration decision in step S1 is used to indicate whether to migrate a service instance from its current host drone edge node to the target drone edge node, and to update the service deployment status after executing the migration decision; the update includes: when a service instance is migrated from the source drone to the target drone, canceling its deployment identifier on the source drone and establishing its deployment identifier on the target drone.
[0022] Preferably, the average end-to-end service latency in step S2 is obtained by averaging the set of valid user requests within the current time slot;
[0023] The end-to-end service latency includes: user access transmission latency, cross-drone forwarding transmission latency when the service host drone and the user-associated drone are inconsistent, critical path dependency latency determined by the dependent directed acyclic graph, and service processing latency.
[0024] The critical path dependency latency is obtained by recursively calculating the dependency directed acyclic graph, taking the maximum value of the dependency access latency accumulated layer by layer along the call chain starting from the entry microservice instance;
[0025] The migration cost of the current time slot is the sum of the transmission time of the microservice instance status data of all migration operations on the inter-unmanned link and the startup or deployment time of the service instance on the target node.
[0026] Preferably, the long-term average migration cost constraint in step S2 is: the time average of the migration cost of each time slot in the entire runtime domain does not exceed the preset migration cost budget C_max; the constraint is guaranteed by the virtual queue introduced in step S3: when the migration cost of a time slot exceeds the budget C_max, the virtual queue backlog increases; when the migration cost of a time slot is lower than the budget C_max, the virtual queue backlog decreases; if the virtual queue remains bounded in long-term operation, the long-term average migration cost can be derived from the queue stability theory to not exceed C_max, that is, the long-term average migration cost constraint is satisfied.
[0027] Preferably, the preset rule formula in step S3 is as follows:
[0028] Q(t+1)=max(0, Q(t)+ C^migration(t)-C_max)
[0029] Where C^migration(t) is the total migration cost actually incurred in this time slot, and C_max is the preset average migration cost budget; by controlling the virtual queue to remain stable in the long term, the long-term average migration cost is guaranteed not to exceed the budget.
[0030] Preferably, the objective function formula for each time slot in step S4 is as follows:
[0031] G(t)=V·delay_avg(t)+Q(t)·C^migration(t)
[0032] Where V is the control parameter, delay_avg(t) is the average end-to-end service delay after executing the time slot migration decision, Q(t) is the virtual queue size of the time slot, and C^migration(t) is the total migration cost of the time slot migration decision.
[0033] Preferably, the control parameter V is used to balance the average end-to-end latency and the degree to which migration cost constraints are met, wherein a larger V emphasizes reducing the average end-to-end latency, and a smaller V emphasizes suppressing virtual queue backlog caused by migration costs.
[0034] Preferably, the migration logic constraints in step S5 include at least the following: the same service instance can only migrate once within the same time slot, and the migration operation only allows the migration from the host drone of the service instance at the beginning of the time slot to the target drone;
[0035] The service deployment status is used to indicate which drone edge node each service instance is deployed on in the current time slot, and each service instance is deployed on only one drone edge node in any time slot.
[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0037] 1. This invention utilizes DAG critical path modeling to directly optimize the end-to-end latency components most critical to user experience through migration decisions, avoiding the problem of "migrating a single point of service but introducing greater cross-node dependent communication";
[0038] 2. This invention achieves constraint control on the long-term average migration cost by introducing a virtual queue and Lyapunov online optimization, thereby suppressing migration oscillations and improving system stability;
[0039] 3. In this invention, each time slot relies solely on the current observable state for online decision-making, without requiring future movement information, making it suitable for dynamic topology scenarios;
[0040] 4. This invention uses an iterative greedy approach to solve the per-slot combination subproblem, reducing the solution complexity and facilitating deployment on resource-constrained UAV edge nodes.
