A method for offloading and scheduling interactive tasks and resources in smart terminals based on edge-side state awareness

Abstract

This invention relates to the field of task scheduling technology, specifically to a method for unloading and scheduling interactive tasks and resources in intelligent terminals based on edge-side state awareness. The method includes the following steps: collecting CPU utilization, GPU utilization, and communication latency data; smoothing communication latency to form an edge-side state indicator set; calculating a suitability index based on the edge-side state indicator set and task characteristics; generating a priority ranking table by scoring and fusing according to rules; selecting the first-ranked path; analyzing computing power and latency to construct an execution strategy; invoking dynamic resource scheduling; monitoring latency and providing feedback to adjust weights to generate resource scheduling instructions. In this invention, by collecting CPU and GPU utilization and communication latency in real time, using a sliding window to smooth and optimize scheduling, weighting the maximum tolerable latency with data size to match edge-side state and generate priorities, and dynamically adjusting resource allocation to match execution paths and network latency, this method avoids resource waste and task delays caused by static rules, thereby improving performance and utilization.

Description

Technical Field

[0001] This invention relates to the field of task scheduling technology, and in particular to a method for unloading and scheduling interactive tasks and resources of intelligent terminals based on terminal-side state awareness. Background Technology

[0002] The field of task scheduling technology involves core aspects of task allocation, execution, and resource management in computing systems. It mainly studies how to efficiently arrange task execution order, allocate computing resources, optimize task completion time, and system throughput in multiprocessor, distributed systems, or cloud computing environments. Key technologies include task queue management, priority setting, load balancing, resource allocation strategies, task migration, and energy consumption control. It is widely used in operating systems, cloud computing platforms, edge computing, and mobile computing scenarios, and achieves optimal system performance and maximized resource utilization through reasonable scheduling strategies.

[0003] Among them, the traditional interactive task offloading and resource scheduling method refers to the technical solution for migrating computationally intensive tasks from terminal devices to cloud servers for execution when smart terminals perform user interaction tasks, which are faced with problems such as insufficient computing power, limited battery life, and network bandwidth fluctuations. This is achieved by using preset task classification rules to determine the amount of computation of the task, selecting the offloading time based on fixed network status parameters, determining the cloud virtual machine configuration according to a predefined resource allocation table, scheduling the execution order of multiple concurrent tasks using a polling method, and estimating the task transmission time and execution time based on historical averages to complete the offloading decision and resource allocation.

[0004] Existing technologies rely on static rules and fixed parameters for task offloading and resource scheduling, making it difficult to cope with dynamically changing computing and network environments. This leads to delays or misjudgments in task offloading decisions. Because real-time terminal load and network latency information is not available, task scheduling decisions cannot reflect the actual state of devices or changes in the network. Estimation methods using raw data cannot provide accurate predictions of transmission and execution times in complex and uncertain network environments, easily resulting in uneven resource allocation and impacting task execution efficiency and overall performance. Existing solutions also fail to effectively optimize task scheduling strategies, cannot adapt to load fluctuations or resource constraints, and reduce system flexibility and resource utilization efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely on static rules and fixed parameters for task offloading and resource scheduling, which struggle to cope with dynamically changing computing and network environments and lead to delayed or misjudged task offloading decisions, and to prevent task scheduling decisions from reflecting actual device status or network changes due to the lack of real-time terminal load and network latency information, historical data estimation methods cannot provide accurate transmission and execution time predictions in complex and uncertain network environments, easily resulting in uneven resource allocation and impacting task execution efficiency and overall performance, and to prevent existing solutions from effectively optimizing task scheduling strategies and adapting to load fluctuations or resource constraints, thus reducing system flexibility and resource utilization efficiency, this invention provides a smart terminal interaction task offloading and resource scheduling method based on edge-side state awareness.

