Mobile edge computing task unloading method based on priority ranking
By adopting a priority-based mobile edge computing task offloading method, the problem of low resource utilization in existing systems is solved, and efficient task processing and resource utilization are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-29
- Publication Date
- 2026-04-07
AI Technical Summary
Existing mobile edge computing systems lack accurate resource prediction and dynamic priority linkage mechanisms, resulting in delays and resource waste of high-priority tasks and low resource utilization.
A priority-based mobile edge computing task offloading method is adopted. Through task feature collection and classification, multi-dimensional dynamic ranking of task priorities, real-time perception and prediction of edge node resources, hierarchical offloading decision and resource fragment filling scheduling, it ensures that high-priority tasks are processed first and edge node resources are utilized efficiently.
This effectively avoids resource overload, improves the stable execution of high-priority tasks, and enhances the resource utilization efficiency and task processing throughput of the edge computing system.
Smart Images
Figure CN121815337A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of edge computing technology, specifically a method for offloading mobile edge computing tasks based on priority sorting. Background Technology
[0002] Mobile edge computing (MEC) is a new computing paradigm that moves cloud computing capabilities and IT service environments from the network core to the edge of the mobile network, closer to the data source or user. It provides localized computing, data processing, and storage services for terminal applications by deploying servers, storage, and network resources at edge locations such as base stations and aggregation nodes. This architecture significantly reduces data transmission latency and network backhaul bandwidth pressure, meeting the stringent requirements of applications such as autonomous driving, industrial IoT, and AR / VR, which demand real-time performance, high bandwidth, and data security. Mobile edge computing task offloading is a key technology in MEC architecture. It refers to the process by which resource-constrained terminal devices (such as smartphones and IoT sensors) migrate some or all of their computationally intensive or latency-sensitive tasks to nearby edge servers via wireless networks. By processing computing tasks closer to the data source, MEC task offloading significantly reduces end-to-end response time and alleviates pressure on the core network. It is a core mechanism for achieving low-latency applications, extending terminal battery life, and improving user experience, providing strong technical support for scenarios such as connected vehicles, industrial automation, and augmented reality.
[0003] However, existing technologies often lack accurate prediction of edge node resources and dynamic priority linkage mechanisms, and often cause delays in high-priority tasks due to passive adjustments after resource saturation. At the same time, there is a lack of efficient utilization solutions for idle resource fragments on edge nodes, which either leads to resource waste or makes it difficult to handle low-priority small tasks reasonably, resulting in low overall resource utilization and task processing efficiency. Summary of the Invention
[0004] The purpose of this invention is to provide a priority-based mobile edge computing task offloading method to solve the problems mentioned above.
[0005] The technical solution adopted in this invention is as follows: a mobile edge computing task offloading method based on priority sorting, characterized in that: the method includes the following steps: S1: Task Feature Collection and Classification: First, collect the core features of user-submitted tasks, including computational complexity (such as the number of floating-point operations), latency threshold (such as whether it is less than 200ms), data transmission volume (such as the size of input data), and reliability requirements (such as whether data loss is tolerated). After classifying these features, synchronize them to the S2 priority evaluation module and the S3 edge resource awareness module to provide a basis for subsequent priority ranking and resource matching.
[0006] S2: Multi-dimensional dynamic task priority ranking: Based on the task characteristics provided by S1, a weighted comprehensive scoring method is used to calculate the priority score (latency sensitivity has the highest weight, followed by computational load and reliability); after generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module; at the same time, it receives abnormal feedback from the S6 monitoring module. If the execution of a high-priority task is blocked, its priority coefficient is temporarily increased to ensure that critical tasks are processed first.
[0007] S3: Real-time perception and prediction of edge node resources: Periodically collect the real-time resource status (CPU utilization, memory remaining, network bandwidth, storage capacity) of all edge nodes, and predict the resource change trend in the near future through historical data; synchronize the perception results to the S4 offload decision module. If it is predicted that the resources of a certain node are about to be saturated, notify S2 in advance to adjust the priority of the tasks to be processed on that node to avoid resource overload.
[0008] S4: Layered Unloading Decision Generation: Combining the priority list of S2 and the resource status of S3, a three-layer unloading strategy is formulated: high-priority and latency-sensitive tasks (such as AR interaction) are given priority to local or nearby edge nodes; medium-priority tasks (such as video transcoding) are selected for unloading based on the resource status of edge nodes; low-priority tasks (such as data backup) can be unloaded to the remote cloud; the decision results are passed to the S5 scheduling module for execution, and the execution anomaly feedback from S6 is received for re-decision.