[0041] 5. This invention addresses the dynamic optimization challenges arising from the dual mobility of drones and users, as well as the dependency coupling of microservices. It incorporates critical path dependency latency and long-term migration cost constraints into a unified online optimization framework, achieving a balance between low latency and migration stability without requiring future mobility information. It is applicable to drone-assisted edge computing scenarios such as disaster emergency response, temporary coverage, and dynamic hotspots. Attached Figure Description
[0042] Figure 1 This is an overall flowchart of a drone-assisted mobile edge computing system and a dependency-aware service migration method for the system according to the present invention.
[0043] Figure 2 This is a schematic diagram of an emergency communication network system for unmanned aerial vehicles (UAVs).
[0044] Figure 3 A flowchart of the iterative greedy algorithm executed by the central controller.
[0045] Figure 4 This is a schematic diagram of information and decision-making flow in a drone network. Detailed Implementation
[0046] The present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and detailed implementation methods are given, but the protection scope of the present invention is not limited to the following embodiments.
[0047] like Figure 1 As shown, the UAV-assisted mobile edge computing system involved in this embodiment includes: multiple UAV edge nodes, one or more user terminals, microservice instances, and a central controller;
[0048] The drone edge node provides limited CPU and memory resources to host service instances and provide edge computing services to mobile users.
[0049] The user terminal is used by mobile users within the ground area. It accesses the edge nodes of the UAV within the coverage area via a wireless link, sends service requests to the UAV edge nodes, and receives response results processed by edge computing.
[0050] Microservice instances are deployed on the edge nodes of drones to receive and process service requests from user terminals. Each microservice instance cascades calls through the call dependency relationship described by the directed acyclic graph to jointly complete end-to-end service processing.
[0051] The central controller is used to collect information such as the real-time location of each UAV edge node, user location and association, remaining capacity of resources such as CPU and service instance deployment status through the uplink at the beginning of each discrete time slot. It runs the dependency-aware service migration algorithm to generate a migration plan and distributes the migration instructions to the relevant UAV edge nodes for execution through the downlink. At the same time, it maintains the global service deployment status and virtual queue.
[0052] A method for migrating dependency-aware services in a drone-assisted mobile edge computing system includes the following steps:
[0053] Step S1: System State Modeling
[0054] This invention divides the operation of a UAV-assisted mobile edge computing system into discrete time slots. At the beginning of each time slot, online decisions are made based on the current observable state. System state modeling includes the following:
[0055] (1) System composition: The system includes: multiple UAV edge nodes as mobile edge servers, one or more user terminals, multiple mobile users, multiple service instances and a central controller; the UAV edge nodes provide limited CPU and memory resources to support service instances and provide edge computing services to users.
[0056] (2) Movement and association: Users and drones move in the airspace and their positions change with time slots; the association between users and drones can be determined by principles such as "nearest coverage" or "best signal quality", so that users can connect to a certain available drone edge node in each time slot.
[0057] (3) Deployment and Migration Representation: The deployment status of a service instance indicates which drone edge node each service instance is deployed on in the current time slot, and each service instance is deployed on only one drone edge node in any time slot; the migration decision indicates whether to migrate a service instance from its current host drone to the target drone edge node, and updates the service deployment status after executing the migration decision. The update includes: when a service instance migrates from the source drone to the target drone, canceling its deployment identifier on the source drone and establishing its deployment identifier on the target drone.
[0058] (4) Communication and computation: The access transmission delay of the user-UAV link and the communication delay of the UAV-UAV link are determined by factors such as bandwidth allocation and link distance; the service processing delay is determined by the computational workload of the service instance and the computational resources allocated / reserved for it.
[0059] (5) Dependencies and critical paths: Microservice dependencies are represented by a directed acyclic graph (DAG), and edge weights are used to characterize the amount of data transmitted during dependency calls; critical path dependency latency is characterized by the critical path, and is the maximum cumulative dependency access latency obtained by accumulating along the dependency call chain starting from the entry service instance.