[0006] To achieve the above objectives, this invention employs a method for intelligent terminal interaction task offloading and resource scheduling based on edge-side state awareness, comprising the following steps: S1: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, a sliding window is used for smoothing to form a set of end-side status indicators; S2: Based on the set of terminal state indicators, collect the maximum tolerable latency and data volume of the interactive task, and use a weighted adaptation calculation model to normalize and jointly map the task characteristics and terminal state, and output the adaptation index. S3: Call the adaptation index, score the end-side status according to the preset mapping rules of computing power utilization and communication latency, and perform weighted fusion with the adaptation score and sort in descending order to generate a priority sorting table; S4: Call the priority sorting table, select the execution path at the top, judge the resource matching conditions, analyze the computing power capacity and latency indicators of the path, and construct a task execution strategy scheme. S5: Based on the task execution strategy, call the computing power and bandwidth resources of the corresponding execution path, dynamically adjust the resource allocation, monitor the task execution latency in real time, input the closed-loop feedback adjustment algorithm, adjust the adaptation weight parameters, and generate resource scheduling instructions.

[0007] As a further aspect of the present invention, the set of end-side status indicators includes processor utilization statistics, graphics processor utilization statistics, and communication latency stability indicators; the adaptability indicators include task latency sensitivity, resource utilization matching degree, and task data volume weight; the priority ranking table includes path score value, path ranking number, and scheduling factor value; the task execution strategy scheme includes path selection rule set, resource allocation threshold, and latency requirement range; and the resource scheduling instructions include resource allocation instruction set, execution monitoring indicators, and feedback adjustment parameters.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, the continuous sampled values ​​of the data frame are processed by sliding window smoothing, compared with the smoothing threshold, and sampled values ​​less than the threshold are aggregated to generate CPU and GPU smoothing sequence sets. S102: Call the CPU and GPU smooth sequence set, retrieve the sampled values ​​based on the two types of communication delay data frames, compare them with the communication delay filtering threshold, filter the sampled values ​​below the threshold, perform serialization aggregation, and align them with the CPU and GPU smooth sequence set to obtain the multi-source synchronization index sequence; S103: Based on the multi-source synchronization index sequence, for the CPU utilization, GPU utilization and two types of communication delay sample values ​​in the sequence, compare them with the synchronization-related benchmark value according to the sample value distribution, and aggregate the sample values ​​that are less than the benchmark value to establish a set of end-side status indicators.

[0009] As a further aspect of the present invention, the smoothing threshold is the upper limit of the load corresponding to the thermal design power consumption of the device under non-full load conditions, which is determined by adding the fluctuation margin to the original average utilization rate of the device when it is running under standard test load at 25℃±2℃. The communication latency filtering threshold is set based on the service quality requirements of real-time interactive applications. The synchronization and association benchmark value is dynamically determined based on the segmented intervals of the terminal's current remaining battery power.

[0010] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Based on the set of terminal status indicators, collect the maximum tolerable latency and data volume of the interactive task, compare the maximum tolerable latency with the latency-related sample values ​​in the set of terminal status indicators, and perform a normalization operation based on the difference sequence to obtain the task normalized load vector. S202: Call the normalized load vector of the task, and based on the performance parameter set extracted from the end-side status index set, match the normalized delay term with the corresponding sample value of the performance parameter set item by item, perform point-to-point operation and mapping to obtain the characteristic mapping parameter set; S203: Call the feature mapping parameter set, perform weighted operation on the feature mapping parameter set and the corresponding performance sample value in the end-side state index set based on the weighted adaptation calculation model, and perform weighted summation based on the preset weight configuration set to obtain the adaptation index.