[0009] S5: Priority-driven resource fragment filling scheduling: First, according to the priority order of S2, high-priority tasks are allocated to edge nodes with sufficient resources to ensure that they exclusively occupy critical resources; for low-priority small tasks (such as sensor data preprocessing), a "resource fragment pool" mechanism is designed—collecting idle resource fragments (such as CPU idle time slots, small blocks of remaining memory) from edge nodes, packaging multiple low-priority small tasks into combined tasks that meet the size of the fragment resources, and filling them for execution without affecting high-priority tasks; this step requires real-time access to the resource status data of S3. If fragment resources are insufficient, low-priority tasks are temporarily stored locally for waiting, or feedback is sent to S4 to adjust the offloading strategy.
[0010] S6: Task execution monitoring and anomaly response: Track the execution progress, resource consumption and latency of all tasks on local and edge nodes in real time; if a task's latency exceeds the threshold set by S1 (e.g., real-time video task latency exceeds 150ms), immediately push the anomaly information to S2 to adjust the priority and trigger S4 to regenerate the uninstallation decision for uninstallation; at the same time, synchronize the monitoring data to the feedback module of S7 for subsequent optimization.
[0011] S7: Result Integration and Closed-Loop Optimization: Integrate the task results completed by local and edge nodes, and return the high-priority task results to the user in the priority order of S2; summarize the key data in the task execution process (such as unloading success rate, resource utilization rate, and delay compliance rate) into the feature library of S1 and the priority model of S2, continuously optimize the feature classification and priority evaluation logic of subsequent tasks, and form a self-iterative task unloading closed loop.
[0012] In a preferred embodiment, in step S1, the core characteristics of the user task, such as the computational complexity (number of floating-point operations), the maximum allowable delay threshold, the data transmission volume (number of input data bytes), and the reliability requirements (packet loss rate tolerance limit), are obtained in real time through the task submission interface. The acquisition process is triggered immediately upon task submission to ensure the real-time nature of the data. Then, the task is divided into different categories based on these characteristics. The classification operation is completed within 10ms, and the classification results are synchronized to the S2 priority evaluation module and the S3 edge resource awareness module to provide an accurate basis for subsequent priority ranking and resource matching.
[0013] In a preferred embodiment, in step S2, a weighted comprehensive scoring method is used to calculate the priority score based on the features provided in S1. The weight of the latency sensitivity dimension is set to 0.5, the weight of the computational complexity dimension is set to 0.3, and the weight of the reliability requirement dimension is set to 0.2. When calculating the score, the lower the latency threshold, the higher the score; the higher the computational complexity, the higher the score; and the higher the reliability requirement, the higher the score. After generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module. At the same time, it receives abnormal feedback from the S6 monitoring module. If the execution of a high-priority task is blocked, its priority coefficient is temporarily increased by 0.1. The time interval between the adjustment and re-sorting does not exceed 50ms to ensure that critical tasks are processed first.
[0014] In a preferred embodiment, in step S3, an exponential smoothing-Markov chain hybrid prediction model is used to acquire the real-time resource status (including CPU utilization, memory remaining, network uplink bandwidth, and local storage available capacity) of all edge nodes through a lightweight acquisition agent at a fixed period of 10 seconds. The acquired data of each node is first processed by a 5-period sliding window noise reduction to filter instantaneous fluctuations. Then, an exponential smoothing algorithm is used to extract the long-term trend of resource changes, and a Markov chain model is used to calculate the state transition probability of the residual of the trend term. The data are then fused to obtain the predicted resource status value for the next 30 seconds. The real-time acquired raw data and the prediction results are packaged into a multi-dimensional resource status matrix and synchronized to the S4 offload decision module through a low-latency message queue. If the predicted CPU utilization of a node exceeds 90% or the memory remaining is less than 10% of the total capacity, a saturation warning mechanism is triggered, and a signal is sent to the S2 task priority adjustment module, suggesting that the priority score of the low-priority tasks to be processed on that node be reduced (e.g., from 3.5 to 2.0), or the priority of the migrated tasks be increased so as to be unloaded in advance to avoid resource overload leading to task execution delays or failures. The calculation formula for determining the risk level of impending saturation of quantifiable node resources, which serves as the core basis for triggering early warnings, is as follows: ; In the formula: This represents the resource saturation risk index of node j (range 0~1, warning is triggered when ≥0.9). This represents the predicted resource value (e.g., CPU utilization percentage) for node j at a future time Δt. This represents the maximum resource threshold for node j (e.g., CPU utilization of 100%). γ represents the weighting coefficient of the predicted value (with a value of 0.7 to balance the influence of the predicted trend and real-time changes). This indicates the change in resources of node j between the current period and the previous period (e.g., a 5% increase in CPU utilization). Δt represents the duration of the data acquisition cycle (fixed at 10 seconds); This formula creatively combines future forecasts with real-time change rates, enabling it to more accurately capture early signs of resource saturation and avoid the lag in predictions based on a single indicator.
