Adaptive priority based hierarchical task offloading method based on mobile edge computing

Abstract

The application belongs to the technical field of mobile edge computing, and particularly relates to a layered task offloading method with adaptive priority based on mobile edge computing, which comprises the following steps: step 1, initial division of mobile edge computing resources; step 2, task classification; step 3, offloading decision making; and step 4, adaptive division of mobile edge computing resources. The application divides automatic driving tasks into two types of real-time tasks and ordinary tasks, and sets priority offloading queues for the two types of tasks respectively; meanwhile, the application performs logical deployment on the edge computing node cluster in a grouping and layering mode, and designs an adaptive resource division algorithm to dynamically adjust the number of nodes among different groups, so as to improve the offloading success rate of automatic driving real-time tasks.

Description

Technical Field

[0001] This invention belongs to the field of mobile edge computing technology, specifically relating to a hierarchical task offloading method based on adaptive priority in mobile edge computing. Background Technology

[0002] In recent years, research in the field of autonomous driving has increased significantly, and global companies are continuously increasing their spending on this area. Traditional cloud computing communication links are long, and autonomous vehicles generate massive amounts of data during operation. Continuing to use a cloud computing architecture would greatly extend task processing time, failing to meet the deadline requirements of real-time tasks and threatening passenger safety. Therefore, autonomous vehicles typically have onboard computing units to process some of the data. However, as the tasks generated by autonomous vehicles become increasingly complex and the data volume grows larger, onboard computing units alone cannot fully handle the processing. Therefore, offloading a large number of autonomous driving tasks to the edge is an inevitable path.

[0003] However, different types of autonomous driving tasks have different latency sensitivities. Real-time autonomous driving tasks are directly related to passenger safety and have strict deadline constraints, while the latency of ordinary autonomous driving tasks only affects the passenger's riding experience and has more lenient latency constraints. Faced with a massive number of autonomous driving tasks to be unloaded and a complex, heterogeneous cluster of edge computing nodes, how to select appropriate edge computing nodes for task unloading to improve the success rate of real-time task unloading is key to the research of task unloading algorithms combining autonomous driving and mobile edge computing. Considering the different latency constraints of different types of tasks in autonomous driving, an adaptive priority hierarchical task unloading method is proposed from the perspective of task classification. Summary of the Invention

[0004] The purpose of this invention is to address the problem that existing task offloading methods in the field of mobile edge computing do not consider the success rate of task offloading and therefore cannot guarantee passenger safety, thus making them unsuitable for autonomous driving scenarios. This invention proposes a hierarchical task offloading method based on adaptive priority in mobile edge computing. Compared with other existing task offloading algorithms, this invention can effectively improve the success rate of real-time task offloading.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The adaptive priority hierarchical task offloading method based on mobile edge computing includes the following steps:

[0007] Step 1: Initial allocation of mobile edge computing resources;

[0008] The computing performance parameters of the edge computing servers were obtained through quantitative analysis of computing performance, and the edge computing servers were ranked accordingly. The edge computing nodes were then divided into C groups based on the server resource allocation ratio η. high and C low Two parts; of which, C high Some edge computing nodes have higher computing performance and are used to offload real-time tasks in autonomous driving; C low Some edge computing nodes have relatively lower computing performance and are responsible for handling ordinary tasks;

[0009] Step 2: Task classification;

[0010] After receiving the task uploaded by the autonomous vehicle, the base station or SDN controller places the real-time task and the ordinary task into different edge task adaptive priority queues. The priority of the real-time task in the edge task adaptive priority queue is dynamically calculated by the initial priority, the latest deadline and the offloading waiting time, and is sorted in descending order of priority to wait for offloading.

[0011] Step 3: Unloading decision-making;

[0012] The base station or SDN controller makes offloading decisions based on task information; for real-time tasks, tasks with high adaptive priority are preferentially offloaded to the C stage. high In some node groups, edge computing nodes with stronger computing performance and the shortest real-time task waiting queues are selected; for ordinary tasks, high-priority tasks are preferentially offloaded to C. low Edge computing nodes with stronger task computing performance and fewer ready tasks are part of the node grouping.