[0011] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Call the fitness index, perform interval retrieval and match mapping entries for the computing power utilization parameter, perform boundary pruning and discretization processing on the matching results, aggregate and normalize the discretized data by time slice to form a sequence, and generate a computing power quantized sequence. S302: Call the computational quantization sequence, perform segmented retrieval and align mapping entries according to the communication delay parameters, construct a communication delay quantization sequence, perform item-by-item proportional conversion on the two types of quantization sequences according to the weight matrix, and merge the corresponding items after conversion into vectors to obtain a weighted fusion sequence; S303: Based on the weighted fusion sequence, perform value range verification and unify the units for the quantized values, remove abnormal indexes from the verified data and calculate the sorting key, re-index the sorting key and arrange it in structured descending order to generate a priority sorting table.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Based on the priority sorting table, retrieve the computing power capacity parameter carried by the first path and compare it with the computing capacity threshold. If it is greater than the threshold, compare the latency index parameter with the transmission latency threshold. If the latency index parameter is less than the threshold, aggregate the path parameters to generate a computing power latency parameter set. S402: Call the computing power latency parameter set, perform a ratio calculation on the computing power capacity parameter and the latency index parameter, compare the ratio with the path optimization benchmark value, record the path index that is greater than the path optimization benchmark value, and obtain the path performance quantification value. S403: Call the path performance quantification value, compare the path performance quantification value with the strategy construction benchmark value, and if it is greater than the strategy construction benchmark value, perform a combination and arrangement of the computing power capacity parameter, latency index parameter and path index to generate a task execution strategy scheme.

[0013] As a further embodiment of the present invention, the computational capacity threshold is statically calibrated based on the number of floating-point operations required by the current interactive task and the processing capability of the device. The transmission delay threshold is a value dynamically set according to the network environment where the path is located. If the delay index parameter is less than the transmission delay threshold, the path parameters are aggregated to generate a set of computing power delay parameters that includes computing power capacity parameters and delay index parameters. The path optimization benchmark value is calculated based on the average original path performance of successfully unloaded tasks. When the ratio between the computing power capacity parameter and the latency index parameter carried by the path is greater than the path optimization benchmark value, the path index is recorded and the corresponding path performance quantification value is generated. The strategy construction benchmark value is a dynamic threshold set according to the task type and network conditions. When the path performance quantification value is greater than the strategy construction benchmark value, the computing power capacity parameter and the latency index parameter are combined and arranged to generate a task execution strategy scheme.

[0014] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: According to the task execution strategy scheme, call the computing power parameter set and bandwidth parameter set of the execution path, obtain the occupancy value, filter items higher than the benchmark value based on the resource allocation benchmark value, perform sorting and aggregation operations, and obtain the computing power and bandwidth occupancy matrix. S502: Based on the computing power and bandwidth occupancy matrix, perform weighted operations on the computing power occupancy column and the bandwidth occupancy column, monitor the task execution latency value in real time, normalize the task execution latency value, perform differential calculation with the weighted result, and feed back to adjust the adaptation weight parameters to generate an adaptation offset parameter set. S503: Call the adaptation offset parameter set, compare the offset parameter with the preset adjustment threshold, filter the indexes that are lower than the adjustment threshold, perform encoding mapping and organize the results to obtain the resource scheduling instruction.