[0015] In a preferred embodiment, in step S4, a three-tiered offloading strategy is formulated by combining the priority list in S2 and the resource status in S3: high-priority tasks with a latency threshold of less than 200ms are preferentially selected for execution on edge nodes within 100 meters of the user device or locally; medium-priority tasks with a latency threshold between 200ms and 500ms are selected for offloading on edge nodes with CPU utilization below 70% and more than 2GB of remaining memory; low-priority tasks with a latency threshold greater than 500ms are directly offloaded to the remote cloud; the decision result is transmitted to the scheduling module in S5 for execution, and the execution anomaly feedback in S6 is received. If the task execution is abnormal, the time for re-decision is no more than 100ms to ensure that the task's offloading path is adjusted in a timely manner.
[0016] In a preferred embodiment, step S5 employs a priority-weighted greedy allocation algorithm for high-priority task allocation. First, the task priority ranking sequence (arranged in descending order of priority score) output by S2 is read, and tasks are processed sequentially. The edge node resource status data output by S3 is retrieved. The CPU, memory, and bandwidth resource requirements for the current task are defined. All edge nodes are traversed, and the resource satisfaction level of each node for the current task is calculated, selecting a set of nodes that fully meet the task requirements. Within this set, a resource redundancy calculation formula is used to select the node with the fewest remaining resources after task allocation, and resources are allocated to the task with a resource exclusivity flag set. If no node can meet the task requirements, a strategy adjustment signal is sent to S4, and the task is temporarily stored in the local execution queue. A specially designed resource redundancy formula will be used subsequently to quantify the core criteria for node selection.
[0017] In a preferred embodiment, step S5, the low-priority task fragment filling stage, employs a multi-dimensional fragment knapsack dynamic programming algorithm. First, the resource fragment set of each edge node is obtained from S3 (each fragment corresponds to idle CPU, memory, and bandwidth resources). The resource requirements and priority-based utility value are defined for each task in the low-priority task set. The optimal task combination is solved using dynamic programming, which defines a state representing the maximum utility of filling the first few tasks with the first few fragments, and uses state transition logic to determine whether to add the current task to the combination (ensuring that the task requirement does not exceed the resource amount of the target fragment). After generating the optimal combined task set, it is allocated to the corresponding resource fragments for execution. The resource state data of S3 is synchronized in real time. If a high-priority task on a node needs resource expansion, the combined task on that node is immediately terminated, fragment resources are released, and the fragment set is updated. The core logic of this algorithm is subsequently supported by a fragment resource availability formula and a combined task optimal utility formula.
[0018] In a preferred embodiment, the resource redundancy calculation formula for selecting high-priority task nodes in step S5 is as follows: ; In the formula: This indicates the resource redundancy after node j is allocated to task t_i (the smaller the value, the more fully the resources are utilized). This represents the remaining amount of resource r in node j; Represents task t i The demand for resources r; Fragment resource availability formula (quantifying the effective utilization value of fragments); ; In the formula: A f This indicates the availability of fragment f (0~1, the higher the value, the more suitable it is for filling tasks). w c ,w m ,w b Represents the resource weights of CPU, memory, and bandwidth; f c ,f m ,f b This represents the CPU idle percentage, memory idle size, and bandwidth idle rate (normalized values) of fragment f. β f This represents the fragment stability coefficient (0~1, reflecting the probability that the fragment remains usable). The optimal utility formula for a combined task using the objective function of dynamic programming is: ; Its constraints are:
[0019] ; In the formula: U opt This represents the maximum utility of the combined tasks that fragment f can carry; C represents a subset of low-priority tasks; T low This represents a set of low-priority tasks. d k [r] represents task t k The demand for resources r; F f [r] represents the resource capacity r of fragment f; u k Represents task t k The normalized utility value; d k[r] represents the resource type r required by task tk; F f [r] represents the available capacity of resource type r in resource fragment f.