[0060] In the dependent directed acyclic graph, nodes represent service instances, edges represent service call dependencies, and edge weights represent the amount of data transmitted during dependency calls. The critical path dependency latency is obtained by recursively calculating the dependent directed acyclic graph, and the maximum cumulative dependency access latency starting from the target service instance is taken as the critical path dependency latency.
[0061] (6) End-to-end latency definition: End-to-end service latency includes user access transmission latency, cross-UAV communication latency, critical path dependency latency, and service processing latency;
[0062] The average end-to-end service latency is obtained by averaging the set of end-to-end service latencies of valid user requests within each time slot. The end-to-end service latency of a single request consists of the following components: user access transmission latency, cross-drone transmission latency when the service host drone is different from the user's associated drone, critical path dependency latency, and service processing latency.
[0063] The communication rates between users and drones, as well as between drones, are calculated based on the Shannon capacity formula and channel gain; and the system bandwidth is allocated proportionally to user requests to obtain allocated bandwidth, thereby determining the user access transmission delay.
[0064] (7) Migration cost model: Migration cost is represented by service interruption time, including the transmission time of status data on the UAV-UAV link and the start-up or deployment time of the service at the target node; the total migration cost within a time slot is the sum of the costs of all migration operations within that time slot.
[0065] like Figure 2 As shown, the online migration decision-making method of this invention is uniformly managed and executed by a central controller (such as a ground base station or a pre-set navigation drone). The central controller and the drone edge nodes exchange information via a bidirectional link. At the beginning of each time slot t, each drone edge node aggregates its current real-time location, user location / association, remaining CPU / memory capacity, and list of currently supported service instances to the central controller via the uplink. The system status information specifically includes: the current location of each user terminal, the current location of each drone edge node, the association between each user terminal and the drone edge node, the bandwidth allocated to each user terminal, the transmit power of the user terminal, the current deployment status and average request data size of each microservice instance, its status data size, the CPU, memory, and bandwidth resource status of each drone edge node, and a directed acyclic graph of microservice dependencies.
[0066] Step S2: Construct the optimization problem
[0067] Construct an optimization problem with the objective of minimizing long-term average end-to-end service latency and constrained by long-term average migration cost budget, and calculate latency and cost metrics for decision-making based on the state information in step S1 in each time slot.
[0068] (1) End-to-end service delay: For any user request, the end-to-end service delay includes user access transmission delay, cross-UAV communication delay, critical path dependency delay and service processing delay; the average end-to-end service delay is obtained by averaging the end-to-end service delay of valid user requests within the time slot.
[0069] (2) Critical path dependency latency: The critical path dependency latency is determined as follows: For the entry service instance, the dependency access latency is accumulated layer by layer along the dependency call chain in the directed acyclic graph of the dependency, starting from the service instance, and the maximum accumulated value is taken as the critical path dependency latency; the dependency access latency includes the dependency communication latency and the processing latency of the dependent service; when there is no dependency call relationship for the entry service instance, the critical path dependency latency is zero.
[0070] (3) Migration cost: The cost is obtained by summing the migration operations of the service instance. The single migration cost of any service instance from the source UAV to the target UAV includes the sum of the transmission time of state data on the UAV-UAV link and the start-up or deployment time of the target node. The state data transmission time is determined by the state size of the service instance and the transmission rate of the UAV-UAV link. The total migration cost of time slot t is the sum of the migration costs of all migration operations in that time slot.
[0071] (4) Immediate constraints: including UAV CPU and memory resource capacity constraints, service deployment uniqueness constraints and migration logic constraints; among which, the resource capacity constraint is that the sum of the CPU requirements of all service instances deployed on any UAV in any time slot t does not exceed the CPU capacity of the UAV and the sum of the memory requirements does not exceed the memory capacity of the UAV; the migration logic constraint includes at least: the same service instance can be migrated at most once in the same time slot, and the migration operation is only allowed to migrate from the host UAV at the beginning of the time slot to the target UAV.
