Multi-access edge computing resource scheduling method and system for 5G communication system

Abstract

The invention relates to the technical field of communication, discloses a 5G communication system multi-access edge computing resource scheduling method and system, and aims to solve the problems that service interruption of MEC service migration in a high-speed mobile scene is too long due to a reactive strategy, and the requirement that uRLLC service end-to-end delay is lower than 10 milliseconds cannot be met. The method comprises the following steps: acquiring a wireless signal intensity sequence, a historical track and a service state of user equipment; constructing a space-time coupling trajectory prediction model to generate a high-confidence target cell sequence; a lightweight container mirror image parallel flow type synchronous increment state is preloaded at the candidate edge node; and after base station switching is completed, a new container is immediately activated to take over service, and source node resources are recycled. According to the invention, through a prediction-driven closed-loop scheduling mechanism, the service interruption time is shortened to be within 5 milliseconds, and the resource utilization efficiency and the service reliability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a method and system for scheduling multi-access edge computing resources in a 5G communication system. Background Technology

[0002] Multi-access edge computing (MEC), as a key component of 5G networks, effectively meets the low latency and high reliability requirements of scenarios such as connected vehicles and industrial control by pushing computing power down to the network edge. When user equipment moves between base stations, ensuring the continuous and smooth migration of its running edge services to edge nodes near the new access point is a core challenge in guaranteeing a smooth user experience.

[0003] Current mainstream service migration methods are mostly "reactive" strategies, meaning the migration process only begins after the user handover is complete. This process typically involves service interruption, full state transmission, and target node reconstruction, resulting in a long overall time consumption and difficulty in meeting the requirements of ultra-low latency services. Some improved solutions attempt to introduce mobility prediction, such as prediction based on historical handover probabilities, but often fail to fully integrate real-time wireless signal changes with the user's actual movement trajectory, leading to insufficient prediction accuracy in high-speed mobile environments. Furthermore, existing methods often directly migrate the entire service image after prediction, consuming significant bandwidth and storage resources, and lack resource recovery mechanisms after prediction failures, easily leading to resource waste. From a system architecture perspective, the existing MEC standards lack sufficient coordination among mobility management and resource scheduling modules, failing to form an efficient closed-loop scheduling capability.

[0004] Therefore, existing technologies still lack a MEC service scheduling solution that can simultaneously achieve accurate prediction, lightweight pre-provisioning, seamless synchronization, and efficient resource recovery in high-speed mobile scenarios. Summary of the Invention

[0005] This invention provides a method and system for scheduling multi-access edge computing resources in 5G communication systems. It aims to address the technical problem that existing MEC service migration mechanisms, due to their reactive strategies, suffer from excessively long service interruptions and fail to meet the end-to-end latency requirement of less than 10 milliseconds for ultra-reliable low-latency communication services in high-speed user mobility scenarios. This invention achieves seamless switching of MEC services between edge nodes by constructing a closed-loop scheduling architecture that integrates wireless-side mobility awareness, high-precision trajectory prediction, service state pre-synchronization, and dynamic resource reclamation.

[0006] According to one aspect of the present invention, a method for scheduling multi-access edge computing resources in a 5G communication system is provided, comprising: S1 collects real-time wireless signal strength sequences, historical movement trajectory data, and status information of currently running service instances from user equipment; S2, Based on the wireless signal strength sequence and historical mobile trajectory data, a spatiotemporally coupled trajectory prediction model is constructed to generate a high-confidence target cell sequence within a future preset time window; S3, based on the target cell sequence, determine at least one candidate edge node, and preload a lightweight container image that is isomorphic to the current service instance on the candidate edge node; S4, synchronize the incremental status data of the current service instance to the preloaded container of the candidate edge node in a streaming manner; S5, after the user equipment enters the target cell and completes the base station handover, the preloaded container on the candidate edge node is immediately activated to take over the service request; S6 initiates a resource reclamation process on the source edge node after the service switch is completed, releasing the computing and storage resources occupied by the migrated service instances.

[0007] Furthermore, in step S1: The acquisition of the wireless signal strength sequence specifically includes: obtaining the reference signal received power and reference signal received quality values ​​reported by the user equipment in multiple consecutive measurement cycles, with each measurement cycle having a time interval of 20 milliseconds; The collection of the historical movement trajectory data specifically includes: acquiring the sequence of latitude and longitude coordinate points recorded by the user device every 10 seconds in the past hour; The collection of service instance status information specifically includes: obtaining the memory snapshot, network connection table, session context, and application layer state variable set of the current service instance through the container runtime interface.

[0008] Furthermore, the construction of the spatiotemporally coupled trajectory prediction model in step S2 includes: The wireless signal strength sequence is processed by a sliding window, and the first-order difference and second-order difference are calculated for the signal strength value in each window to form a gradient feature vector of signal strength change. The historical movement trajectory data is converted to a local Cartesian coordinate system with the current serving base station as the origin, and the velocity vector and acceleration vector are calculated. The signal intensity change gradient feature vector is time-aligned and feature-concatenated with the velocity vector and acceleration vector to form a fused feature vector sequence, which is then input into the sequence-to-sequence prediction network for training and prediction.