[0013] Step 4: Adaptive allocation of mobile edge computing resources;

[0014] The base station or SDN controller periodically obtains the current real-time task offloading request density and adjusts the resource allocation ratio η according to the adaptive server resource allocation algorithm, modifying C. high With C low The number of edge computing nodes in a partial group.

[0015] As a preferred embodiment, in step 1, the computing performance parameters of the edge computing server include processor speed, memory space, I / O rate, static indicators, and dynamic indicators.

[0016] Static metrics include CPU frequency, number of cores, maximum number of threads, and cache capacity, while dynamic metrics include the system's IPS (Integrated Performance Index).

[0017] As a preferred embodiment, in step 2, the priority of the real-time task is dynamically calculated using the following formula based on the initial priority, the latest deadline, and the unloading waiting time:

[0018]

[0019] Among them, P ori The initial priority of the task is represented by dl2P(dl), and the priority is represented by the deadline. The initial priority and the priority represented by the deadline are given the same weight to calculate the adaptive priority. WT represents the unloading wait time of the task. When the unloading wait time of the task exceeds the average unloading wait time of the current priority, the priority of the task is increased.

[0020] As a preferred option, the priority of conversion based on task deadlines includes:

[0021] Calculate the difference between the current task's deadline and the current time to determine if the current task is the first task; if so, assign it the lowest priority.

[0022] For tasks that arrive later, calculate the urgency of their deadlines using deltaDL;

[0023] When deltaDL is less than the historical deltaDL maximum value maxDeltaDL, the priority is calculated according to the ratio of deltaDL to the historical deltaDL maximum value maxDeltaDL, and the maximum maxDeltaDL is updated at the same time.

[0024] When deltaDL is not less than the historical maximum deltaDL maxDeltaDL, it is given the lowest priority, and the maximum maxDeltaDL is updated.

[0025] The lowest priority is the priority corresponding to the deadline of the current tail task, and the maximum maxDeltaDL is updated at the same time.

[0026] As a preferred approach, the hierarchical offloading of edge computing nodes includes:

[0027] The edge computing nodes are sorted in two ways. The first round is to layer them according to the length of the task queue on the edge computing nodes, always keeping the task queue length of the top edge computing node the shortest. The second round is to sort them in descending order according to the computing performance of the nodes, that is, to prioritize unloading the real-time tasks with higher priority to the edge computing nodes with stronger computing performance and more idle time.

[0028] Traverse the edge computing nodes in the above order. If a task can be unloaded, that is, if the node meets the task's deadline requirement, then the computation is unloaded. If not, continue to look for a suitable node. If a task cannot find an edge computing node that meets the deadline requirement, then the unloading fails and the task is removed from the edge task adaptive priority queue.

[0029] As a preferred embodiment, step 4 specifically includes:

[0030] Obtain the current task unloading request density λ, and determine whether the difference between it and the request density in the previous period is greater than the detection threshold θ. The larger the detection threshold θ, the more frequently the detection frequency changes with the request density. When a detection is triggered, first determine whether the change in the current unloading success rate is within the allowable fluctuation range. If not, it indicates that the server resource allocation ratio η needs to be adjusted. The step size of η changes starts from 1%. Determine whether the task unloading success rate meets expectations after changing the resource allocation ratio. If not, increase the step size. If it meets the expectations, adjust the step size growth threshold to the median of the resource allocation ratio at the beginning and end.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] The adaptive priority hierarchical task unloading method of the present invention divides autonomous driving tasks into two types: real-time tasks and ordinary tasks, and sets up priority unloading queues for each type. At the same time, it performs grouped hierarchical logical deployment of edge computing node clusters and designs an adaptive resource partitioning algorithm to dynamically adjust the number of nodes between different groups, effectively improving the unloading success rate of real-time autonomous driving tasks. Attached Figure Description