[0015] As a further aspect of the present invention, the resource allocation benchmark value is obtained by weighted average calculation of the original computing power and bandwidth usage records and real-time task demand data. The preset adjustment threshold is a threshold that is dynamically adjusted based on resource usage and task execution status.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by real-time collection of the terminal's CPU and GPU utilization, as well as end-to-edge and end-to-cloud communication latency, and combined with sliding window smoothing processing, task scheduling becomes more precise. The maximum tolerable latency and data size of a task are matched with the terminal's state using a weighted adaptation model to generate a priority ranking, ensuring more efficient resource scheduling. Furthermore, resource allocation is dynamically adjusted based on real-time feedback, matching task execution paths and network latency, avoiding resource waste and task delays caused by static rules, improving performance and resource utilization, and solving the problems of lag and inaccuracy in task scheduling in traditional solutions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0021] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0022] In this embodiment of the invention, sometimes the subscript such as W1 is written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0023] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0024] Please see Figure 1 This invention provides a method for offloading and scheduling interactive tasks and resources on intelligent terminals based on terminal-side state awareness, comprising the following steps: S1: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, a sliding window is used for smoothing to form a set of end-side status indicators; S2: Based on the set of edge-side state indicators, collect the maximum tolerable latency and data volume of the interactive task, and use a weighted adaptation calculation model to normalize and jointly map the task characteristics and edge-side state, and output the adaptation index. S3: Call the adaptability index, score the end-side status according to the preset mapping rules of computing power utilization and communication latency, and weight and merge it with the adaptability to generate a priority ranking table; S4: Sort the priority sorting table in descending order, select the execution path at the top, determine the resource matching conditions, analyze the computing capacity and latency of the path, and construct a task execution strategy. S5: Based on the task execution strategy, call upon the computing power and bandwidth resources of the corresponding execution path, dynamically adjust resource allocation, monitor task execution latency in real time, input closed-loop feedback adjustment algorithm, adjust the adaptation weight parameters, and generate resource scheduling instructions.

[0025] The set of edge-side status indicators includes processor utilization statistics, graphics processor utilization statistics, and communication latency stability indicators. Adaptability indicators include task latency sensitivity, resource utilization matching degree, and task data volume weight. The priority ranking table includes path score value, path ranking number, and scheduling factor value. The task execution strategy scheme includes path selection rule set, resource allocation threshold, and latency requirement range. The resource scheduling instructions include resource allocation instruction set, execution monitoring indicators, and feedback adjustment parameters.

[0026] Please see Figure 2 The specific steps of S1 are as follows: S101: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, the continuous sampled values ​​of the data frame are processed by sliding window smoothing, compared with the smoothing threshold, and sampled values ​​less than the threshold are aggregated to generate CPU and GPU smoothing sequence sets. First, the built-in hardware performance monitoring interface of the smart terminal is activated. For the actual operating scenarios of high-computing-power-demand devices such as industrial augmented reality (AR) inspection glasses, the sampling frequency at the core level is set to trigger a read operation every 20 milliseconds, and the data is read separately. In the file , , , Fields for calculation Instantaneous utilization, while also provided by the graphics card driver API Acquisition The core load percentage, for the communication link, sends 64-byte data to the preset edge computing nodes and central cloud server. probe packets, record sending timestamps With the received response timestamp ,pass The round-trip time is calculated, and an original sampling vector containing four dimensions of data is constructed. Fifty consecutively collected sampling vectors are stored in a first-in-first-out (FIFO) circular buffer queue. For the data in the queue... Utilization data to define the sliding window size The weight is set to 5, and a weighted moving average algorithm is used, with the weight coefficient of the latest data within the window set. It is 0.4, the latest data. The value is 0.3, and the remaining data decrease sequentially, according to... The logic calculates the smooth value of the current window to eliminate glitches and noise caused by sudden scheduling of background processes. Similarly, it performs the same calculation on the smooth value of the current window. Both utilization and communication latency data undergo the same smoothing process, with a set smoothing threshold. The setting reference value is the upper limit of the load corresponding to the thermal design power (TDP) of the device under non-full load conditions, for example, setting... The threshold is 75%, calculated by adding a 15% fluctuation margin to the historical average utilization rate of the equipment under standard test load at 25℃±2℃. Smoothing value If the percentage is 72%, then the judgment will be executed. If the result is true, the sampled value is retained; otherwise, if the smoothing value is 78%, it is discarded. A smoothing threshold for edge communication latency is set for communication latency. The smoothing threshold for cloud communication latency is 30ms. The time delay data after smoothing is compared one by one with a time interval of 100ms.