[0020] In a preferred embodiment, in step S6, the monitoring frequency is to collect data once every 10ms; if it is found that the current delay of the task exceeds the set threshold of 10%, the abnormal information is immediately pushed to S2 to adjust the priority and trigger S4 to regenerate the uninstallation decision, wherein the time for S2 to adjust the priority does not exceed 20ms, the time for S4 to re-determine the decision does not exceed 30ms, and the monitoring data is synchronized to the feedback module of S7 to provide data support for subsequent optimization.
[0021] In a preferred embodiment, in step S7, key data such as unloading success rate, resource utilization rate, and latency compliance rate during task execution are summarized. The unloading success rate is statistically analyzed once per hour. The resource utilization rate is the average utilization rate of CPU, memory, and bandwidth of each edge node. The latency compliance rate is the proportion of each task whose actual latency is within a set threshold. The feature library of S1 and the priority model of S2 are updated once every 1,000 task data are accumulated, with an update time of no more than 5 minutes. The feature classification and priority evaluation logic of subsequent tasks are continuously optimized to form a self-iterative task unloading closed loop.
[0022] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, by collecting the resource status of edge nodes in real time and combining it with a hybrid prediction model, the system accurately captures resource change trends and identifies the risk of node resource saturation in advance. When it is predicted that node resources are about to become insufficient, the system will adjust the priority of pending tasks in a timely manner, either by lowering the priority of low-priority tasks to release resources or by increasing the priority of migrated tasks to unload them in advance, effectively avoiding task blocking or delays caused by resource overload and ensuring the stable execution of high-priority tasks.
[0023] 2. In this invention, the idle resource fragments of edge nodes are efficiently utilized through a resource fragmentation pool mechanism and a dynamic programming algorithm. The system combines low-priority small tasks into task packages adapted to fragmented resources, and fills them for execution without interfering with high-priority tasks. This fully taps into the remaining resource potential of edge nodes and allows low-priority tasks to be handled reasonably, thereby improving the overall resource utilization efficiency and task processing throughput of the edge computing system. Attached Figure Description
[0024] Figure 1 This is a schematic diagram illustrating the process principle of the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0026] Example: Refer to Figure 1 A method for offloading mobile edge computing tasks based on priority sorting, comprising the following steps: S1: Task Feature Collection and Classification: First, collect the core features of user-submitted tasks, including computational complexity (such as the number of floating-point operations), latency threshold (such as whether it is less than 200ms), data transmission volume (such as the size of input data), and reliability requirements (such as whether data loss is tolerated). After classifying these features, synchronize them to the S2 priority evaluation module and the S3 edge resource awareness module to provide a basis for subsequent priority ranking and resource matching.
[0027] S2: Multi-dimensional dynamic task priority ranking: Based on the task characteristics provided by S1, a weighted comprehensive scoring method is used to calculate the priority score (latency sensitivity has the highest weight, followed by computational load and reliability); after generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module; at the same time, it receives abnormal feedback from the S6 monitoring module. If the execution of a high-priority task is blocked, its priority coefficient is temporarily increased to ensure that critical tasks are processed first.
[0028] S3: Real-time perception and prediction of edge node resources: Periodically collect the real-time resource status (CPU utilization, memory remaining, network bandwidth, storage capacity) of all edge nodes, and predict the resource change trend in the near future through historical data; synchronize the perception results to the S4 offload decision module. If it is predicted that the resources of a certain node are about to be saturated, notify S2 in advance to adjust the priority of the tasks to be processed on that node to avoid resource overload.
[0029] S4: Layered Unloading Decision Generation: Combining the priority list of S2 and the resource status of S3, a three-layer unloading strategy is formulated: high-priority and latency-sensitive tasks (such as AR interaction) are given priority to local or nearby edge nodes; medium-priority tasks (such as video transcoding) are selected for unloading based on the resource status of edge nodes; low-priority tasks (such as data backup) can be unloaded to the remote cloud; the decision results are passed to the S5 scheduling module for execution, and the execution anomaly feedback from S6 is received for re-decision.
[0030] S5: Priority-driven resource fragment filling scheduling: First, according to the priority order of S2, high-priority tasks are allocated to edge nodes with sufficient resources to ensure that they exclusively occupy critical resources; for low-priority small tasks (such as sensor data preprocessing), a "resource fragment pool" mechanism is designed—collecting idle resource fragments (such as CPU idle time slots, small blocks of remaining memory) from edge nodes, packaging multiple low-priority small tasks into combined tasks that meet the size of the fragment resources, and filling them for execution without affecting high-priority tasks; this step requires real-time access to the resource status data of S3. If fragment resources are insufficient, low-priority tasks are temporarily stored locally for waiting, or feedback is sent to S4 to adjust the offloading strategy.