[0033] Figure 1 This is a common task diagram for autonomous driving according to an embodiment of the present invention;

[0034] Figure 2 This is a diagram illustrating a single-base station MEC server deployment scheme according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the task unloading process according to an embodiment of the present invention;

[0036] Figure 4 This is an overall framework diagram of task unloading according to an embodiment of the present invention;

[0037] Figure 5 This is a comparison diagram of IPS for different edge devices according to embodiments of the present invention;

[0038] Figure 6 This is a hierarchical edge computing node task unloading diagram according to an embodiment of the present invention. Detailed Implementation

[0039] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0040] Typical autonomous driving tasks in this invention include vehicle detection, pedestrian detection, scene tracking, and path planning, with most detection tasks relying on data from various onboard sensors. For example... Figure 1 As shown, these include vehicle-mounted radar (Lidar), Global Positioning System (GPS), sonar (SONAR), and cameras. Due to the nature of autonomous driving, these tasks typically have strict time constraints and require significant computing power. For real-time tasks, the execution result directly relates to passenger safety; exceeding the time limit could potentially lead to irreversible consequences. A summary and analysis of the tasks generated by autonomous vehicles reveals the following characteristics:

[0041] (1) The tasks in autonomous driving scenarios are usually known and require the edge computing server to deploy services in advance before the unloading task can be run.

[0042] (2) The resource consumption of most real-time tasks in autonomous driving scenarios is usually known or can be calculated.

[0043] (3) Real-time tasks in autonomous driving scenarios mostly exhibit computationally intensive characteristics.

[0044] (4) Most real-time tasks in autonomous driving scenarios are initiated by the vehicle and unloaded to the edge computing center for processing, and finally the calculation results are obtained and processed accordingly.

[0045] Based on the above analysis, after classifying tasks in autonomous driving scenarios, real-time tasks require completion before the latest deadline, while ordinary tasks do not have such strict time requirements and can adopt a concurrent execution strategy. However, due to the limited resources of edge computing nodes and the increasing amount of task data generated by autonomous vehicles, ordinary tasks should maximize the utilization of edge computing resources; that is, the execution goal of ordinary tasks should be to improve the concurrent throughput of edge computing nodes.

[0046] In edge computing scenarios for autonomous driving, autonomous vehicles typically first communicate with nearby base stations to send edge tasks. The base stations then determine how to distribute these tasks among edge computing servers. For example... Figure 2The diagram shows an example deployment of a single-base station edge computing server under "FMPC," one of the mainstream MEC deployment solutions. SDN, short for Software-Defined Networking, is a key technology in the 5G era, just like MEC. Generally, the core of SDN is to separate the control and forwarding layers. With the introduction of SDN, edge computing servers can report their status information, including remaining computing resources, storage resources, and bandwidth pressure, to achieve unified and elastic management of the network and services.

[0047] In mobile edge computing task offloading systems, with the increasing number of autonomous vehicles, the scale of tasks in the network is also growing rapidly. To avoid task congestion at the base station, a task queue is added to the edge computing server, and a task offloading algorithm selects a suitable node for direct offloading. The edge computing server periodically reports its status information to the SDN controller at the base station via a Ping signal. Subsequently, the SDN controller executes the offloading algorithm based on the received task information and the edge computing server's status information, matching tasks with edge computing nodes. Figure 3 The diagram shown illustrates the task unloading process in an autonomous driving scenario.

[0048] The detailed uninstallation process is as follows:

[0049] (1) Autonomous vehicles connect to nearby base stations via the network and upload data of the task to be unloaded, including various metadata of the task such as the amount of business data, the latest deadline, and the amount of task computation.

[0050] (2) The SDN controller periodically receives status information reported by the edge computing server and maintains a status table to save the server status information.

[0051] (3) The SDN controller performs certain preprocessing work on the received mobile terminal tasks, combines the maintained edge computing server status table, runs the corresponding task offloading algorithm to calculate the task offloading request result, and sends the result back to the base station.

