AI task optimization method and system based on edge dynamic routing

Abstract

The application discloses an AI task optimization method and system based on edge dynamic routing. The method comprises the following steps: constructing an edge dynamic routing network composed of a central controller, an edge node cluster and a programmable data plane; the central controller performs dynamic monitoring and self-calibration of node resource health, obtains the effective computing capacity of each edge node through integral method and multi-node redundant calibration mechanism of the detection task, and estimates the remaining healthy service life in combination with temperature compensation and aging curve; after receiving an AI task, a feature vector is extracted, which is input into a dynamic routing decision model based on deep reinforcement learning and graph attention network fusion together with the global state, an optimal routing decision is generated, and a forwarding rule is issued; the edge node executes the task, and triggers online adjustment and task migration when the health is abnormal. The application deeply integrates resource health monitoring and AI task routing, and improves the long-term reliability and task execution quality of the edge computing network.

Description

Technical Field

[0001] This invention relates to the fields of edge computing and artificial intelligence technology, specifically to an AI task optimization method and system based on edge dynamic routing. Background Technology

[0002] With the rapid development of artificial intelligence technology, AI inference tasks are becoming increasingly prevalent in various application scenarios. Edge computing, with its advantages of low latency, high bandwidth, and privacy protection, has become a key infrastructure architecture for AI task processing. However, executing AI tasks in a distributed edge environment faces numerous technical challenges.

[0003] On the one hand, edge nodes exhibit significant differences in computing power, storage resources, and network bandwidth, and task traffic is highly dynamic and bursty, making it difficult for traditional static routing strategies to adapt to these dynamic changes. On the other hand, AI tasks themselves have complex resource requirements, and a single scheduling strategy cannot accommodate the differentiated needs of all tasks. More importantly, edge nodes operate in complex physical environments for extended periods, and their hardware resources (such as GPUs, memory, and storage) can experience performance degradation due to aging, temperature fluctuations, and prolonged high loads. This health degradation is difficult to detect accurately in real time using traditional resource monitoring methods, potentially causing routing decision models to assign critical tasks to an edge node that appears to have sufficient resources but is actually close to failure, leading to task failures or a sharp decline in service quality.

[0004] Existing technologies include several task scheduling and routing optimization schemes for edge computing, such as dynamic task offloading through multi-dimensional perception or task scheduling through deep reinforcement learning. However, these schemes primarily focus on the instantaneous resource status of nodes (e.g., CPU utilization, available memory), lacking perception and modeling of long-term node health degradation trends, and also lacking self-calibration mechanisms to eliminate the interference of environmental factors on resource measurement accuracy. Furthermore, in existing schemes, task routing decision-making and node resource status monitoring are usually two separate processes, failing to form a closed-loop collaborative optimization of "monitoring-calibration-prediction-decision".

[0005] Therefore, there is an urgent need for an intelligent routing optimization method that can integrate dynamic monitoring of node health, adaptive calibration of environmental factors, and semantic perception of AI tasks, so as to improve the reliability and resource utilization efficiency of the system in the long term while ensuring the quality of real-time services. Summary of the Invention

[0006] This invention aims to provide an AI task optimization method and system based on edge dynamic routing. By introducing a dynamic monitoring and self-calibration mechanism for node resource health, and combining AI task feature perception and deep reinforcement learning routing decision-making, it achieves long-term reliable and efficient scheduling of AI tasks in edge networks.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] In a first aspect, the present invention provides an AI task optimization method based on edge dynamic routing, the method comprising the following steps:

[0009] Step S1: Construct an edge dynamic routing network architecture, which consists of a central controller and multiple edge nodes. The edge nodes are interconnected through a programmable data plane to form a distributed edge computing network.

[0010] Step S2: The central controller performs dynamic monitoring and self-calibration of node resource health to obtain the effective service capability assessment value of each edge node;

[0011] Step S3: Receive AI task requests from smart terminal devices, and collect the current global system status information by the global status perception module of the central controller;

[0012] Step S4: Perform task parsing and feature extraction on the AI ​​task request received in step S3, and extract the task feature vector of the AI ​​task;

[0013] Step S5: The global state vector obtained in step S3 and the task feature vector obtained in step S4 are jointly input into the routing strategy generation module. The routing strategy generation module adopts a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision for the current AI task.