[0031] S6: Task execution monitoring and anomaly response: Track the execution progress, resource consumption and latency of all tasks on local and edge nodes in real time; if a task's latency exceeds the threshold set by S1 (e.g., real-time video task latency exceeds 150ms), immediately push the anomaly information to S2 to adjust the priority and trigger S4 to regenerate the uninstallation decision for uninstallation; at the same time, synchronize the monitoring data to the feedback module of S7 for subsequent optimization.
[0032] S7: Result Integration and Closed-Loop Optimization: Integrate the task results completed by local and edge nodes, and return the high-priority task results to the user in the priority order of S2; summarize the key data in the task execution process (such as unloading success rate, resource utilization rate, and delay compliance rate) into the feature library of S1 and the priority model of S2, continuously optimize the feature classification and priority evaluation logic of subsequent tasks, and form a self-iterative task unloading closed loop.
[0033] In step S1, the core features of the user task are obtained in real time through the task submission interface, such as the computational complexity (number of floating-point operations), the maximum allowable delay threshold, the data transmission volume (number of input data bytes), and the reliability requirements (upper limit of packet loss rate tolerance). The acquisition process is triggered immediately upon task submission to ensure the real-time nature of the data. Then, based on these features, the task is divided into different categories. The classification operation is completed within 10ms, and the classification results are synchronized to the S2 priority evaluation module and the S3 edge resource awareness module, providing an accurate basis for subsequent priority ranking and resource matching.
[0034] In step S2, a weighted comprehensive scoring method is used to calculate the priority score based on the features provided in S1. The weight of the latency sensitivity dimension is set to 0.5, the weight of the computational complexity dimension is set to 0.3, and the weight of the reliability requirement dimension is set to 0.2. When calculating the score, the lower the latency threshold, the higher the score; the higher the computational complexity, the higher the score; and the higher the reliability requirement, the higher the score. After generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module. At the same time, it receives the abnormal feedback from the S6 monitoring module. If the execution of a high-priority task is blocked, its priority coefficient is temporarily increased by 0.1. The time interval between the adjustment and re-sorting does not exceed 50ms to ensure that critical tasks are processed first.
[0035] In step S3, an exponential smoothing-Markov chain hybrid prediction model is used. A lightweight acquisition agent acquires the real-time resource status (including CPU utilization, remaining memory, network uplink bandwidth, and available local storage capacity) of all edge nodes at a fixed 10-second interval. The acquired data for each node first undergoes a 5-period sliding window denoising process to filter out instantaneous fluctuations. Then, an exponential smoothing algorithm is used to extract the long-term trend of resource changes. A Markov chain model is then used to calculate the state transition probability of the residuals of the trend, and the data is fused to obtain the predicted resource status for the next 30 seconds. The real-time acquired raw data and the prediction results are packaged into a multi-dimensional resource status matrix and synchronized to the S4 offload decision module via a low-latency message queue. If the predicted CPU utilization of a node exceeds 90% or the remaining memory is less than 10% of the total capacity, a saturation warning mechanism is triggered. A signal is sent to the S2 task priority adjustment module, suggesting that the priority score of the low-priority tasks pending processing on that node be lowered (e.g., from 3.5 to 2.0), or that the priority of migrated tasks be increased to facilitate early offload and avoid resource overload leading to task execution delays or failures. The calculation formula for determining the risk level of impending saturation of quantifiable node resources, which serves as the core basis for triggering early warnings, is as follows: ; In the formula: This represents the resource saturation risk index of node j (range 0~1, warning is triggered when ≥0.9). This represents the predicted resource value (e.g., CPU utilization percentage) for node j at a future time Δt. This represents the maximum resource threshold for node j (e.g., CPU utilization of 100%). γ represents the weighting coefficient of the predicted value (with a value of 0.7 to balance the influence of the predicted trend and real-time changes). This indicates the change in resources at node j between the current period and the previous period (e.g., a 5% increase in CPU utilization). Δt represents the duration of the data acquisition cycle (fixed at 10 seconds). This formula creatively combines future forecasts with real-time change rates, enabling it to more accurately capture early signs of resource saturation and avoid the lag in predictions based on a single indicator.