[0052] (4) The base station performs different operations based on the offloading algorithm processing result. If it does not offload, it informs the mobile terminal to process the task in the local VCU. If the offloading result gives the location of an edge computing node, it sends the task data to the corresponding edge computing node.

[0053] (5) The edge computing server receives the task execution and finally transmits the task execution result back to the mobile terminal.

[0054] like Figure 4As shown, the specific process of the adaptive priority hierarchical task offloading method (APHTO) of this invention is as follows:

[0055] Step 1: Initial Allocation of MEC (Mobile Edge Computing) Resources: The computing performance parameters of the edge computing servers are obtained through quantitative analysis of computing performance, and the edge computing servers are ranked accordingly. The edge computing nodes are then divided into C groups based on the server resource allocation ratio η. high and C low Two parts. C high Some edge nodes have stronger computing performance, which is used to offload real-time tasks in autonomous driving, while C low Some nodes have relatively lower computational performance and are responsible for handling ordinary tasks. (Regarding C...) high Some nodes are configured with real-time task waiting queues, and nodes are deployed in a hierarchical manner according to the queue length.

[0056] Step 2, Task Classification: After receiving tasks uploaded by autonomous vehicles, the base station or SDN controller places real-time tasks and ordinary tasks into different task unloading queues (i.e., edge task adaptive priority queues). The priority of real-time tasks is dynamically calculated based on the initial priority, the latest deadline, and the unloading waiting time, and they are sorted in descending order of priority for unloading.

[0057] Step 3: Offload Decision Making: The base station or SDN controller makes an offload decision based on task information. For real-time tasks, tasks with high adaptive priority are preferentially offloaded to C. high In node grouping, edge computing nodes with stronger computing performance and the shortest real-time task waiting queue are prioritized; for ordinary tasks, high-priority tasks are offloaded to C first. low The node group is the one with the strongest task computation performance and the fewest ready tasks.

[0058] Step 4: MEC Resource Adaptive Allocation: The base station or SDN controller periodically obtains the current real-time task offload request density and adjusts the resource allocation ratio η according to the adaptive server resource allocation algorithm, modifying C. high With C low Number of edge computing nodes in the group.

[0059] The design of the edge task adaptive priority queue is as follows:

[0060] Considering the relationship between task priority and deadline, an adaptive priority queue was designed. The adaptive priority is calculated from three factors: initial task priority, task deadline, and average unloading wait time. The specific calculation formula is shown below:

[0061]

[0062] Where P ori Let represent the initial priority of the task, and Dl2P(dl) represent the priority converted from the task's deadline. The initial priority and the priority converted from the deadline are given the same weight to calculate the adaptive priority. WT represents the task's unloading wait time. When the task's unloading wait time exceeds the average unloading wait time for the current priority, the task's priority is appropriately increased. The conversion between task deadline and priority is described in the following algorithm:

[0063]

[0064] For the first real-time task to arrive, it is assigned the lowest priority. For tasks that arrive afterward, the urgency of the deadline is calculated as deltaDL. When deltaDL is smaller, the priority is calculated according to the ratio of deltaDL to the historical maximum deltaDL maxDeltaDL. When deltaDL is larger, it is assigned the lowest priority, which is the priority corresponding to the deadline of the current tail task. At the same time, the maximum maxDeltaDL is updated.

[0065] Based on adaptive priority, the ordering of tasks in the unloading queue should follow the following principles:

[0066] (1) Tasks of equal priority are sorted in ascending order according to their arrival time, and the task that arrives first is selected for unloading.

[0067] (2) Tasks with different priorities that arrive at the same time are sorted in descending order of priority, and the task with higher priority is selected for unloading first.

[0068] (3) Set dynamic average unloading wait time for different priorities. When the unloading wait time of a task exceeds the average unloading wait time, the priority of the task is dynamically adjusted to ensure that the task can be successfully unloaded.