[0009] Furthermore, generating a high-confidence target cell sequence includes: The sequence-to-sequence prediction network outputs a probability matrix of user equipment accessing each adjacent cell within multiple future measurement periods; The probability matrix is ​​summed along the time dimension to obtain the cumulative access probability of each cell within the prediction window. Cells whose cumulative probability exceeds a preset threshold are selected and arranged in descending order of their cumulative probability values ​​to form the target cell sequence.

[0010] Furthermore, the preloading of the lightweight container image in the S3 step includes: Pull a standardized container image corresponding to the current service instance from the image repository. The image adopts a layered structure. The standardized container image is layered and stripped, retaining only the operating system base layer and application dependency library layer, while removing the application code layer and configuration file layer to generate a lightweight base image; Deploy the lightweight base image to candidate edge nodes and mount read-only storage volumes for the preloaded containers.

[0011] Furthermore, the incremental state data synchronization in streaming mode during step S4 includes: Scan the memory page tables of service instances at fixed intervals on the source edge nodes to identify dirty pages that have been modified since the last cycle; Difference encoding is performed on the identified dirty pages, and only the data blocks that have changed are generated into binary difference patch streams containing offsets, lengths and new content; The binary difference patch stream is pushed to the candidate node through a dedicated transport control protocol connection between the source edge node and the candidate edge node; At candidate edge nodes, the received differential patch stream is written into a pre-allocated memory-mapped region in the preloaded container that is consistent with the virtual address layout of the source container.

[0012] Furthermore, the dedicated transmission control protocol connection employs low-latency optimization parameters, including disabling the Nagle algorithm and setting the minimum retransmission timeout to 10 milliseconds.

[0013] Furthermore, the S5 steps include: Listen to the service activation command issued by the core network session management function unit after the user equipment completes the base station handover; According to the service activation instruction, on the corresponding candidate edge node, the network interface of the preloaded container is bound to the service IP address assigned to the user device, and the container state is switched to the running state; Return a service redirection response containing the address of the new container instance service endpoint to the user device.

[0014] Furthermore, the trigger condition for starting the resource reclamation process in step S6 is: confirming that the newly activated container instance has successfully processed at least 10 business requests consecutively. The resource reclamation process includes: terminating the original service instance process on the source edge node, releasing the computing and network resources it occupies, marking the persistent storage volume used by the original service instance as pending cleanup and physically deleting it after a delay time, and updating the global resource view in the MEC orchestrator.

[0015] According to another aspect of the present invention, a multi-access edge computing resource scheduling system for a 5G communication system is provided, comprising: The multi-source data acquisition unit is used to collect in real time the wireless signal strength sequence of user equipment, historical movement trajectory data, and status information of currently running service instances; A high-precision trajectory prediction unit is used to construct a spatiotemporally coupled trajectory prediction model based on the wireless signal strength sequence and historical mobile trajectory data, and generate a high-confidence target cell sequence within a future preset time window. An edge resource preloading unit is used to determine at least one candidate edge node based on the target cell sequence, and preload a lightweight container image that is isomorphic to the current service instance on the candidate edge node; The service status pre-synchronization unit is used to synchronize the incremental status data of the current service instance to the pre-loaded container of the candidate edge node in a streaming manner. The seamless service handover unit is used to immediately activate the preloaded container on the candidate edge node to take over the service request after the user equipment enters the target cell and completes the base station handover. The dynamic resource reclamation unit is used to initiate a resource reclamation process on the source edge node after the service switch is completed, so as to release the resources occupied by the migrated service instances.

[0016] In summary, this application includes at least one of the following beneficial technical effects: (1): This invention fundamentally changes the lagging process of traditional "reactive" migration by constructing a proactive scheduling closed loop of "prediction-preloading-synchronization-switching". Based on high-precision prediction that integrates wireless signal gradients and motion trajectories, the system pre-configures the lightweight service environment to the target node before the actual user handover occurs, and maintains real-time consistency of the state through streaming incremental synchronization. When a handover event occurs, container activation and traffic redirection can be completed in milliseconds, making the end-to-end service interruption time much less than 10 milliseconds, meeting the stringent continuity requirements of services such as uRLLC, and ensuring user experience.

[0017] (2): This invention abandons the crude approach of migrating virtual machines or complete container images as a whole, and adopts a lightweight image preloading strategy of layered stripping, which significantly reduces the amount of data transferred by the image. At the same time, through dirty page identification and differential encoding technology, only the state fragments that change are synchronized, further reducing network bandwidth consumption. Combined with intelligent node screening based on resource thresholds and a dynamic resource reclamation mechanism after the switch is completed, the system effectively avoids resource redundancy reservation and idle waste, and improves the overall resource utilization of the edge computing cluster.

[0018] (3): This invention improves the accuracy of the target cell sequence through a spatiotemporally coupled prediction model, and provides fault tolerance space for the prediction results by setting a confidence threshold and a multi-candidate node mechanism. Streaming state synchronization and delayed resource reclamation strategies provide a buffer for possible rollback or anomaly handling. In addition, each functional unit adopts a loosely coupled design and degradation processing mechanism, so that even when some modules are abnormal, the system can still maintain basic service capabilities, thereby improving the robustness and environmental adaptability of the overall scheduling system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the 5G communication system multi-access edge computing resource scheduling method of the present invention; Figure 2 This is a schematic diagram of the framework of the high-precision spatiotemporal coupled trajectory prediction model of the present invention; Figure 3 This is a flowchart of the multi-access edge computing resource scheduling system for the 5G communication system of the present invention. Detailed Implementation

[0020] This invention provides a method and system for scheduling multi-access edge computing resources in 5G communication systems. Its core lies in constructing a closed-loop scheduling process by integrating wireless-side mobility awareness, high-precision trajectory prediction, service state pre-synchronization, and dynamic resource reclamation mechanisms. This addresses the technical problem of excessively long service interruptions and inability to meet the end-to-end latency requirement of less than ten milliseconds for ultra-reliable low-latency communication services in scenarios with high-speed user mobility, due to the reactive strategy employed by traditional MEC service migration mechanisms. The following detailed implementation of the 5G communication system multi-access edge computing resource scheduling method will be provided in conjunction with the specific steps of this invention.