[0072] Step S3: Construct a virtual queue
[0073] To ensure that the long-term average migration cost does not exceed the budget C_max, this invention introduces a virtual queue Q(t) to track the cumulative amount of "overspending / surplus of migration cost relative to the budget". The virtual queue does not store real data but serves as an "ledger" for migration cost constraints: the long-term average migration cost constraint is that the time average of migration costs in each time slot throughout the entire runtime domain does not exceed the preset migration cost budget C_max. This constraint is guaranteed by the virtual queue introduced in step S3: when the migration cost of a time slot exceeds the budget C_max, the virtual queue backlog increases; when the migration cost of a time slot is lower than the budget C_max, the virtual queue backlog decreases. If the virtual queue remains bounded during long-term operation, the long-term average migration cost can be deduced from queue stability theory to not exceed C_max, meaning the long-term average migration cost constraint is satisfied.
[0074] Preferably, at the end of the time slot, the update is performed according to the following preset rule and kept non-negative:
[0075] Q(t+1)=max(0, Q(t)+C^migration(t)-C_max)
[0076] Where C^migration(t) is the total migration cost of the migration decision in this time slot; by controlling Q(t) to remain stable in the long run, the constraint on the long-term average migration cost is satisfied.
[0077] The long-term average migration cost constraint is: the long-term average migration cost constraint does not exceed the migration cost budget, and the stability of the virtual queue ensures that the long-term average migration cost constraint is met.
[0078] Step S4, Drift plus penalty conversion
[0079] This step transforms a constrained optimization problem requiring long-term statistical information into an online decision subproblem where each time slot depends only on the currently observable state. Based on the drift-penalty concept, the objective function G(t) in each time slot is a combination of the "weighted average end-to-end service latency based on control parameter V" and the "migrating cost weighted by virtual queue backlog," as shown in the following formula:
[0080] G(t)=V·delay_avg(t)+Q(t)·C^migration(t)
[0081] Where V is the control parameter, delay_avg(t) is the average end-to-end service delay after executing the time slot migration decision, Q(t) is the virtual queue size of the time slot, and C^migration(t) is the total migration cost of the time slot migration decision;
[0082] The control parameter V is used to balance the degree to which the average end-to-end latency and migration cost constraints are met, thereby balancing the reduction of latency and the suppression of migration costs: a larger V focuses more on reducing the average end-to-end latency, which may lead to more aggressive migration; a smaller V focuses more on suppressing queue backlog caused by migration costs, which may sacrifice some latency performance.
[0083] Step S5: Solve and execute the migration plan for each time slot using a greedy algorithm.
[0084] Since the subproblem in each time slot is a combinatorial optimization problem, this invention employs an iterative greedy search to generate a migration plan, which may include migration operations for one or more service instances. To ensure feasibility and consistency, the migration plan must satisfy CPU and memory resource capacity constraints, service deployment uniqueness, and migration logic constraints. Among these, the migration logic constraints include at least the following: the same service instance can only be migrated once within the same time slot, and the migration operation is only allowed to migrate from the host drone at the beginning of that time slot to the target drone.
[0085] like Figure 3 As shown, the greedy solution process includes:
[0086] (1) Initialization: Using an empty migration plan as the initial solution, calculate the corresponding objective function value;
[0087] (2) Candidate generation: Traverse the migration operations of the candidate "single service instance - target drone" and remove candidates that do not meet the CPU and memory resource capacity constraints or migration logic constraints;
[0088] (3) Evaluation and selection: For each candidate migration, evaluate the degree of improvement to the objective function after it is added to the current migration plan, and select the candidate migration with the greatest improvement to update the migration plan;
[0089] (4) Iteration termination: Repeat the candidate generation and evaluation selection process until the objective function can no longer be reduced or the preset maximum number of iterations is reached;
[0090] (5) Execution and Update: Execute the migration plan in time slot t, update the service deployment status, and update the virtual queue according to the migration cost of this time slot.