[0069] In summary, APHTO designed a task scheduling queue for the edge by combining factors such as the priority of real-time edge tasks and the average unloading wait time of tasks. The algorithm is as follows:

[0070]

[0071] The performance quantification design for edge nodes is as follows:

[0072] Edge computing server resources can be divided into computing resources and storage resources. Computing resources include CPU and GPU, while storage resources can be divided into internal storage resources such as RAM and external I / O read / write resources. Different resource types have different evaluation metrics. According to the evaluation metrics, the performance of the device can be roughly evaluated from the following aspects.

[0073] (1) Processor speed. Processing speed is a core indicator for measuring the computing power of a computer device. It can be expressed by the CPU's clock speed or the number of instructions executed per second.

[0074] (2) Memory space (storage capacity). Memory is the bridge between the CPU and external storage devices, and it is also the guarantee for the correct operation of applications and interaction with users. Generally speaking, the more memory space there is, the more tasks the computer can store and run concurrently.

[0075] (3) I / O rate (read and write capability). Applications involve writing and writing data to some extent, and a faster disk I / O rate means that data can be read or written more quickly, thus achieving a shorter CPU wait time.

[0076] (4) Real-time tasks in autonomous driving scenarios are mostly computationally intensive. Edge node computing performance metrics can be divided into two categories: static metrics and dynamic metrics. Static metrics refer to the basic attributes or routine operating metrics of the device, such as CPU frequency, number of cores, maximum number of threads, and cache capacity. The most common dynamic metric is CPU utilization. However, since CPU utilization cannot be directly used to measure the computing performance of the device, this invention uses the system's IPS as a performance metric. IPS represents the number of instructions executed per second in the system, which can effectively measure the CPU's computing power. Figure 5 The figure shows the changes in system IPS data collected when different edge devices (PCs, Raspberry Pis, and remote servers) perform the same CPU-intensive task. The IPS values ​​fluctuate slightly but remain relatively stable during the execution of the task on the same device. The differences between different devices are also basically fixed, making it suitable as an indicator to measure the computing power of edge devices.

[0077] The following algorithm describes the process of quantifying the computing performance of edge computing servers:

[0078]

[0079]

[0080] The design of the hierarchical offloading queue for edge computing nodes is as follows:

[0081] After adaptively prioritizing edge tasks and assessing the computing performance of edge computing servers, it's necessary to further partition edge computing nodes to maximize average resource utilization. Considering the massive amounts of communication data in the 5G era and the increasing number of autonomous vehicles, the number of tasks generated is also growing exponentially. Therefore, it's crucial to control the time complexity of the task offloading algorithm to maximize its execution efficiency; otherwise, congestion in the task offloading queue and even network congestion may occur, as some vehicles may have a mechanism for resending tasks after timeout. Congestion in the task offloading queue can lead to increased task execution latency or even the collapse of the mobile edge computing system, with unacceptable consequences for autonomous driving scenarios. Therefore, obtaining the task offloading algorithm results and distributing tasks as early as possible is essential for the edge task offloading queue. Since this method has already prioritized tasks arriving at the edge, the task offloading order is relatively stable. However, this doesn't mean that only single-threaded task distribution is possible; as long as the task traversal speed of subsequent threads doesn't surpass that of the parent thread or earlier threads, it's acceptable.

[0082] During task unloading, to minimize computational latency, edge computing nodes with higher computing power are prioritized. Simultaneously, to reduce the pressure on edge terminals for task distribution, task queues are first set up at the edge computing nodes to receive the assigned tasks. The length of the node's task queue is also taken into account. The final task distribution principle is to prioritize unloading high-priority tasks to edge computing nodes with stronger computing performance and shorter real-time task waiting queues. This principle applies to both real-time and ordinary tasks. When ordinary and real-time tasks coexist, since ordinary tasks have lower priority than real-time tasks, unloading ordinary tasks does not affect the success rate of real-time task unloading. Taking real-time tasks as an example, the specific unloading process is as follows: Figure 6 As shown.