[0021] The 5G communication system multi-access edge computing resource scheduling method includes the following steps: S1 collects real-time wireless signal strength sequences, historical movement trajectory data, and status information of currently running service instances from user equipment; S2, Based on the wireless signal strength sequence and historical mobile trajectory data, a spatiotemporally coupled trajectory prediction model is constructed to generate a high-confidence target cell sequence within a future preset time window; S3, based on the target cell sequence, determine at least one candidate edge node, and preload a lightweight container image that is isomorphic to the current service instance on the candidate edge node; S4, synchronize the incremental status data of the current service instance to the preloaded container of the candidate edge node in a streaming manner; S5, after the user equipment enters the target cell and completes the base station handover, the preloaded container on the candidate edge node is immediately activated to take over the service request; S6 initiates a resource reclamation process on the source edge node after the service switch is completed, releasing the computing and storage resources occupied by the migrated service instances.

[0022] In step S1, the system uses a multi-source data acquisition unit to collect real-time wireless signal strength sequences, historical movement trajectory data, and current service instance status information of user equipment, providing complete and dynamic input data for subsequent trajectory prediction and resource scheduling. This step specifically includes the following sub-steps: S101: The wireless signal acquisition subunit obtains the reference signal received power and reference signal received quality values ​​reported by the user equipment in multiple consecutive measurement cycles through the wireless resource management module on the base station side. The time interval between each measurement cycle is 20 milliseconds, and data from 100 consecutive cycles is collected to form a signal strength time series with a length of 2000 milliseconds.

[0023] S102: The trajectory data acquisition subunit acquires the sequence of latitude and longitude coordinates recorded by the user equipment every 10 seconds in the past hour through the core network's location management function unit, resulting in a total of 360 coordinate points. The coordinate points are arranged in chronological order to form the historical movement trajectory of the user equipment.

[0024] S103: The service status reading subunit obtains the following status information of the current service instance through the container runtime interface on the MEC host: memory snapshot (organized in units of pages, each page is 4096 bytes in size), network connection table (records all active TCP connections and their status), session context (including user identity, authentication token and session validity period), and application layer state variable set (such as the current playback frame sequence number, buffer fill rate, etc. in the video streaming service).

[0025] The above steps S101-S103 collect real-time and historical data from three dimensions: wireless side, mobility management, and service operation status, respectively, to jointly construct a multi-dimensional dynamic user profile, providing a structured and quantifiable input basis for subsequent high-precision trajectory prediction and service status synchronization.

[0026] S104: The three types of data collected are time-aligned and format-unified within the data acquisition unit, converted into a structured data stream with timestamps, and transmitted to the high-precision trajectory prediction unit via the internal bus.

[0027] In summary, step S1, through a multi-source collaborative acquisition mechanism, achieves real-time, continuous, and structured acquisition of user equipment signal strength, movement trajectory, and service status, laying a reliable data input foundation for the initiation of the entire "prediction-pre-migration" process. Next, the system will perform high-precision trajectory prediction in step S2 based on this data.

[0028] In step S2, based on the wireless signal strength sequence, historical mobile trajectory data, and service instance status information obtained in step S1, step S2 constructs a spatiotemporally coupled trajectory prediction model based on multidimensional data and outputs the sequence of cells that the user equipment is most likely to access in the future, providing accurate target guidance for subsequent resource preloading. Step S2 is executed by the high-precision trajectory prediction unit and specifically includes the following ordered sub-steps: S201: The signal gradient calculation subunit performs sliding window processing on the wireless signal strength sequence (RSRP time series) obtained in step S101. The window length is set to 10 measurement periods (i.e., 200 milliseconds), and the sliding step size is 5 measurement periods. Thus, for a sequence containing 100 sampling points, a total of 19 overlapping time windows are generated. For the RSRP numerical sequence within each window, its first-order difference (reflecting the signal change trend) and second-order difference (reflecting the acceleration of the change trend) are calculated respectively. The first-order and second-order difference values ​​calculated for each window are combined to form a 30-dimensional feature vector, which is used to characterize the dynamic change pattern of the signal strength within that time period.

[0029] S202: The trajectory feature extraction subunit processes the historical latitude and longitude trajectory data obtained in step S102.

[0030] First, transform all coordinate points to a local Cartesian coordinate system with the user's current serving base station as the origin (e.g., using UTM projection).

[0031] Next, using the central difference method, the user's velocity vector (magnitude and direction) is calculated point by point based on the time interval (10 seconds) and position difference between adjacent coordinate points.

[0032] Then, the acceleration vector is calculated based on the changes in adjacent velocity vectors. Finally, the corresponding velocity and acceleration vector features are generated for each historical trajectory point.