[0014] Step S6: The task distribution module of the central controller, based on the routing decision generated in step S5, sends the corresponding forwarding rules to the programmable data plane to route the AI ​​task request to one or more selected edge nodes for execution.

[0015] Step S7: After the selected edge node receives the AI ​​task request, the local AI task queue management module puts the task into the processing queue of the corresponding priority, and the local computing resource scheduler allocates computing resources for inference execution.

[0016] Step S8: During the execution process, the dynamic routing execution module of the edge node monitors the status changes of the node and its neighboring nodes in real time; when it detects node overload, link congestion or a sharp deterioration in node health, it initiates an online adjustment mechanism and generates dynamic adjustment instructions.

[0017] Step S9: After each edge node completes its task, it returns the inference results to the terminal device and reports the task execution quality indicators and the node self-check data collected during the execution process to the central controller, which serve as training samples for the subsequent online updates of the health calibration model and the routing decision model.

[0018] Furthermore, the central controller performs dynamic monitoring and self-calibration of node resource health to obtain an effective service capability assessment value for each edge node. The specific process is as follows:

[0019] The node health monitoring and calibration module of the central controller periodically sends probing task instructions to each edge node. The probing task is a series of standardized, lightweight benchmark test tasks, including computationally intensive micro-tasks, memory read / write test tasks, and network loopback test tasks.

[0020] The node health monitoring and calibration module calculates the current effective computing power capacity of each edge node based on the raw measurement data reported in multiple detection cycles.

[0021] A multi-node redundant calibration mechanism is used to correct measurement errors.

[0022] Furthermore, the current effective computing power capacity of each edge node is calculated, and the specific process is as follows:

[0023] For computationally intensive tasks, the cumulative number of task units processed by node i during continuous discharge testing is... The processing rate varies with time t; the cumulative processing volume is obtained by integrating the processing rate over the test period [t1, t2], as shown in the following formula:

[0024]

[0025] in, Instantaneous processing rate;

[0026] Combined with node voltage changes and frequency change The effective computing power capacity is calculated using the following formula:

[0027]

[0028] in, Let be the voltage change at node i. Let be the frequency change of node i.

[0029] Furthermore, the multi-node redundancy calibration mechanism includes:

[0030] Select at least two edge nodes to execute the same calibration task in parallel, and measure the task response frequency of each node. ;

[0031] The capacity is estimated by inversely calculating the formula based on the resonance principle. The specific formula is as follows:

[0032]

[0033] in, The inherent load characteristic coefficients for the calibration task;

[0034] The difference between the estimated capacities of different nodes is calculated as the estimation error. When the error exceeds a preset threshold, a retest is triggered or a voting mechanism is used to correct it.

[0035] Furthermore, step S2 also includes temperature effect compensation, as detailed below:

[0036] After obtaining effective computing power capacity, the ambient temperature of the current environment of the edge node and the historical peak ambient temperature of the edge node in the long-term operation history are collected.

[0037] According to the pre-stored temperature influence coefficient table, which records the attenuation ratio of the node's effective computing power under different ambient temperatures relative to the baseline computing power under standard temperatures, the current temperature influence coefficient corresponding to the current ambient temperature is queried, and the effective computing power capacity is divided by the current temperature influence coefficient to obtain the standardized effective computing power capacity after excluding the influence of the current temperature.

[0038] Meanwhile, the peak temperature influence coefficient corresponding to the historical peak ambient temperature is queried, and the preset node failure computing power threshold is divided by the peak temperature influence coefficient to obtain the differentiated failure threshold under the peak temperature condition.

[0039] Then, based on the pre-stored node aging curve data, which records the relationship between the standardized effective computing power capacity of edge nodes of the same model under standard temperature and the cumulative running time, and combined with the current standardized effective computing power capacity, the differentiated failure threshold and the initial effective computing power capacity of the node, the remaining time required for the edge node to decay from the current state to the failure threshold is estimated through the inverse function of the aging curve, which is taken as the remaining healthy service life of the edge node.