[0027] Table 1: Data Table of Intelligent Terminal Resource Status Acquisition and Smoothing Processing Table 1 lists some of the collected data and processing results. Smoothed sample values ​​that meet the condition of being less than a set threshold are extracted and re-push into a new dynamic array according to the timestamp order to form a cleaned resource availability sequence, generating CPU and GPU smoothed sequence sets. S102: Call the CPU and GPU smooth sequence sets, retrieve the sampled values ​​based on the two types of communication delay data frames, compare them with the communication delay filtering threshold, filter the sampled values ​​below the threshold, perform serialization aggregation, and align them with the CPU and GPU smooth sequence sets to obtain the multi-source synchronization index sequence. Read the processed sequence of computing resources stored in memory; this sequence now contains a series of timestamps. and Load data is filtered based on network communication status, and a communication latency filtering threshold is defined. This threshold is set based on the minimum Quality of Service (QoS) requirements for real-time interactive applications, thus setting the edge-side filtering threshold. The cloud-side filtering threshold is 25ms. Given 80ms, iterate through the smoothed delay data frames for a given moment. Edge delay sampling value Perform comparison operations ,like If the time is 22ms, then the condition is met and the data is added to the aggregation list. If the latency is 28ms, it is discarded. The same operation is performed on the cloud latency data, and the filtered latency data is sorted according to timestamp. Sort the data in ascending order, perform serialization aggregation operations, and build a hash map (HashMap) with timestamps as keys. The value is a structure object containing edge latency and cloud latency. Start the position alignment procedure and set the time alignment tolerance window. Iterate for 5ms. and Extract the timestamp from each element in the smoothed sequence set. Search the hash map table of latency data to see if the key-value pair exists. satisfy If network data that meets the conditions exists, the computational resource data and the network resource data will be merged to form a complete multidimensional state vector. If no corresponding network data is found within the tolerance range, it indicates that the network state at that moment is unstable or has been filtered. The computational resource data for that moment is then discarded to ensure the integrity of the data dimensions. For example, at time [time value missing]... Place, The utilization rate is 45%, but the corresponding network latency is rejected because it exceeds the threshold. Therefore, at this point in time... The data does not participate in subsequent steps. Through this bidirectional matching and intersection operation, the dispersed heterogeneous data sources are strictly locked in the time dimension to obtain a multi-source synchronization indicator sequence.

[0028] S103: Based on the multi-source synchronization index sequence, for the CPU utilization, GPU utilization and two types of communication delay sample values ​​in the sequence, compare them with the synchronization correlation benchmark value according to the sample value distribution, and aggregate the sample values ​​that are less than the benchmark value to establish a set of end-side status indicators. Load the aligned multidimensional state vector list from memory, perform deep distribution analysis on each quadruple in the list, calculate the dispersion of the indicators in the sequence, and calculate the variance using data from the 10 most recent synchronization cycles. Set stability criteria, for example The variance threshold for utilization should be less than 5 to exclude non-steady-state data with acceptable mean but significant fluctuations. A synchronous correlation benchmark value is defined. This benchmark value is a dynamically adjusted four-dimensional vector. Its setup process depends on the current remaining battery power of the terminal. Define a piecewise power function, when At that time, the relaxed benchmark value is ,when At that time, the tightening benchmark value is Assuming the current battery level is 90%, using a relaxed baseline value, the vector in the current sequence... Perform item-by-item comparisons and execute logical operations. If all four comparison results are true, the state vector is determined to meet the task offloading condition under high battery conditions. If any dimension of the vector exceeds the corresponding baseline value, for example... If the utilization rate is 75%, it is determined that the edge state at that moment does not meet the task execution requirements, and the vector will be verified through a full-dimensional comparison. The data is stored in the final result container. Each element in the container represents the terminal's complete, stable computing and communication capabilities at a certain moment, which meet the current energy efficiency strategy. The entire sequence is traversed until the data processing is completed, and a set of terminal status indicators is established.