[0033] S203: The sequence prediction network subunit first aligns the signal gradient feature vector sequence generated in step S201 with the velocity and acceleration vector sequence generated in step S202 in terms of time. The alignment rule is as follows: the starting timestamp of each signal gradient feature window (corresponding to a 200-millisecond time period) is matched with the motion features (velocity, acceleration) of the trajectory point closest to that timestamp in the historical trajectory data.

[0034] Subsequently, the signal feature vector (30-dimensional) and motion feature vector (4-dimensional) at each alignment time are concatenated along the feature dimension to form a 34-dimensional fused feature vector. This yields an input sequence of length 19 (corresponding to 19 historical time windows), where each time step has a feature dimension of 34.

[0035] The sequence is input into a pre-trained sequence-to-sequence prediction network. This network consists of an encoder and a decoder, both composed of three stacked Gated Recurrent Units (GRUs), with each GRU having a hidden state dimension of 128. The encoder processes the input sequence of length 19 sequentially and uses its final hidden state as the encoded context vector.

[0036] The decoder uses the context vector as the initial hidden state and receives a fixed start tag as the initial input. It then predicts the output for the next 30 measurement periods in an autoregressive manner.

[0037] At each time step of the decoder, its output passes through a fully connected layer (with an input dimension of 128 and an output dimension equal to the number of neighboring cells m), and then through a Softmax activation function, thereby generating the probability distribution of the user's access to each neighboring cell at the current and future time steps.

[0038] The goal of training the network is to make the predicted probability distribution approximate the true switching sequence. During training, the cross-entropy loss function and the Adam optimizer are used.

[0039] S204: After forward propagation through the GRU network, the network's output layer outputs the probability, via the Softmax function, that a user equipment (UE) will access each cell in its current neighbor list (assuming there are m neighbor cells) within a preset time window (e.g., 30 future measurement periods, or 600 milliseconds). The output is a 30-row, m-column probability matrix P. Here, the matrix element P[i][j] represents the predicted probability that the UE will access the j-th neighbor cell (denoted as cⱼ) in the i-th future measurement period.

[0040] S205: Generate a high-confidence target cell sequence.

[0041] The target cell generation sub-unit processes the probability matrix P to filter out high-confidence migration targets. Specifically, it accumulates and sums the predicted probabilities of each cell cⱼ along the time dimension to obtain the total cumulative access probability of that cell over the entire 30 prediction periods. Then, a confidence threshold (e.g., 90%) is set, and all cells with cumulative probabilities exceeding this threshold are selected. Finally, these eligible cells are sorted from highest to lowest cumulative probability value, forming an ordered target cell sequence T={t1,t2,...,t...}. k}, where k is the number of selected cells. This sequence is the list of cells that the user equipment is most likely to switch to in the near future, output by the prediction model.

[0042] In summary, step S2 constructs and runs a spatiotemporal prediction model that deeply integrates the dynamic characteristics of wireless signals and the physical motion characteristics of users. The raw data collected in step S1 is transformed into a sequence of target cells with clear physical meaning and quantified confidence. The output (sequence T) determines the selection of candidate edge nodes and the timing of resource preloading in step S3. This is the core technical link to realize "predictive migration" and break through the bottleneck of delay in traditional reactive migration.

[0043] In step S3, based on the high-confidence target cell sequence T={t1,t2,...,t... output in step S2... k Step S3 involves preparing a lightweight runtime environment for the service instance on the edge node where the user device may switch, thus laying the foundation for seamless state synchronization and service switching. Step S3 is executed by the edge resource preloading unit and specifically includes the following sub-steps: S301: The image layering processing subunit pulls the standardized container image corresponding to the user's current service instance from the image repository managed by the MEC orchestrator. This image adopts a layered structure, including an operating system base layer, an application dependency library layer, an application code layer, and a configuration file layer.

[0044] To reduce the amount of data and time required for preloading, this sub-unit performs the following operations to generate a lightweight base container image: Parse image layer information: Use container image tools to obtain a list of all layers of the standardized image and their dependency order.

[0045] Extract the target layer: Based on the layer description information, identify the specific image layer files that constitute the "operating system base layer" and the "application dependency library layer".

[0046] Create a new image manifest and configuration file. In the new image manifest, only reference the target layer files identified in step 2. Modify the new configuration file based on the original image configuration, removing or clearing the startup commands and working directory settings related to the "Application Code Layer" and "Configuration File Layer".

[0047] The newly generated lightweight image is pushed to a dedicated preloaded image cache repository for candidate nodes to pull.

[0048] For example, a complete image that was originally 500MB in size may be reduced to a new image containing only about 80MB of base and dependency layers after the application code and configuration layer are stripped away using the above method.

[0049] S302: The candidate node selection subunit determines the specific nodes that need to be preloaded based on the target cell sequence T and the real-time resource status of the edge nodes. Its selection logic is as follows: Based on the network topology mapping relationship, obtain one or more physically adjacent edge nodes associated with each cell tᵢ in the target cell sequence T, forming a list of nodes to be checked, and the checking order is consistent with the sequence order of T.

[0050] For each edge node in the list to be inspected, obtain the current real-time resource metrics of that node by calling the resource query interface provided by the MEC orchestrator (or centralized resource manager). These metrics should include at least: CPU utilization (percentage), remaining memory capacity (GB), and network bandwidth idle rate (percentage).