[0040] Furthermore, the state space of the dynamic routing decision model in step S5 includes the standardized effective computing power capacity and remaining health lifetime of each edge node. The reward function of the model includes a health penalty term, which is calculated as follows: when the remaining health lifetime of a node is greater than or equal to the preset health lifetime threshold, the penalty term is zero; when the remaining health lifetime of a node is less than the preset health lifetime threshold, the penalty term takes a negative value, and its absolute value is the difference between the preset penalty coefficient multiplied by one and the ratio of the remaining health lifetime to the health lifetime threshold.

[0041] Furthermore, the online adjustment mechanism in step S8 includes: when the node self-checking module of the edge node detects that the standardized effective computing power capacity decreases by more than a preset drop threshold within two consecutive detection cycles, the edge node actively sends a health alarm to the central controller; after receiving the alarm, the central controller removes the edge node from the candidate list of high-criticality tasks and triggers the migration of low-criticality tasks already assigned to the edge node to adjacent healthy nodes for execution.

[0042] Secondly, the present invention provides an AI task optimization system based on edge dynamic routing for implementing the AI ​​task optimization method. The system includes a central controller, a distributed edge computing cluster composed of multiple edge nodes, and a programmable data plane. The programmable data plane is composed of several programmable switches and connects the central controller and each edge node.

[0043] The central controller includes:

[0044] The global state awareness module is configured to receive and aggregate state information from all edge nodes to construct a global system state vector.

[0045] The task parsing and feature extraction module is configured to perform semantic parsing on received AI task requests and extract task feature vectors.

[0046] The node health monitoring and calibration module is configured to periodically send detection tasks to edge nodes, receive raw measurement data, calculate the effective service capability assessment value of each node and eliminate the influence of temperature, and estimate the remaining health lifespan.

[0047] The routing strategy generation module is configured to use a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision based on the global state vector, node health features and task feature vector.

[0048] The task distribution module is configured to distribute forwarding rules to the programmable data plane based on the output of the routing policy generation module; and

[0049] The model training and update module is configured to perform offline pre-training and online fine-tuning and updating of the dynamic routing decision model using historical data and real-time feedback samples.

[0050] Each edge node includes:

[0051] The local computing resource monitoring module is configured to monitor the CPU, GPU, memory, and disk resource usage of this node in real time and periodically report to the central controller.

[0052] The node self-test module is configured to execute the detection tasks issued by the central controller, collect raw data such as node voltage, frequency, temperature and task processing rate, and monitor health change events.

[0053] The AI ​​task queue management module is configured to maintain a multi-priority task queue and to queue and schedule received AI tasks.

[0054] The local computing resource scheduler is configured to allocate computing resources to execute AI inference tasks based on task priority and resource requirements;

[0055] The dynamic routing execution module is configured to execute the forwarding rules issued by the central controller, monitor the status changes of the local node and its neighboring nodes, and trigger the local adjustment mechanism when an anomaly is detected.

[0056] The inference result feedback module is configured to return the task execution results to the terminal device and report the execution quality indicators.

[0057] Thirdly, the present invention also provides a computer-readable storage medium for storing one or more programs, said one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any of the methods described.

[0058] Fourthly, the present invention provides a computing device for an AI task optimization method based on edge dynamic routing, characterized in that it includes:

[0059] One or more processors, memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for any of the methods described.

[0060] The beneficial effects of this invention are:

[0061] This invention obtains a dynamic estimate of the true computing power capacity of a node by integrating the processing rate of a continuous detection task and combining it with a voltage-frequency change factor, thus overcoming the deficiency that simple CPU utilization indicators cannot reflect performance degradation.

[0062] This invention establishes a mapping relationship between the processing response frequency of standard tasks and the computing power of nodes, and estimates and corrects measurement errors by having multiple nodes execute the same calibration task in parallel, which greatly improves the accuracy of health assessment.

[0063] This invention introduces a temperature influence coefficient to standardize the measured values ​​and combines it with historical peak temperatures to determine the failure threshold. Finally, it uses aging curves to predict the remaining service life of nodes, enabling routing decisions to proactively avoid nodes that are about to fail. Attached Figure Description

[0064] The invention will now be further described with reference to the accompanying drawings.