[0091] The service deployment status is used to indicate which drone edge node each service instance is deployed on in the current time slot, and each service instance is deployed on only one drone edge node in any time slot.
[0092] In short, such as Figure 4 As shown, at the beginning of each discrete time slot, the central controller uses the algorithm described above to solve for the migration plan based on the current observable system state. Then, the central controller generates a migration instruction set containing the source node, target node, and service instance ID, and distributes it to the relevant drone nodes via the downlink. Upon receiving the instruction, the source drone involved in the migration packages the corresponding microservice state data and transmits it to the target drone via the inter-drone link. After completing the deployment switch, the target drone reports its execution status back to the central controller, which then updates the global deployment status matrix and virtual queue accordingly. User terminals access drone edge nodes within their coverage area via wireless links, send service requests to these edge nodes, receive response results processed by edge computing, and report their status to the central controller.
[0093] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A drone-assisted mobile edge computing system, characterized in that, include: Multiple drone edge nodes, one or more user terminals, microservice instances, and a central controller; The drone edge node provides limited CPU and memory resources to host service instances and provide edge computing services to mobile users. The user terminal is used by mobile users within the ground area. It accesses the edge nodes of the UAV within the coverage area via a wireless link, sends service requests to the UAV edge nodes, and receives response results processed by edge computing. Microservice instances are deployed on the edge nodes of drones to receive and process service requests from user terminals. Each microservice instance cascades calls through the call dependency relationship described by the directed acyclic graph to jointly complete end-to-end service processing. The central controller is used to collect system status information of each UAV edge node through the uplink at the beginning of each discrete time slot, run the dependent perception service migration algorithm to generate a migration plan, and distribute the migration instructions to the relevant UAV edge nodes for execution through the downlink, while maintaining the global service deployment status and virtual queues.
2. A dependency-aware service migration method for a UAV-assisted mobile edge computing system based on claim 1, characterized in that, The method operates within discrete time slots and includes the following steps: Step S1, System State Modeling: At the beginning of each time slot, the central controller collects the currently observable system state information; Step S2: Construct the optimization problem: Construct an optimization problem with the objective of minimizing the long-term average end-to-end service latency and constrained by the long-term average migration cost. Based on the system state information obtained in step S1, calculate the end-to-end service latency and migration cost of each user request in the current time slot. Step S3: Introduce a virtual queue: Construct a virtual queue for the long-term average migration cost budget constraint in step S2 to track the cumulative overrun of migration costs relative to the budget; update the virtual queue at the end of each time slot according to preset rules and keep it non-negative. Step S4, Drift plus penalty transformation: Based on the virtual queue introduced in step S3, the Lyapunov drift plus penalty method is used to transform the long-term optimization problem constrained by the long-term average migration cost budget in step S2 into a per-slot objective function that is solved only by the current observable state in each slot. Minimizing the per-slot objective function is equivalent to approximately minimizing the upper bound of the drift plus penalty under the current observable state. Step S5: Greedy solution and execution of the migration plan for each time slot: Using the empty migration plan as the initial solution, under the premise of satisfying the CPU and memory resource capacity constraints of each drone, the uniqueness constraint of microservice instance deployment, and the migration logic constraint, iteratively traverse the migration operations of candidate single microservice instances and target drones, evaluate the improvement of the objective function of each time slot in step S4 after adding the candidate migration to the current migration plan, select the candidate migration with the largest improvement to add to the current plan and update the resource projection status, repeat the above process until the objective function can no longer be reduced or the preset maximum number of iterations is reached; execute the obtained migration plan in the current time slot, update the deployment status of each microservice instance, and update the virtual queue according to the preset rules in step S3 based on the actual migration cost of this time slot.