[0029] Please see Figure 3 The specific steps of S2 are as follows: S201: Based on the set of edge-side status indicators, collect the maximum tolerable latency and data volume of the interactive task, compare the maximum tolerable latency with the latency-related sample values ​​in the set of edge-side status indicators, and perform a normalization operation based on the difference sequence to obtain the normalized load vector of the task. First, the task descriptor of the current task to be processed is read. This descriptor clearly defines the task type as a high-precision image recognition task. From this, the maximum tolerable latency of the interactive task is extracted. The settings are 150ms and the size of the data to be transmitted and processed. The memory size is 2.5MB. The system retrieves the edge communication latency sample value at the current moment from the set of edge-side status indicators stored in memory. Sampling value of communication latency with the cloud The latency correlation comparison logic is initiated, and the difference between the maximum tolerable latency and the measured latency is calculated to obtain the latency margin on the edge side. Cloud-side latency margin Normalization is performed on the difference sequence using the inverse load mapping method, and the delay-normalized load formula is defined as follows: The advantage of this formula lies in converting the time delay margin into a dimensionless load pressure value. This represents the latency margin of the corresponding node, with a constant of 1 representing the full load threshold. Subtracting the margin percentage directly reflects the latency urgency of the current communication link for a specific task. Substituting the specific values ​​into the calculation yields the normalized load on the edge side. Cloud-side normalized load Meanwhile, regarding the scale of data volume Normalization is performed to set a standard processing capacity benchmark for a single task. 10MB, compute data load The calculation results of the above three dimensions are arranged in order to obtain the task normalized load vector.

[0030] S202: Call the task normalized load vector, based on the performance parameter set extracted from the end-side status index set, match the normalized delay term with the corresponding sample value of the performance parameter set item by item, perform point-to-point operation and mapping, and obtain the characteristic mapping parameter set; Call the task normalized load vector Extract the current device's performance parameter set, including CPU idle rate, from the edge-side status indicator set. and GPU idle rate Assuming that it is currently detected (i.e., 65% idle) (i.e., 45% idle), while simultaneously acquiring the currently available uplink bandwidth utilization margin. (i.e., 80% availability), perform item-by-item allocation normalization of the corresponding sampled values ​​of the latency term and performance parameter set, and define the resource supply and demand matching calculation formula as follows: ,in For the first Performance indicators For the corresponding load vector elements, As the adjustment coefficient, set This direct mapping formula quantifies the degree to which resources meet task requirements by calculating the algebraic difference between available resources and task load. A positive result indicates resource surplus, while a negative result indicates resource deficit. Point-to-point computation is performed, and for edge computing scenarios, the ratio of CPU performance to edge latency load is calculated. For cloud computing scenarios, considering that cloud tasks often involve graphics rendering or deep learning inference, the GPU performance is matched with the cloud latency load, and the computation... For data transmission, the bandwidth margin is matched with the data load, and calculations are performed. The negative result This indicates that the current cloud processing path has bottlenecks in GPU resources or latency constraints, and does not meet the ideal execution conditions of the task. The above-calculated mapping values ​​are encapsulated to obtain a feature mapping parameter set.

[0031] S203: Call the feature mapping parameter set, perform a weighted operation on the feature mapping parameter set and the corresponding performance sample value in the end-side status index set based on the weighted adaptation calculation model, and perform a weighted summation based on the preset weight configuration set to obtain the adaptation index; Invoke Feature Mapping Parameter Set Based on the current task attributes, a weighted adaptation calculation model is configured. The task attribute is identified as "computation-intensive - real-time rendering type". It is necessary to focus on GPU processing power and data transmission efficiency. The preset weight configuration table is queried to set the edge computing weight. Cloud-based weight calculation Data transmission weight And satisfy The characteristic mapping parameter set and the corresponding performance sample values ​​in the edge-side state index set are weighted and the fitness calculation formula is defined as follows: ,in Let be the penalty function, when hour To amplify the negative impact of resource scarcity, when hour The advantage of this formula lies in its introduction of a non-linear penalty mechanism, which reduces the overall score of schemes with resource bottlenecks (i.e., negative mapping values) and avoids the risk of masking local shortcomings by relying solely on the mean. Substituting the parameters into the formula calculation, the contribution of the marginal terms is... Cloud-based contributions The contribution of the transmission item is Perform weighted summation Table 2 shows examples of weight configuration and calculation results for different task types. By comparing the adaptability indicators of different scheduling paths, the task unloading strategy is finally determined and the adaptability indicators are obtained.