[0065] Figure 1 This is a flowchart of the AI ​​task optimization method based on edge dynamic routing in an embodiment of the present invention;

[0066] Figure 2 This is a block diagram of an AI task optimization system based on edge dynamic routing in an embodiment of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figure 1 This invention provides an AI task optimization method based on edge dynamic routing, the method comprising:

[0069] Step S1: Construct an edge dynamic routing network architecture, which consists of a central controller and multiple edge nodes. The edge nodes are interconnected through a programmable data plane to form a distributed edge computing network.

[0070] The central controller is equipped with a global status awareness module, a routing policy generation module, a task distribution module, and a node health monitoring and calibration module; each edge node is equipped with a local computing resource monitoring module, an AI task queue management module, a dynamic routing execution module, and a node self-test module; edge nodes exchange data through a programmable data plane, which supports dynamic distribution of packet processing and forwarding policies based on stream granularity.

[0071] Step S2: The central controller performs dynamic monitoring and self-calibration of node resource health to obtain the effective service capability assessment value of each edge node;

[0072] The central controller's node health monitoring and calibration module periodically sends probing task instructions to each edge node. These probing tasks are a series of standardized, lightweight benchmark tests, including computationally intensive micro-tasks, memory read / write tests, and network loopback tests. Upon receiving the probing task instructions, each edge node's self-test module executes the corresponding tests and reports the following raw measurement data:

[0073] Changes in processing time and power consumption when node i performs computationally intensive tasks;

[0074] Effective bandwidth and access latency when node i performs memory read / write tests

[0075] The minimum round-trip time and jitter when node i performs a network loopback test;

[0076] The node health monitoring and calibration module calculates the current effective computing power capacity of each edge node based on the raw measurement data reported in multiple detection cycles.

[0077] For computationally intensive tasks, the cumulative number of task units processed by node i during continuous discharge testing is... The processing rate varies with time t; the cumulative processing volume is obtained by integrating the processing rate over the test period [t1, t2], as shown in the following formula:

[0078]

[0079] in, Instantaneous processing rate;

[0080] Combined with node voltage changes and frequency change Calculate effective computing power capacity The formula is as follows:

[0081]

[0082] in, Let be the voltage change at node i. Let be the frequency change of node i.

[0083] The node's processing capacity decreases as the supply voltage and operating frequency degrade. This reflects the overall performance degradation factor of the node during the test.

[0084] A redundant calibration mechanism is introduced to reduce measurement errors. In the edge node cluster, a node with a known performance benchmark is selected as the reference node (or a virtual set of calibration nodes is constructed), and both nodes perform the same probing task. Specifically, the node health monitoring and calibration module maintains a standard calibration task library, and during each probing, a calibration task is randomly selected and assigned to at least two different edge nodes for parallel execution. Assuming node i and node j execute the same calibration task simultaneously, their processing times are obtained. and and the system response frequency observed during task execution. and ;

[0085] According to the principle of resonance response, for a given calibration task, its expected processing frequency and the node's baseline computing power satisfy the following:

[0086]

[0087] in, For effective computing capacity, The inherent load characteristic coefficients for the calibration task.

[0088] The capacity is estimated by inversely calculating the formula based on the resonance principle. The specific formula is as follows:

[0089]

[0090] in, The inherent load characteristic coefficients for the calibration task.

[0091] Similarly, compute for node j The estimation error of node i relative to node j is If the absolute value of the error exceeds the preset error threshold, it is judged as a measurement anomaly, triggering a retest or a multi-node voting mechanism for correction. Finally, the effective service capability evaluation value of node i is the median or weighted average of multiple calibration results.

[0092] After obtaining effective computing power capacity, the ambient temperature of the current environment of the edge node and the historical peak ambient temperature of the edge node in the long-term operation history are collected.

[0093] According to the pre-stored temperature influence coefficient table, which records the attenuation ratio of the node's effective computing power under different ambient temperatures relative to the baseline computing power under standard temperatures, the current temperature influence coefficient corresponding to the current ambient temperature is queried, and the effective computing power capacity is divided by the current temperature influence coefficient to obtain the standardized effective computing power capacity after excluding the influence of the current temperature.

[0094] Meanwhile, the peak temperature influence coefficient corresponding to the historical peak ambient temperature is queried, and the preset node failure computing power threshold is divided by the peak temperature influence coefficient to obtain the differentiated failure threshold under the peak temperature condition.