[0083] The edge computing nodes are sorted in two stages. The first stage involves stratifying the nodes based on the length of their task queues, ensuring that the task queues within the top-level edge computing nodes are always the shortest. Generally, the shorter the queue length of a new task, the earlier it can be put into operation. The second stage involves sorting the nodes in descending order based on their computing performance as measured in the previous section. This prioritizes unloading higher-priority real-time tasks to edge computing nodes with stronger computing performance and more available space. The edge computing nodes are traversed in this order. If a task can be unloaded (meaning the node meets the task's deadline), then the computation is unloaded; otherwise, the search continues for a suitable node. If a task ultimately cannot find an edge computing node that meets its deadline, the unloading fails, and the task is removed from the unloading task queue. The specific implementation process can be described by the following algorithm:

[0084]

[0085]

[0086] The above algorithm is suitable for unloading real-time tasks in autonomous driving. For ordinary autonomous driving tasks, there is no latest deadline constraint, and task unloading can be performed directly. Considering that when a single thread performs computational unloading, if the current edge computing node cannot meet the latest deadline requirement of the current real-time task, an edge computing node traversal operation is required. Other real-time tasks should not incur the cost of waiting for the current traversal time, so a multi-threaded distribution design is needed. Although the introduction of multi-threading can speed up task distribution, attention must also be paid to thread synchronization issues. The principle of adaptively unloading higher-priority tasks to more powerful nodes cannot be violated. Therefore, a token bucket or direct locking is required. Considering that the process of determining whether a node meets the task's time requirement is not complicated and does not take long, locking the edge computing nodes is directly chosen. The final task distribution strategy is as described in the following algorithm:

[0087]

[0088]

[0089] The adaptive server resource partitioning algorithm is designed as follows:

[0090] When edge computing servers are not grouped, the offloading of ordinary tasks and real-time tasks are mixed together. Since real-time tasks and ordinary tasks have different latency sensitivities and scheduling methods, it is difficult to find the optimal offloading strategy. Grouping nodes allows for separate processing of different task types. However, when the density of real-time task offloading requests increases, such as during rush hour when traffic volume increases significantly, the number of real-time tasks also increases. To improve the success rate of real-time task offloading, more available edge nodes are needed; therefore, the node grouping ratio η needs to be adjusted. When changes in node grouping cause real-time tasks and ordinary tasks to coexist, since ordinary tasks have a lower priority than real-time tasks, it will not affect the execution latency of real-time tasks. The adaptive server resource allocation algorithm is shown below:

[0091]

[0092]

[0093] The algorithm periodically obtains the current task unloading request density λ and determines whether the difference between it and the request density in the previous period is greater than a detection threshold θ. A larger detection threshold θ indicates a more frequent detection frequency that changes with request density. When a detection is triggered, it first checks whether the change in the current unloading success rate is within the allowable fluctuation range. If not, it indicates that the server resource allocation ratio η needs to be adjusted, with η changing in increments of 1%. It then checks whether the task unloading success rate meets expectations after changing the resource allocation ratio. If not, the increment is increased, following a principle similar to the TCP congestion window (cwnd): when the increment is within the threshold ω, it increases rapidly; otherwise, it increases gradually. When the task unloading success rate meets expectations, the increment threshold is adjusted to the median of the resource allocation ratios at the beginning and end of the algorithm.