[0032] Table 2: Task Type Suitability Weight Configuration and Calculation Results As shown in Table 2, the comparison reveals that, under the current resource conditions, the overall adaptability of real-time rendering tasks is low, mainly due to the negative mapping of cloud GPU resources. In contrast, text processing tasks, which focus on edge CPU resources (weight 0.6), have a significantly higher overall adaptability than rendering tasks. This result indicates that the current state is more suitable for executing local text processing or storage tasks with high bandwidth requirements, rather than for offloading high-load rendering tasks to the cloud.

[0033] Please see Figure 4 The specific steps of S3 are as follows: S301: Call the adaptability index, perform interval retrieval and match mapping entries for the computing power utilization parameter, perform boundary pruning and discretization processing on the matching results, aggregate and normalize the discretized data by time slice to form a sequence, and generate a computing power quantized sequence. The system reads the adaptation score data of the candidate computing node set for the current industrial AR inspection task from the cache in memory. This set includes the computing adaptation values ​​of five candidate objects, including the currently connected edge node, neighboring edge nodes, and the central cloud node. The first element Based on the calculated current edge node adaptability, perform interval retrieval and match mapping entries for the computing power utilization parameter, set up a computing power scoring mapping table, and define intervals. Mapping level is 1, interval The mapping level is 2, and so on up to Mapping level is 5, traversal For each element in, for element ,judge The original level of the match is 1 for the element. ,judge The original level of 5 is obtained through matching. Boundary clipping and discretization are performed on the matching results, and a clipping function is introduced. To prevent outlier overflow, all values ​​here are within the legal range, and then the discretization enhancement formula is applied. ,in The up-rounding symbol indicates that this formula is beneficial because it converts consecutive decimal scores into higher-granularity integer step values ​​to amplify performance differences between nodes. Transform into ,Will Transform into Similarly, discrete sequences are obtained. Discretized data is aggregated and normalized by time slice to form a sequence, and a normalization reference upper limit is set. , apply formula Calculations yielded , Finally, a standardized computing power feature vector is constructed, generating a computing power quantification sequence.

[0034] S302: Call the computational power quantization sequence, perform segmented retrieval and align mapping entries according to the communication delay parameters, construct the communication delay quantization sequence, perform item-by-item proportional conversion on the two types of quantization sequences according to the weight matrix, and merge the corresponding items after conversion into vectors to obtain the weighted fusion sequence; Calling computational power sequence Simultaneously, read the real-time communication latency data corresponding to these 5 candidate nodes. Based on the communication delay parameters, segmented retrieval and alignment of mapping entries are performed, and a delay utility function is defined. ,in With a baseline delay set to 50ms, this formula utilizes the non-linear decay characteristic of the Sigmoid function to rapidly reduce the weighting of delays exceeding the baseline value. Substitute into the calculation, Similarly, calculate hour, Construct communication delay quantization sequence The two types of quantized sequences are proportionally converted item by item according to the weight matrix, and the corresponding items after conversion are vector-merged. A resource preference weight matrix is ​​set. Based on the current AR task's need for a balance between real-time performance and computing power, the following settings are made: (Computing power weight) (Network weights), the weighted fusion formula is defined as follows: This formula uses a linear weighted superposition mechanism to comprehensively evaluate the computing and transmission capabilities of nodes under a unified dimension, and calculates the results by substituting the first set of data. Substitute the third set of data into the calculation. The node scores are calculated sequentially to obtain the weighted fusion sequence.