[0095] Then, based on the pre-stored node aging curve data, which records the relationship between the standardized effective computing power capacity of edge nodes of the same model under standard temperature and the cumulative running time, and combined with the current standardized effective computing power capacity, the differentiated failure threshold and the initial effective computing power capacity of the node, the remaining time required for the edge node to decay from the current state to the failure threshold is estimated through the inverse function of the aging curve, which is taken as the remaining healthy service life of the edge node.

[0096] Step S3: Receive AI task requests from smart terminal devices, and collect the current global system status information by the global status perception module of the central controller;

[0097] The global system status information includes: available computing resources (real-time dynamic value), available bandwidth resources, current task queue length, terminal device location information, network topology, and the effective service capability assessment value and remaining health lifespan of each node obtained in step S2.

[0098] Step S4: Perform task parsing and feature extraction on the AI ​​task request received in step S3, and extract the task feature vector of the AI ​​task;

[0099] The task parsing and feature extraction specifically include: performing session-level semantic parsing on AI task requests to identify task type, input data scale, and expected quality of service constraints (including maximum tolerable latency limit, minimum available bandwidth requirement, and reliability requirement).

[0100] Extract multidimensional feature vectors for the task. For complex AI tasks (such as multimodal fusion tasks), the task parsing module recursively decomposes the task into several atomic operators with dependencies, forming a task structure represented by a directed acyclic graph.

[0101] Step S5: The global state vector obtained in step S3 and the task feature vector obtained in step S4 are jointly input into the routing strategy generation module. The routing strategy generation module adopts a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision for the current AI task.

[0102] The state space of the dynamic routing decision model includes the standardized effective computing power capacity and remaining health lifetime of each edge node. The reward function of the model includes a health penalty term, which is calculated as follows: when the remaining health lifetime of a node is greater than or equal to the preset health lifetime threshold, the penalty term is zero; when the remaining health lifetime of a node is less than the preset health lifetime threshold, the penalty term takes a negative value, and its absolute value is the difference between the preset penalty coefficient multiplied by one and the ratio of the remaining health lifetime to the health lifetime threshold.

[0103] Step S6: The task distribution module of the central controller, based on the routing decision generated in step S5, sends the corresponding forwarding rules to the programmable data plane to route the AI ​​task request to one or more selected edge nodes for execution.

[0104] The central controller's task distribution module transforms the routing decisions generated in the previous step into flow table entries that can be executed by the programmable data plane. The specific operation is as follows:

[0105] For each atomic task, the task dispatch module constructs a flow table entry. The flow table entry consists of two parts:

[0106] Matching fields include, for example, the source IP address, destination IP address, protocol type, and a specially defined task tag (which is encapsulated by the controller when the task request enters the network to identify the task).

[0107] Actions performed include forwarding data to a specific port on a switch, encapsulating task metadata (such as priority and target node ID), and modifying packet headers.

[0108] Priority: A higher value indicates a higher priority, used to resolve matching conflicts.

[0109] The controller uses a standard protocol for controlling P4 switches to distribute these flow table entries to the switches located on the task entry path in the programmable data plane. The switches take effect immediately upon receiving the entries.

[0110] For complex tasks with DAG structures, it is necessary to issue multiple interrelated flow entries simultaneously to construct a "virtual data path" to ensure that intermediate results between atomic operators are transmitted to the designated node in the correct order.

[0111] Once the flow table entry is deployed, subsequent AI request data packets belonging to this task will be automatically forwarded to the target edge node along the preset path, without having to go through the controller, thus achieving a high-speed data channel.

[0112] Step S7: After the selected edge node receives the AI ​​task request, the local AI task queue management module puts the task into the processing queue of the corresponding priority, and the local computing resource scheduler allocates computing resources for inference execution.

[0113] After the target edge node receives the AI ​​task request data packet:

[0114] First, the node's AI task queue management module parses task information, including task type, input data, and priority in feature vectors;

[0115] Based on the latency sensitivity level in the task characteristics, tasks are placed into the corresponding priority queues. High-sensitivity tasks are placed in the high-priority queue and preemptive scheduling is used (i.e., low-priority tasks that are currently executing can be interrupted); medium- and low-sensitivity tasks are placed in the normal or low-priority queues and weighted fair queues are used (i.e., execution time slices are allocated according to weights).