[0036] In step S4, a three-tiered offloading strategy is formulated based on the priority list in S2 and the resource status in S3: high-priority tasks with a latency threshold below 200ms are preferentially selected for execution on edge nodes within 100 meters of the user device or locally; medium-priority tasks with a latency threshold between 200ms and 500ms are selected for offloading on edge nodes with CPU utilization below 70% and more than 2GB of remaining memory; low-priority tasks with a latency threshold greater than 500ms are directly offloaded to the remote cloud; the decision result is transmitted to the S5 scheduling module for execution, and the execution anomaly feedback is received from S6. If a task execution anomaly occurs, the time for re-decision is no more than 100ms to ensure that the task's offloading path is adjusted in a timely manner.
[0037] In step S5, the high-priority task allocation stage employs a priority-weighted greedy allocation algorithm. First, the task priority ranking sequence output from S2 (arranged in descending order of priority score) is read, and tasks are processed one by one in this order. The edge node resource status data output from S3 is retrieved. The CPU, memory, and bandwidth resource requirements for the current task are determined. All edge nodes are traversed, and the resource satisfaction level of each node for the current task is calculated, filtering out a set of nodes that can fully meet the task's requirements. Within this set, the node with the fewest remaining resources after task allocation is selected using a resource redundancy calculation formula. Resources are allocated to the task, and a resource exclusivity flag is set. If no node can meet the task's requirements, a strategy adjustment signal is sent to S4, and the task is temporarily stored in the local execution queue. A specially designed resource redundancy formula will be used later to quantify the core criteria for node selection.
[0038] In step S5, the low-priority task fragment filling stage employs a multi-dimensional fragment knapsack dynamic programming algorithm. First, the resource fragment set of each edge node is obtained from S3 (each fragment corresponds to idle CPU, memory, and bandwidth resources). The resource requirements and priority-based utility value of each task in the low-priority task set are defined. The optimal task combination is solved using dynamic programming, which defines a state representing the maximum utility of filling the first few tasks with the first few fragments, and uses state transition logic to determine whether to add the current task to the combination (ensuring that the task requirement does not exceed the resource amount of the target fragment). After generating the optimal combined task set, it is allocated to the corresponding resource fragments for execution. The resource state data of S3 is synchronized in real time. If a high-priority task on a node needs resource expansion, the combined task on that node is immediately terminated, fragment resources are released, and the fragment set is updated. The core logic of this algorithm will be supported by the fragment resource availability formula and the optimal utility formula for combined tasks.
[0039] In step S5, the formula for calculating resource redundancy used to select high-priority task nodes is: ; In the formula: This indicates the resource redundancy after node j is allocated to task t_i (the smaller the value, the more fully the resources are utilized). This represents the remaining amount of resource r in node j; Represents task t i The demand for resources r; Fragment resource availability formula (quantifying the effective utilization value of fragments) ; In the formula: A f This indicates the availability of fragment f (0~1, the higher the value, the more suitable it is for filling tasks). w c ,w m ,w b Represents the resource weights of CPU, memory, and bandwidth; f c ,f m ,f b This represents the CPU idle percentage, memory idle size, and bandwidth idle rate (normalized values) of fragment f. β f This represents the fragment stability coefficient (0~1, reflecting the probability that the fragment remains usable). The optimal utility formula for a combined task using the objective function of dynamic programming is: ; Its constraints are:
[0040] ; In the formula: U opt This represents the maximum utility of the combined tasks that fragment f can carry; C represents a subset of low-priority tasks; T low This represents a set of low-priority tasks. d k [r] represents task t k The demand for resources r; F f [r] represents the resource capacity r of fragment f; u kRepresents task t k The normalized utility value; d k [r] represents the resource type r required by task tk; F f [r] represents the available capacity of resource type r in resource fragment f.
[0041] In step S6, the monitoring frequency is to collect data once every 10ms. If the current delay of the task is found to exceed the set threshold of 10%, the abnormal information is immediately pushed to S2 to adjust the priority and trigger S4 to regenerate the uninstallation decision. The time for S2 to adjust the priority does not exceed 20ms, and the time for S4 to re-determine the decision does not exceed 30ms. At the same time, the monitoring data is synchronized to the feedback module of S7 to provide data support for subsequent optimization.
[0042] In step S7, key data such as unloading success rate, resource utilization rate, and latency compliance rate during task execution are summarized. The unloading success rate is statistically analyzed once per hour. The resource utilization rate is the average utilization rate of CPU, memory, and bandwidth of each edge node. The latency compliance rate is the proportion of each task whose actual latency is within the set threshold. The feature library of S1 and the priority model of S2 are updated once every 1,000 task data are accumulated, with an update time of no more than 5 minutes. The feature classification and priority evaluation logic of subsequent tasks are continuously optimized to form a self-iterative task unloading closed loop.