[0094] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A hierarchical task offloading method based on adaptive priority of mobile edge computing, characterized in that, Includes the following steps: Step 1: Initial allocation of mobile edge computing resources; The computing performance parameters of the edge computing servers are obtained through quantitative analysis of computing performance, and the edge computing servers are ranked accordingly. Based on server resource allocation ratio Divide edge computing nodes into and Two parts; of which, Some edge computing nodes are used to offload real-time tasks for autonomous driving; Some edge computing nodes are responsible for handling routine tasks; Step 2: Task classification; After receiving the task uploaded by the autonomous vehicle, the base station or SDN controller places the real-time task and the ordinary task into different edge task adaptive priority queues. The priority of the real-time task in the edge task adaptive priority queue is dynamically calculated by the initial priority, the latest deadline and the offloading waiting time, and is sorted in descending order of priority to wait for offloading. The priority of real-time tasks is dynamically calculated using the following formula: initial priority, latest deadline, and unloading wait time. ； in, Indicates the initial priority of the task. This indicates the priority derived from the task's deadline; the initial priority and the priority derived from the deadline are given the same weight to calculate the adaptive priority. This indicates the task's unloading wait time; when the task's unloading wait time exceeds the average unloading wait time for the current priority... At that time, increase the priority of the task; Step 3: Unloading decision-making; The base station or SDN controller makes offloading decisions based on task information; for real-time tasks, tasks with high adaptive priority are prioritized for offloading. In some node groups, edge computing nodes with stronger computing performance and the shortest real-time task waiting queues are selected; for ordinary tasks, high-priority tasks are preferentially offloaded to... Edge computing nodes with stronger task computing performance and fewer ready tasks are part of the node grouping. Step 4: Adaptive allocation of mobile edge computing resources; The base station or SDN controller periodically obtains the current real-time task offloading request density and adjusts the resource allocation ratio according to the adaptive server resource allocation algorithm. ,Revise Part and The number of edge computing nodes in a partial node grouping.

2. The hierarchical task offloading method based on adaptive priority using mobile edge computing as described in claim 1, characterized in that, In step 1, the computing performance parameters of the edge computing server include processor speed, memory space, I / O rate, static indicators, and dynamic indicators. Static metrics include CPU frequency, number of cores, maximum number of threads, and cache capacity, while dynamic metrics include the system's IPS (Integrated Performance Index).

3. The hierarchical task offloading method based on adaptive priority using mobile edge computing as described in claim 1, characterized in that, The priority derived from the deadline of the task includes: Calculate the difference between the current task's deadline and the current time to determine if the current task is the first task; if so, assign it the lowest priority. For tasks that arrive later, calculate the urgency of their deadlines. ; when Less than history Maximum value Then according to With history Maximum value The ratio is used to calculate priority, and the maximum is updated simultaneously. ; when No less than history Maximum value If so, assign it the lowest priority, and update the maximum priority. ; The lowest priority is the priority corresponding to the deadline of the current tail task, while the highest priority is updated. .

4. The hierarchical task offloading method based on adaptive priority using mobile edge computing as described in claim 3, characterized in that, Layered offloading of edge computing nodes includes: The edge computing nodes are sorted in two ways. The first round is to stratify them according to the length of the task queues on the edge computing nodes, always keeping the task queue length of the top-level edge computing nodes the shortest. The second round is to sort them in descending order according to the computing performance of the nodes. The edge computing nodes are traversed according to the above sorting. If a task is unloaded, that is, if the node meets the task's deadline requirement, then the computation is unloaded; if not, the search continues for a suitable node; if a task cannot find an edge computing node that meets the deadline requirement, then the unloading fails and the task is removed from the edge task adaptive priority queue.

5. The hierarchical task offloading method based on adaptive priority using mobile edge computing as described in claim 3, characterized in that, Step 4 specifically includes: Get the current task unload request density Determine whether the difference in request density compared to the previous period is greater than a detection threshold. Among them, the detection threshold The larger the value, the more frequently the detection frequency changes with request density. When a detection is triggered, the system first checks if the current unloading success rate is within the allowable fluctuation range. If not, it indicates that the server resource allocation ratio needs to be adjusted. , The change step size starts at 1%, and it is determined whether the task unloading success rate meets expectations after changing the resource allocation ratio. If it does not meet expectations, the step size is increased. If it does meet expectations, the step size growth threshold is adjusted to the median of the resource allocation ratio at the beginning and end.

Images