[0116] The local computing resource scheduler retrieves tasks from the queue, allocates corresponding computing resources (CPU cores, GPU stream processors, memory, etc.), loads the corresponding AI models, and performs inference calculations.

[0117] For tasks that require streaming results (such as real-time video analytics), edge nodes can use an inference-while-output approach to reduce initial packet latency.

[0118] During execution, the local computing resource monitoring module of the edge node continuously records resource usage (such as changes in GPU utilization) and periodically (e.g., every 1 second) reports it to the central controller.

[0119] Step S8: During the execution process, the dynamic routing execution module of the edge node monitors the status changes of the node and its neighboring nodes in real time; when it detects node overload, link congestion or a sharp deterioration in node health, it initiates an online adjustment mechanism and generates dynamic adjustment instructions.

[0120] The online adjustment mechanism includes: when the node self-inspection module of an edge node detects that the standardized effective computing power capacity decreases by more than a preset drop threshold within two consecutive detection cycles, the edge node actively sends a health alarm to the central controller; after receiving the alarm, the central controller removes the edge node from the candidate list of high-criticality tasks and triggers the migration of low-criticality tasks already assigned to the edge node to adjacent healthy nodes for execution.

[0121] Step S9: After each edge node completes its task, it returns the inference results to the terminal device and reports the task execution quality indicators and the node self-check data collected during the execution process to the central controller, which serve as training samples for the subsequent online updates of the health calibration model and the routing decision model.

[0122] After the edge node completes AI inference, it generates a result data packet.

[0123] Result Return: The dynamic routing execution module searches for the optimal return path based on the return address in the original request. If the original return path is still available, the original reverse flow table entry is reused; if the original path is congested, the path with the lowest latency is selected from the local alternative path table to return to the terminal device.

[0124] Quality Reporting: Edge nodes simultaneously report a detailed quality report of the task execution to the central controller, including: actual latency (total time from the request arriving at the controller to the result being sent), resource consumption (GPU time, CPU time, amount of data transmitted), task completion status (success, failure, or timeout), and the latest resource status of the node during task execution (such as changes in GPU utilization after execution).

[0125] Online Model Updates: The model training and update module of the central controller combines the aforementioned quality report with the health data obtained in step S2 to calculate the actual reward value for this round of decision-making (corresponding to the reward function during training). This reward value, along with the current global state and task characteristics, forms an "experience sample" and is stored in the experience replay pool. Every so often (e.g., after collecting 100 new samples), the module triggers a background online fine-tuning training, using these new samples to perform small-batch updates to the parameters of the dynamic routing decision model. This allows the model to continuously adapt to changes in the edge network environment (such as node aging and workload pattern drift), achieving lifelong learning.

[0126] Please see Figure 2 Another embodiment of the present invention provides an AI task optimization system based on edge dynamic routing. The system includes a central controller, a distributed edge computing cluster consisting of multiple edge nodes, and a programmable data plane. The programmable data plane consists of several programmable switches that connect the central controller and each edge node.

[0127] The central controller includes:

[0128] The global state awareness module is configured to receive and aggregate state information from all edge nodes to construct a global system state vector.

[0129] The task parsing and feature extraction module is configured to perform semantic parsing on received AI task requests and extract task feature vectors.

[0130] The node health monitoring and calibration module is configured to periodically send detection tasks to edge nodes, receive raw measurement data, calculate the effective service capability assessment value of each node and eliminate the influence of temperature, and estimate the remaining health lifespan.

[0131] The routing strategy generation module is configured to use a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision based on the global state vector, node health features and task feature vector.

[0132] The task distribution module is configured to distribute forwarding rules to the programmable data plane based on the output of the routing policy generation module; and

[0133] The model training and update module is configured to perform offline pre-training and online fine-tuning and updating of the dynamic routing decision model using historical data and real-time feedback samples.

[0134] Each edge node includes:

[0135] The local computing resource monitoring module is configured to monitor the CPU, GPU, memory, and disk resource usage of this node in real time and periodically report to the central controller.

[0136] The node self-test module is configured to execute the detection tasks issued by the central controller, collect raw data such as node voltage, frequency, temperature and task processing rate, and monitor health change events.

[0137] The AI ​​task queue management module is configured to maintain a multi-priority task queue and to queue and schedule received AI tasks.