[0051] Determine whether the resource data obtained by this node simultaneously meets all of the following preset threshold conditions: CPU utilization <70%; Remaining memory capacity > 2GB; Network bandwidth idle rate > 50%; If a node meets all the threshold conditions, it is added to the candidate node list. Then, the next node in the list is checked. This process continues until the accumulated number of candidate nodes reaches a preset positive integer N, or the entire list of nodes to be checked has been traversed.

[0052] S303: For each candidate edge node selected in S302, the container deployment subunit performs the deployment operation. Specifically, the lightweight container image generated in S301 is deployed to the container environment of that node through the container runtime interface. When creating a container instance, a dedicated read-only storage volume is mounted to it. This volume is in a read-only state during the preloading phase, used to receive and persist incremental service state data synchronized from the source node, while ensuring that the preloaded container does not modify any state before activation, keeping the environment clean.

[0053] In summary, step S3 utilizes three specific technical means—"image layering and stripping," "resource threshold filtering," and "read-only volume pre-deployment"—to achieve predictive preparation of the service instance's runtime environment. This provides the necessary target carrier for achieving real-time, incremental synchronization of service status in step S4 and is a key prerequisite for achieving the goal of "seamless switching."

[0054] In step S4, after completing the resource preloading in step S3, step S4 synchronizes the dynamic running status of the user's current service instance to the preloading container of the candidate edge node in an efficient and real-time manner. Step S4 is executed by the service status pre-synchronization unit and specifically includes the following sub-steps: S401: The dirty page identification subunit initiates a state capture agent process on the source edge node. This process actively scans the memory page table corresponding to the current service instance at fixed time intervals (e.g., every 5 milliseconds). The memory page table is maintained by the operating system kernel and records the mapping relationship and access attributes between virtual memory pages and physical memory pages. By comparing the page table state of the current period with the page table snapshot saved in the previous period, the state capture agent identifies memory pages that have been modified since the last synchronization period, i.e., "dirty pages". These dirty pages carry the state changes of the service instance in the recent period.

[0055] S402: The differential encoding subunit processes the dirty pages identified by S401. To minimize the amount of data transmitted over the network, this subunit uses a byte-level rolling hash algorithm for differential encoding. Specifically, it compares the current content of the dirty page with the baseline content of the page saved in the previous cycle, calculating and locating the consecutive blocks of bytes that have changed. For each changed block, the encoder generates a binary data block containing its starting offset within the page, the length of the change, and the new byte content. All changed data blocks are sequentially combined to form a compact "binary differential patch stream." This method avoids transmitting the entire dirty page, sending only the actually changed portion.

[0056] S403: The State Transmission Subunit is responsible for reliably delivering the difference patch stream to candidate nodes. It establishes a dedicated TCP (Transmission Control Protocol) connection between the source edge node and each candidate edge node, specifically for transmitting state synchronization data. To achieve low-latency transmission, this connection undergoes targeted parameter optimizations, primarily including disabling the Nagle algorithm (to reduce the transmission latency of small data packets) and setting the minimum retransmission timeout (RTO) to a small value (e.g., 10 milliseconds). The binary difference patch stream generated by the state capture agent process is pushed to the corresponding candidate edge nodes in real-time and streaming through this optimized connection.

[0057] S404: The state application subunit runs on the candidate edge node. It receives the binary difference patch stream from S403 and is responsible for applying it to the preloaded container. Specifically, when the candidate node's container runs, it pre-allocates a memory-mapped region that is completely identical to the virtual address layout of the source container for the preloaded container. The state application subunit parses the received patch stream and, based on the offset and length information, directly writes the new byte content to the specified location in the memory-mapped region corresponding to the preloaded container. Through this "receive and apply simultaneously" streaming processing method, progressive, near real-time synchronization of service state from the source node to the candidate node is achieved.

[0058] Step S4 employs a series of sophisticated technical methods to construct an efficient, low-latency service state synchronization channel. It fully utilizes the pre-loaded container environment prepared in Step S3 (especially the memory-mapped area) to continuously and incrementally "warm up" the dynamically changing application state to the target node. This ensures that when the user device triggers the actual switch (Step S5), the state deviation between the candidate container and the source container is minimal, thus providing the crucial state continuity guarantee for achieving millisecond-level "seamless switchover." This mechanism is the core of overcoming the major bottleneck of excessively long state serialization and transmission times in traditional migration schemes.

[0059] After completing the real-time synchronization of service status from the source node to the candidate node in step S4, step S5, at the instant the user equipment actually switches to the target cell, quickly activates the pre-loaded and synchronized container instances on the candidate edge nodes, and completes the takeover and redirection of service requests, achieving seamless service continuity for the user. Specifically: S501: The handover event listening subunit continuously listens for service activation commands from the core network session management function (SMF) unit. This command is generated and issued by the SMF after the user equipment (UE) completes Radio Resource Control (RRC) connection reconfiguration and successfully accesses the target cell. The service activation command includes at least the UE identifier (such as SUPI or GPSI), the target cell identifier, and the identifier of the service instance to be activated.

[0060] S502: Upon receiving the service activation instruction, the container activation subunit locates the pre-loaded container instance on the corresponding candidate edge node based on the service instance identifier in the instruction. It then performs the following operations: Bind Service IP Address: Call the network namespace operation interface of the container runtime to bind the network interface of the preloaded container to the service IP address pre-assigned to the user device at this target edge node. This service IP address is dynamically assigned by the IP address management module of the MEC host during the preloading phase or during activation.