3. The dependency-aware service migration method according to claim 2, characterized in that, The system status information mentioned in step S1 includes: the current location of each user terminal, the current location of each UAV edge node, the association between each user terminal and the UAV edge node, the bandwidth allocated to each user terminal, the transmission power of the user terminal, the current deployment status and average request data size of each microservice instance and its status data size, the CPU, memory, and bandwidth resources of each UAV edge node, and the directed acyclic graph of microservice dependencies. In the dependent directed acyclic graph, nodes represent service instances, edges represent service call dependencies, and edge weights represent the amount of data transmitted during dependency calls. The critical path dependency latency is obtained by recursively calculating the dependent directed acyclic graph, and the maximum cumulative dependency access latency starting from the target service instance is taken as the critical path dependency latency.
4. The dependency-aware service migration method according to claim 2, characterized in that, The migration decision in step S1 is used to indicate whether to migrate a service instance from its current host drone edge node to the target drone edge node, and to update the service deployment status after executing the migration decision; the update includes: when a service instance is migrated from the source drone to the target drone, canceling its deployment identifier on the source drone and establishing its deployment identifier on the target drone.
5. The dependency-aware service migration method according to claim 2, characterized in that, The average end-to-end service latency mentioned in step S2 is obtained by averaging the set of valid user requests within the current time slot; The end-to-end service latency includes: user access transmission latency, cross-drone forwarding transmission latency when the service host drone and the user-associated drone are inconsistent, critical path dependency latency determined by the dependent directed acyclic graph, and service processing latency. The critical path dependency latency is obtained by recursively calculating the dependency directed acyclic graph, taking the maximum value of the dependency access latency accumulated layer by layer along the call chain starting from the entry microservice instance; The migration cost of the current time slot is the sum of the transmission time of the microservice instance status data of all migration operations on the inter-unmanned link and the startup or deployment time of the service instance on the target node.
6. The dependency-aware service migration method according to claim 2, characterized in that, The long-term average migration cost constraint mentioned in step S2 is: the time average of the migration cost of each time slot in the entire runtime domain does not exceed the preset migration cost budget C_max; the constraint is guaranteed by the virtual queue introduced in step S3: when the migration cost of a time slot exceeds the budget C_max, the virtual queue backlog increases; when the migration cost of a time slot is lower than the budget C_max, the virtual queue backlog decreases; if the virtual queue remains bounded in long-term operation, it can be deduced from the queue stability theory that the long-term average migration cost does not exceed C_max, that is, the long-term average migration cost constraint is satisfied.
7. The dependency-aware service migration method according to claim 2, characterized in that, The preset rule formula mentioned in step S3 is as follows: Q(t+1)=max(0, Q(t)+ C^migration(t)-C_max) Where C^migration(t) is the total migration cost actually incurred in this time slot, and C_max is the preset average migration cost budget; by controlling the virtual queue to remain stable in the long term, the long-term average migration cost is guaranteed not to exceed the budget.
8. The dependency-aware service migration method according to claim 2, characterized in that, The objective function formula for each time slot mentioned in step S4 is as follows: G(t)=V·delay_avg(t)+Q(t)·C^migration(t) Where V is the control parameter, delay_avg(t) is the average end-to-end service delay after executing the time slot migration decision, Q(t) is the virtual queue size of the time slot, and C^migration(t) is the total migration cost of the time slot migration decision.
9. The dependency-aware service migration method according to claim 8, characterized in that, The control parameter V is used to balance the average end-to-end latency and the degree to which migration cost constraints are met. The larger V is, the more emphasis is placed on reducing the average end-to-end latency, and the smaller V is, the more emphasis is placed on suppressing virtual queue backlog caused by migration costs.
10. The dependency-aware service migration method according to claim 2, characterized in that, The migration logic constraints in step S5 include at least the following: the same service instance can only migrate once within the same time slot, and the migration operation is only allowed to migrate from the host drone at the beginning of the time slot to the target drone. The service deployment status is used to indicate which drone edge node each service instance is deployed on in the current time slot, and each service instance is deployed on only one drone edge node in any time slot.