[0138] The local computing resource scheduler is configured to allocate computing resources to execute AI inference tasks based on task priority and resource requirements;

[0139] The dynamic routing execution module is configured to execute the forwarding rules issued by the central controller, monitor the status changes of the local node and its neighboring nodes, and trigger the local adjustment mechanism when an anomaly is detected.

[0140] The inference result feedback module is configured to return the task execution results to the terminal device and report the execution quality indicators.

[0141] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0143] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0144] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0145] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0146] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. An AI task optimization method based on edge dynamic routing, characterized in that, The method includes the following steps: Step S1: Construct an edge dynamic routing network architecture, which consists of a central controller and multiple edge nodes. The edge nodes are interconnected through a programmable data plane to form a distributed edge computing network. Step S2: The central controller performs dynamic monitoring and self-calibration of node resource health to obtain the effective service capability assessment value of each edge node; Step S3: Receive AI task requests from smart terminal devices, and collect the current global system status information by the global status perception module of the central controller; Step S4: Perform task parsing and feature extraction on the AI ​​task request received in step S3, and extract the task feature vector of the AI ​​task; Step S5: The global state vector obtained in step S3 and the task feature vector obtained in step S4 are jointly input into the routing policy generation module. The routing policy generation module adopts a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision for the current AI task. Step S6: The task distribution module of the central controller, based on the routing decision generated in step S5, sends the corresponding forwarding rules to the programmable data plane to route the AI ​​task request to one or more selected edge nodes for execution. Step S7: After the selected edge node receives the AI ​​task request, the local AI task queue management module puts the task into the processing queue of the corresponding priority, and the local computing resource scheduler allocates computing resources for inference execution. Step S8: During the execution process, the dynamic routing execution module of the edge node monitors the status changes of the node and its neighboring nodes in real time; when it detects node overload, link congestion or a sharp deterioration in node health, it initiates an online adjustment mechanism and generates dynamic adjustment instructions. Step S9: After each edge node completes its task, it returns the inference results to the terminal device and reports the task execution quality indicators and the node self-check data collected during the execution process to the central controller, which serve as training samples for the subsequent online updates of the health calibration model and the routing decision model.

2. The AI ​​task optimization method based on edge dynamic routing according to claim 1, characterized in that, The central controller performs dynamic monitoring and self-calibration of node resource health to obtain an effective service capability assessment value for each edge node. The specific process is as follows: The node health monitoring and calibration module of the central controller periodically sends probing task instructions to each edge node. The probing task is a series of standardized, lightweight benchmark test tasks, including computationally intensive micro-tasks, memory read / write test tasks, and network loopback test tasks. The node health monitoring and calibration module calculates the current effective computing power capacity of each edge node based on the raw measurement data reported in multiple detection cycles. A multi-node redundant calibration mechanism is used to correct measurement errors.

3. The AI ​​task optimization method based on edge dynamic routing according to claim 2, characterized in that, The current effective computing power capacity of each edge node is calculated as follows: For computationally intensive tasks, the cumulative number of task units processed by node i during continuous discharge testing is... The processing rate varies with time t; the cumulative processing volume is obtained by integrating the processing rate over the test period [t1, t2], as shown in the following formula: in, Instantaneous processing rate; Combined with node voltage changes and frequency change Calculate effective computing power capacity The formula is as follows: in, Let be the voltage change at node i. Let be the frequency change of node i.

4. The AI ​​task optimization method based on edge dynamic routing according to claim 2, characterized in that, The multi-node redundancy calibration mechanism includes: Select at least two edge nodes to execute the same calibration task in parallel, and measure the task response frequency of each node. ; The capacity is estimated by inversely calculating the formula based on the resonance principle. The specific formula is as follows: in, The inherent load characteristic coefficients for the calibration task; The difference between the estimated capacities of different nodes is calculated as the estimation error. When the error exceeds a preset threshold, a retest is triggered or a voting mechanism is used to correct it.