[0043] From the above, we can conclude that: In this invention, by collecting the resource status of edge nodes in real time and combining it with a hybrid prediction model, the system accurately captures resource change trends and identifies the risk of node resource saturation in advance. When it is predicted that node resources are about to become insufficient, the system will promptly adjust the priority of pending tasks, either by lowering the priority of low-priority tasks to release resources or by increasing the priority of migrated tasks to unload them in advance, effectively avoiding task blocking or delays caused by resource overload and ensuring the stable execution of high-priority tasks.
[0044] This invention utilizes a resource fragmentation pool mechanism and a dynamic programming algorithm to efficiently utilize idle resource fragments on edge nodes. The system combines low-priority small tasks into task packages adapted to fragmented resources, filling and executing them without interfering with high-priority tasks. This fully taps into the remaining resource potential of edge nodes while ensuring that low-priority tasks are handled appropriately, thereby improving the overall resource utilization efficiency and task processing throughput of the edge computing system.
[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0046] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for offloading mobile edge computing tasks based on priority sorting, characterized in that: The method includes the following steps: S1: Task Feature Collection and Classification: First, collect the core features of user-submitted tasks, including computational complexity, latency threshold, data transmission volume, and reliability requirements; S2: Multi-dimensional dynamic task priority ranking: Based on the task features provided by S1, a weighted comprehensive scoring method is used to calculate the priority score; after generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module; at the same time, it receives abnormal feedback from the S6 monitoring module. If the execution of a high-priority task is blocked, its priority coefficient is temporarily increased to ensure that critical tasks are handled first. S3: Real-time perception and prediction of edge node resources: Periodically collect the real-time resource status of all edge nodes and predict the resource change trend in the near future through historical data; synchronize the perception results to the S4 unloading decision module; if it is predicted that the resources of a certain node are about to be saturated, notify S2 in advance to adjust the priority of the tasks to be processed on that node. S4: Layered offloading decision generation: Combining the priority list of S2 and the resource status of S3, a three-layer offloading strategy is formulated: high-priority and latency-sensitive tasks are given priority to local or nearby edge nodes; medium-priority tasks are selected for offloading based on the resource status of edge nodes; low-priority tasks can be offloaded to the remote cloud; the decision results are passed to the S5 scheduling module for execution, and the execution anomaly feedback from S6 is received for re-decision. S5: Priority-driven resource fragment filling scheduling: First, according to the priority order of S2, high-priority tasks are allocated to edge nodes with sufficient resources to ensure that they exclusively occupy critical resources; for low-priority small tasks, a "resource fragment pool" mechanism is designed - collect idle resource fragments from edge nodes, package multiple low-priority small tasks into combined tasks that meet the size of fragment resources, and fill the execution without affecting high-priority tasks. S6: Task execution monitoring and anomaly response: Real-time tracking of the execution progress, resource consumption, and latency of all tasks on local and edge nodes; if a task's latency exceeds the threshold set by S1, the anomaly information is immediately pushed to S2 to adjust its priority, and S4 is triggered to regenerate the uninstallation decision for uninstallation; at the same time, the monitoring data is synchronized to the feedback module of S7 for subsequent optimization. S7: Result Integration and Closed-Loop Optimization: Integrate the task results completed by local and edge nodes, and return the high-priority task results to the user in the priority order of S2; summarize the key data in the task execution process into the feature library of S1 and the priority model of S2, and continuously optimize the feature classification and priority evaluation logic of subsequent tasks.
2. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S1, the computational complexity of the user task, the number of floating-point operations, the maximum allowable value of the delay threshold, the number of bytes of input data for data transmission, and the upper limit of the packet loss rate tolerance for reliability requirements are obtained in real time through the task submission interface.
3. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S2, a weighted comprehensive scoring method is used to calculate the priority score based on the features provided in S1. The weight of the delay sensitivity dimension is set to 0.5, the weight of the computational complexity dimension is set to 0.3, and the weight of the reliability requirement dimension is set to 0.
2. When calculating the score, the lower the delay threshold, the higher the score; the higher the computational complexity, the higher the score; and the higher the reliability requirement, the higher the score. After generating the priority ranking list, it is passed to the S4 unloading decision module and the S5 scheduling module, and at the same time, the abnormal feedback from the S6 monitoring module is received.
4. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S3, an exponential smoothing-Markov chain hybrid prediction model is used to acquire the real-time resource status of all edge nodes through a lightweight acquisition agent at a fixed period of 10 seconds. The acquired data of each node is first processed by a 5-period sliding window noise reduction to filter instantaneous fluctuations. Then, an exponential smoothing algorithm is used to extract the long-term trend term of resource changes, and a Markov chain model is used to calculate the state transition probability of the residual of the trend term. The data are then fused to obtain the predicted value of the resource status in the next 30 seconds. The real-time acquired raw data and the prediction results are packaged into a multi-dimensional resource status matrix and synchronized to the S4 offloading decision module through a low-latency message queue. The calculation formula for determining the risk level of impending saturation of quantifiable node resources, which serves as the core basis for triggering early warnings, is as follows: ; In the formula: This represents the resource saturation risk index of node j; This represents the predicted resource value for node j at a future time Δt. This represents the maximum resource threshold for node j; γ represents the weighting coefficient of the predicted value; This indicates the change in resources of node j between the current period and the previous period; Δt represents the duration of the data acquisition cycle.
5. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S4, a three-tiered offloading strategy is formulated by combining the priority list in S2 and the resource status in S3: high-priority tasks with a latency threshold of less than 200ms are preferentially selected for execution on edge nodes within 100 meters of the user device or locally; medium-priority tasks with a latency threshold between 200ms and 500ms are selected for offloading on edge nodes with CPU utilization of less than 70% and more than 2GB of remaining memory; low-priority tasks with a latency threshold greater than 500ms are directly offloaded to the remote cloud.
6. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S5, the high-priority task allocation process adopts a priority-weighted greedy allocation algorithm. First, the task priority sorting sequence output by S2 is read, and tasks are processed one by one in this order. The edge node resource status data output by S3 is called. The resource requirements of CPU, memory, and bandwidth for the current task are defined. All edge nodes are traversed, the resource satisfaction of each node for the current task is calculated, and a set of nodes that can fully meet the task requirements is selected.
7. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S5, the low-priority task fragment filling step adopts a multi-dimensional fragment knapsack dynamic programming algorithm. First, the resource fragment set of each edge node is obtained from S3. The resource requirements and priority-based utility value of each task in the low-priority task set are defined. The optimal task combination is solved by dynamic programming. This method defines a state to represent the maximum utility of using the first few fragments to fill the first few tasks, and determines whether to add the current task to the combination through state transition logic.
8. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S5, the formula for calculating resource redundancy used to select high-priority task nodes is: ; In the formula: This indicates the resource redundancy after node j is assigned to task t_i; This represents the remaining amount of resource r in node j; Represents task t i The demand for resources r; Fragment resource availability formula ; In the formula: A f Indicates the availability of fragment f; w c ,w m ,w b Represents the resource weights of CPU, memory, and bandwidth; f c ,f m ,f b This represents the CPU idle percentage, memory idle size, and bandwidth idle rate of fragment f; β f Indicates the fragment stability coefficient; The optimal utility formula for a combined task using the objective function of dynamic programming is: ; Its constraints are: ; ; In the formula: U opt This represents the maximum utility of the combined tasks that fragment f can carry; C represents a subset of low-priority tasks; T low This represents a set of low-priority tasks. d k [r] represents task t k The demand for resources r; F f [r] represents the resource capacity r of fragment f; u k Represents task t k The normalized utility value; d k [r] represents the resource type r required by task tk; F f [r] represents the available capacity of resource type r in resource fragment f.
9. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S6, the monitoring frequency is to collect data once every 10ms. If the current delay of the task is found to exceed the set threshold of 10%, the abnormal information is immediately pushed to S2 to adjust the priority and trigger S4 to regenerate the uninstallation decision. The time for S2 to adjust the priority does not exceed 20ms, and the time for S4 to re-determine the decision does not exceed 30ms. At the same time, the monitoring data is synchronized to the feedback module of S7 to provide data support for subsequent optimization.
10. The mobile edge computing task offloading method based on priority sorting as described in claim 1, characterized in that: In step S7, the uninstallation success rate, resource utilization rate, and latency compliance rate data during the task execution process are summarized. The uninstallation success rate is statistically analyzed once per hour. The resource utilization rate is the average utilization rate of CPU, memory, and bandwidth of each edge node. The latency compliance rate is the proportion of each task whose actual latency is within the set threshold.