[0061] Switch container running state: Change the container's state from "Ready" or "Paused" to "Running" to make it start processing business requests.

[0062] S501 precisely captures the handover completion moment by listening to core network standardized signaling (SMF instructions), providing a reliable event source for triggering subsequent actions. Based on this event, S502 quickly activates the prepared container instances and connects them to the service network through specific container network and state management interfaces, preparing for traffic takeover.

[0063] S503: After confirming successful container activation, the client redirection subunit returns a service redirection response to the user equipment. This response is encapsulated in an HTTP 302 status code, and its response header contains the service endpoint address of the newly activated container instance (e.g., a combination of IP address and port). Upon receiving this redirection response, the user equipment automatically sends all subsequent business requests to this new address.

[0064] Step S5 enables a fast and smooth switchover of service instances. It fully leverages the results of the preceding steps (S3 and S4) – pre-loaded and state-synchronized containers – to activate and connect them to the network at the right time, and guide user traffic through application-layer redirection mechanisms, thereby minimizing the impact on users.

[0065] This step, combined with the S4 state synchronization mechanism, ensures that the entire service migration process appears as a "seamless switch" on the user side, achieving a complete closed loop from prediction, provisioning, synchronization to switching, and meeting the stringent continuity requirements of scenarios such as uRLLC.

[0066] In step S6, after the service switchover is successfully completed in step S5, to ensure the effective utilization of edge computing resources and prevent resource silos, step S6 immediately initiates the reclamation of resources occupied by the old service instances on the source node. This step aims to achieve a closed loop in resource management and is executed by the dynamic resource reclamation unit. The specific process is as follows: S601: The recycling trigger subunit is responsible for determining when to initiate the recycling process. Specifically, it continuously monitors the service processing status of the newly activated container instance in step S5. Once it confirms that the instance has successfully processed 10 consecutive service requests from user devices, it sends a resource recycling trigger signal to the source edge node.

[0067] The criteria for determining whether a business request is "successfully processed" include: an application layer protocol response status code indicating success (e.g., HTTP status code 2xx), and passing data integrity checks (e.g., checksums or hashes) according to business agreements.

[0068] S602: After receiving the trigger signal at the source node, the resource release subunit performs a resource release operation. This operation consists of two steps: Terminate service process: Send a SIGTERM signal to the main process of the original service instance to notify it to terminate gracefully.

[0069] Release resource allocation: After a process terminates, by calling the operating system kernel's resource management interface (e.g., releasing memory page frames, removing from the process scheduling queue, closing network sockets, etc.), the process's occupation of CPU, memory, and network bandwidth is released, allowing these resources to return to the system's global available pool.

[0070] S603: The storage cleanup subunit is responsible for handling persistent data that stateful services may leave behind. Specifically, it marks the persistent storage volume used by the original service instance (usually based on copy-on-write technology, such as Docker's overlay2 driver) as "pending cleanup." After marking, the volume becomes read-only and no longer accepts new writes.

[0071] Subsequently, the system initiates a low-priority background garbage collection task, which will eventually complete the physical deletion of the storage volume within a marked 24-hour delay window. This delayed deletion strategy preserves a recovery window for possible abnormal rollbacks or data audits.

[0072] S604: After completing the above resource release and cleanup operations, the resource view update subunit reports resource changes to the MEC orchestrator. It calls the orchestrator's resource management API to update the global resource view stored therein in key-value pairs.

[0073] Specifically, the process involves using the identifier of the source edge node as the key to update its corresponding available resource vector value (e.g., increasing the number of available CPU cores, memory capacity, and bandwidth), thereby reintegrating the reclaimed resources into the cluster's available resource pool and providing an accurate basis for subsequent scheduling decisions.

[0074] Step S6, as the final step in the entire predictive migration process, employs a cautious and comprehensive resource reclamation mechanism. Reclamation is safely triggered through explicit success confirmation conditions (S601), which then systematically releases computing resources (S602), cleans up persistent data (S603), and synchronously updates the global resource state (S604). This not only solves the common problems of resource residue and waste in traditional migrations and enables elastic resource reuse, but more importantly, its delayed deletion and other strategies enhance fault tolerance for abnormal situations during the switchover process.

[0075] Step S6, in conjunction with the preceding steps, forms a complete resource scheduling closed loop from prediction, pre-setting, synchronization, switching, to recovery. This ensures seamless business continuity while also improving the resource utilization efficiency and reliability of the entire MEC system.

[0076] The aforementioned method constructs a high-precision spatiotemporal coupled trajectory prediction model by deeply fusing real-time signal strength gradients from the wireless side with historical user trajectory data, significantly improving the accuracy of target cell prediction. This model utilizes the first and second-order differences of signal strength to capture the trends in user movement direction and speed, combining this with the velocity and acceleration vectors of historical trajectories to form multi-dimensional dynamic feature inputs. The sequence-to-sequence prediction network effectively analyzes user movement patterns in complex urban environments or along high-speed railway lines through the long-term dependency modeling capabilities of gated recurrent units. The prediction results are output in the form of a probability distribution, providing a quantitative basis for the generation of target cell sequences.