5. The AI ​​task optimization method based on edge dynamic routing according to claim 4, characterized in that, Step S2 further includes temperature effect compensation, as detailed below: After obtaining effective computing power capacity, the ambient temperature of the current environment of the edge node and the historical peak ambient temperature of the edge node in the long-term operation history are collected. According to the pre-stored temperature influence coefficient table, which records the attenuation ratio of the node's effective computing power under different ambient temperatures relative to the baseline computing power under standard temperatures, the current temperature influence coefficient corresponding to the current ambient temperature is queried, and the effective computing power capacity is divided by the current temperature influence coefficient to obtain the standardized effective computing power capacity after excluding the influence of the current temperature. Meanwhile, the peak temperature influence coefficient corresponding to the historical peak ambient temperature is queried, and the preset node failure computing power threshold is divided by the peak temperature influence coefficient to obtain the differentiated failure threshold under the peak temperature condition. Then, based on the pre-stored node aging curve data, which records the relationship between the standardized effective computing power capacity of edge nodes of the same model under standard temperature and the cumulative running time, and combined with the current standardized effective computing power capacity, the differentiated failure threshold and the initial effective computing power capacity of the node, the remaining time required for the edge node to decay from the current state to the failure threshold is estimated through the inverse function of the aging curve, which is taken as the remaining healthy service life of the edge node.

6. The AI ​​task optimization method based on edge dynamic routing according to claim 1, characterized in that, The state space of the dynamic routing decision model in step S5 includes the standardized effective computing power capacity and remaining health lifetime of each edge node. The reward function of the model includes a health penalty term, which is calculated as follows: when the remaining health lifetime of a node is greater than or equal to the preset health lifetime threshold, the penalty term is zero; when the remaining health lifetime of a node is less than the preset health lifetime threshold, the penalty term takes a negative value, and its absolute value is the difference between the preset penalty coefficient multiplied by one and the ratio of the remaining health lifetime to the health lifetime threshold.

7. The AI ​​task optimization method based on edge dynamic routing according to claim 1, characterized in that, The online adjustment mechanism in step S8 includes: when the node self-test module of the edge node detects that the standardized effective computing power capacity decreases by more than the preset drop threshold in two consecutive detection cycles, the edge node actively sends a health alarm to the central controller; after receiving the alarm, the central controller removes the edge node from the candidate list of high-criticality tasks and triggers the migration of low-criticality tasks that have been assigned to the edge node to adjacent healthy nodes for execution.

8. An AI task optimization system based on edge dynamic routing, used to implement the AI ​​task optimization method according to any one of claims 1-7, characterized in that, The system includes a central controller, a distributed edge computing cluster consisting of multiple edge nodes, and a programmable data plane. The programmable data plane consists of several programmable switches that connect the central controller and each edge node. The central controller includes: The global state awareness module is configured to receive and aggregate state information from all edge nodes to construct a global system state vector. The task parsing and feature extraction module is configured to perform semantic parsing on received AI task requests and extract task feature vectors. The node health monitoring and calibration module is configured to periodically send detection tasks to edge nodes, receive raw measurement data, calculate the effective service capability assessment value of each node and eliminate the influence of temperature, and estimate the remaining health lifespan. The routing strategy generation module is configured to use a dynamic routing decision model based on the fusion of deep reinforcement learning and graph attention network to generate the optimal routing decision based on the global state vector, node health features and task feature vector. The task distribution module is configured to distribute forwarding rules to the programmable data plane based on the output of the routing policy generation module; and The model training and update module is configured to perform offline pre-training and online fine-tuning and updating of the dynamic routing decision model using historical data and real-time feedback samples. Each edge node includes: The local computing resource monitoring module is configured to monitor the CPU, GPU, memory, and disk resource usage of this node in real time and periodically report to the central controller. The node self-test module is configured to execute the detection tasks issued by the central controller, collect raw data such as node voltage, frequency, temperature and task processing rate, and monitor health change events. The AI ​​task queue management module is configured to maintain a multi-priority task queue and to queue and schedule received AI tasks. The local computing resource scheduler is configured to allocate computing resources to execute AI inference tasks based on task priority and resource requirements; The dynamic routing execution module is configured to execute the forwarding rules issued by the central controller, monitor the status changes of the local node and its neighboring nodes, and trigger the local adjustment mechanism when an anomaly is detected. The inference result feedback module is configured to return the task execution results to the terminal device and report the execution quality indicators.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1-7.

10. A computing device for an AI task optimization method based on edge dynamic routing, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods according to claims 1-7.

Images