[0077] During the resource preloading phase, this invention abandons the coarse-grained approach of migrating complete containers and instead adopts a lightweight container image preloading strategy. Only the operating system base layer and application dependency library layer are retained, significantly reducing the amount of data transferred by the image. Real-world testing shows that for typical industrial control applications, the size of a complete container image is approximately 500 megabytes, while the lightweight image is only 80 megabytes, reducing the transfer time from 800 milliseconds to 120 milliseconds. Simultaneously, the design of mounting read-only volumes prevents the preloaded container from performing any write operations before state synchronization, ensuring state consistency.

[0078] A streaming synchronization mechanism for service state is key to achieving seamless failover. By periodically scanning dirty pages every five milliseconds, the system can capture state changes with sub-ten-millisecond granularity. Differential encoding further compresses the amount of data transmitted; actual measurements show that under typical load, the average differential patch stream generated per second is only two hundred kilobytes. Low-latency optimization of the dedicated transport control protocol connection ensures real-time state synchronization, keeping the deviation between the state of the preloaded container and the source container within an acceptable range.

[0079] The service handover action is triggered the instant the base station handover is completed, and the entire activation process takes no more than three milliseconds. Container state switching and network interface binding are executed in parallel, and the client redirection response returns immediately after activation. After receiving the redirection response, the user equipment sends subsequent requests directly to the new container, achieving business continuity. The measured end-to-end service interruption time is an average of 4.2 milliseconds, which fully meets the uRLLC service requirement of latency below 10 milliseconds.

[0080] A dynamic resource reclamation mechanism effectively prevents waste of edge resources. The reclamation trigger condition is set to ensure that the source resources are released only after a successful switchover by successfully processing ten consecutive business requests. The delayed deletion strategy for persistent storage volumes accommodates the possibility of abnormal rollbacks, retaining data for 24 hours without consuming active computing resources. Real-time updates to the global resource view ensure that the MEC orchestrator accurately grasps the system resource status, providing a reliable basis for subsequent scheduling decisions.

[0081] In summary, this invention constructs a complete closed-loop scheduling process of "prediction-preloading-seamless handover-reclaiming" through the coordinated execution of six steps, S1 to S6. This process moves service migration actions forward to before base station handover, fundamentally solving the service interruption problem caused by traditional reactive strategies. Simultaneously, the combination of lightweight imagery and incremental state synchronization significantly reduces resource overhead while ensuring handover performance. The dynamic reclamation mechanism ensures system resource resilience in the event of prediction failure or user trajectory deviation from expectations, improving overall service reliability and resource utilization efficiency.

[0082] This invention also provides a 5G communication system multi-access edge computing resource scheduling system, which includes a multi-source data acquisition unit, a high-precision trajectory prediction unit, an edge resource preloading unit, a service status pre-synchronization unit, a seamless service handover unit, and a dynamic resource reclamation unit. The multi-source data acquisition unit is responsible for executing step S1, which collects in real-time the wireless signal strength sequence of user equipment, historical mobile trajectory data, and the status information of currently running service instances. The high-precision trajectory prediction unit is responsible for executing step S2, which constructs a spatiotemporally coupled trajectory prediction model based on the collected data and generates a high-confidence target cell sequence. The edge resource preloading unit is responsible for executing step S3, which determines candidate edge nodes based on the target cell sequence and preloads lightweight container images. The service status pre-synchronization unit is responsible for executing step S4, which streams incremental status data to the preloaded container. The seamless service handover unit is responsible for executing step S5, which immediately activates the new container after base station handover is completed. The dynamic resource reclamation unit is responsible for executing step S6, which releases the source node resources after successful handover.

[0083] The system is deployed on top of the 5G core network and MEC infrastructure, with each unit interacting with existing network functions through standard interfaces. The multi-source data acquisition unit communicates with the wireless resource management module, location management unit, and container runtime via service-oriented interfaces. The high-precision trajectory prediction unit runs as an embedded service within the MEC orchestrator, distributing its prediction results to the edge resource preloading unit via message queues. The edge resource preloading unit interfaces with the container management agent of the MEC host to perform image deployment and container creation. The service state pre-synchronization unit deploys agent processes on both the source and target MEC hosts, achieving state synchronization through point-to-point connections. The seamless service switching unit listens for event notifications from the session management unit and activates containers through the application lifecycle management interface of the MEC host. The dynamic resource reclamation unit works in conjunction with the resource manager to perform resource release and view updates.

[0084] The system employs a microservice architecture, characterized by high cohesion and low coupling. Inter-unit communication utilizes an asynchronous message passing mechanism, ensuring degraded operation even in the event of component failure. For example, if the high-precision trajectory prediction unit is temporarily unavailable, the system can degrade to a default prediction strategy based on nearest neighbor cells, reducing accuracy while maintaining basic functionality. Each unit is equipped with health checks and automatic recovery mechanisms, ensuring service continuity through periodic heartbeat checks and container restarts.

[0085] This embodiment details the specific implementation of the method and system of the present invention, covering the complete technical chain from data acquisition, trajectory prediction, resource preloading, state synchronization, service switching, and resource reclamation. All technical details are described in deterministic language without using any vague or selective terms, ensuring the clarity and implementability of the patent protection scope.

[0086] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0087] 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 method for scheduling multi-access edge computing resources in a 5G communication system, the method comprising: Comprise: S1, real-time acquisition of user equipment wireless signal strength sequence, historical moving track data and current running service instance state information; S2, based on the wireless signal strength sequence and historical moving track data, construct a spatio-temporal coupled trajectory prediction model, generate a high confidence target cell sequence in a future preset time window; S3, according to the target cell sequence, determine at least one candidate edge node, and preload a lightweight container image with the same structure as the current service instance on the candidate edge node; S4, the incremental state data of the current service instance is synchronized to the preloaded container of the candidate edge node in a streaming manner; S5, when the user equipment enters the target cell and completes the base station handover, immediately activate the preloaded container on the candidate edge node to take over the service request; S6, after the service handover is completed, start the resource recycling process on the source edge node to release the computing and storage resources occupied by the migrated service instance. 2.The method of claim 1, wherein, In S1 step: The acquisition of the wireless signal strength sequence specifically includes: obtaining the reference signal received power and reference signal received quality values reported by the user equipment in continuous multiple measurement periods, and the time interval of each measurement period is 20 milliseconds; The acquisition of the historical moving track data specifically includes: obtaining the latitude and longitude coordinate point sequence recorded by the user equipment every 10 seconds in the past one hour; The acquisition of the service instance state information specifically includes: obtaining the memory snapshot, network connection table, session context and application layer state variable set of the current service instance through the container runtime interface. 3.The method of claim 2, wherein, The construction of the spatio-temporal coupled trajectory prediction model in S2 step includes: The signal strength sequence is processed by sliding window, and the first order difference and the second order difference of the signal strength value in each window are calculated to form a signal strength change gradient feature vector; The historical moving track data is converted to a local Cartesian coordinate system with the current service base station as the origin, and the velocity vector and acceleration vector are calculated; The signal strength change gradient feature vector, velocity vector and acceleration vector are time-aligned and feature-spliced to form a fusion feature vector sequence, which is input into a sequence-to-sequence prediction network for training and prediction. 4.The method of claim 3, wherein, Generating a high confidence target cell sequence includes: The sequence-to-sequence prediction network outputs a probability matrix of user equipment accessing each adjacent cell in the future multiple measurement periods; The probability matrix is accumulated and summed along the time dimension to obtain the cumulative access probability of each cell in the prediction window; Select the cells with cumulative probability exceeding the preset threshold, arrange them in descending order of cumulative probability value to form the target cell sequence. 5.The method of claim 1, wherein, The preloading of lightweight container image in S3 step includes: Pull the standardized container image corresponding to the current service instance from the image repository, and the image adopts a layered structure; Strip the standardized container image layer by layer, only keep the operating system base layer and application dependency library layer, remove the application code layer and configuration file layer, and generate a lightweight base image; Deploy the lightweight base image to the candidate edge node and mount a read-only storage volume for the preloaded container. 6.The method of claim 1, wherein, In S4 step, the incremental state data is synchronized in a streaming manner, which includes: Scanning the memory page table of the service instance on the source edge node at a fixed period, identifying the dirty pages modified since the last period; Differential encoding the identified dirty pages, generating a binary patch stream containing offset, length and new content for the changed data block only; Pushing the binary patch stream to the candidate node through a dedicated Transmission Control Protocol connection between the source edge node and the candidate edge node; On the candidate edge node, writing the received differential patch stream into the pre-allocated memory mapping area consistent with the source container virtual address layout in the preloaded container. 7.The method of claim 6, wherein, The dedicated Transmission Control Protocol connection adopts low-delay optimization parameters, including disabling the Nagle algorithm and setting the minimum retransmission timeout to 10 milliseconds. 8.The method of claim 1, wherein, The S5 step includes: Listening to the service activation instruction issued by the core network session management function unit after the user equipment completes the base station switching; According to the service activation instruction, binding the network interface of the preloaded container to the service IP address allocated to the user equipment on the corresponding candidate edge node, and switching the container state to the running state; Returning the service redirection response containing the new container instance service endpoint address to the user equipment. 9.The method of claim 1, wherein, The trigger condition for starting the resource recycling process in the S6 step is that the newly activated container instance has successfully processed at least 10 service requests in succession; The resource recycling process includes: terminating the original service instance process on the source edge node, releasing the computing and network resources occupied by it, marking the persistent storage volume used by the original service instance as to be cleaned up and physically deleting it after a delay time, and updating the global resource view in the MEC orchestrator. 10.A system for scheduling of multi-access edge computing resources in a 5G communication system, the system comprising: It includes: A multi-source data acquisition unit for real-time acquisition of user equipment wireless signal strength sequence, historical mobile trajectory data and current service instance state information; A high-precision trajectory prediction unit for constructing a spatio-temporal coupled trajectory prediction model based on the wireless signal strength sequence and historical mobile trajectory data, and generating a high-confidence target cell sequence within a future preset time window; An edge resource preloading unit for determining at least one candidate edge node based on the target cell sequence, and preloading a lightweight container image isomorphic to the current service instance on the candidate edge node; A service state pre-synchronization unit for synchronizing the incremental state data of the current service instance to the preloaded container on the candidate edge node in a streaming manner; A seamless service switching unit for activating the preloaded container on the candidate edge node to take over service requests immediately after the user equipment enters the target cell and completes the base station switching; A dynamic resource recycling unit for starting a resource recycling process on the source edge node to release the resources occupied by the migrated service instance after service switching is completed.

Images