[0035] S303: Based on the weighted fusion sequence, perform value range verification and unify the units for the quantized values, remove abnormal indexes from the verified data and calculate the sort key, re-index the sort key and structure the sort position list to generate a priority sort table. Based on weighted fusion sequence Perform range verification and unify the units for the quantized values, and traverse the sequence to check for existence. Elements that are valid in the current sequence are forcibly set to zero if they exist. Abnormal indices are removed from the validated data, and the sorting key is calculated. A minimum availability threshold is set. Perform the filtering operation and find If the current edge node (the one currently connected) is deemed to have insufficient overall performance, it is marked as "unavailable" or removed from the index. The remaining four valid nodes are retained. The sorting key is then indexed and rearranged into a structured list. The remaining scores are then sorted in descending order using a quicksort algorithm. The sorting process is as follows: Based on the ranking results, a structured form containing node ID, sub-scores, and comprehensive scores is constructed, as shown in Table 3. Table 3 details the priority information of the candidate nodes after filtering and ranking. The results show that although Node_3 (i.e., the central cloud node) has an extremely low latency score (0.047), it still ranks first after weighting due to its extremely high computing power score (0.9), making it suitable for handling non-real-time but high-computing-power rendering tasks. Node_4, as a neighboring edge node, follows closely behind with its balanced computing power and excellent latency performance, generating a priority ranking table.

[0036] Table 3: Priority Ranking of Candidate Nodes for Task Unloading As shown in Table 3, through sorting and analysis of the weighted fusion sequence, the optimal path for task unloading was finally determined to be Node_3. This result shows that under the current weight configuration (computing power weight 0.6), it is more inclined to sacrifice some network latency in exchange for more powerful cloud processing capabilities, thereby maximizing the success rate of task execution.

[0037] Please see Figure 5 The specific steps of S4 are as follows: S401: Based on the priority sorting table, retrieve the computing power capacity parameter carried by the first path and compare it with the computing capacity threshold. If it is greater than the threshold, compare the latency index parameter with the transmission latency threshold. If the latency index parameter is less than the threshold, aggregate the path parameters to generate a computing power latency parameter set. Read the sorting results shown in Table 3 stored in memory, extract the relevant configuration information of the candidate path Node_3 that is at the top, and identify its computing power capacity parameters. Set a preset computing capacity threshold for 90 computing units (CUs). The threshold is 60 CU. This threshold is statically calibrated based on the minimum number of floating-point operations (FLOPs) required for the current interactive task, combined with the device's historical processing capabilities, and the comparison logic is executed accordingly. If the result is true, it indicates that the node has the basic capability to handle the current high-load rendering task. Subsequently, a network environment sniffing mechanism is triggered to read the signal-to-noise ratio (SNR) and packet loss rate data of the current link and dynamically set the transmission delay threshold. If you are currently in a stable Wi-Fi 6 environment, set Extract the real-time latency parameters corresponding to Node_3. Execute comparison logic The path is determined to meet the latency constraint. For paths that pass the screening, the parameter aggregation function is called to calculate the computing capacity. Latency indicators In addition to path ID packaging, link jitter variance is also introduced. As auxiliary parameters, construct standardized feature vectors. If the first path fails the verification, it will fall back to the second path Node_4 in the priority table for the same logic verification until a usable path that meets the dual constraints is found, and a set of computing power latency parameters will be generated.

[0038] S402: Call the computing power latency parameter set, perform a ratio calculation on the computing power capacity parameter and the latency index parameter, compare the ratio with the path optimization benchmark value, record the path index that is greater than the path optimization benchmark value, and obtain the path performance quantification value. Calling the computing power latency parameter set The computational capacity parameter and latency parameter are proportionally calculated, and the path efficiency ratio is defined as follows: ,in For computing power capacity, For actual delay, The computing power-latency conversion factor is set to 1.5. The network jitter penalty factor is set to 0.1. For jitter variance, the advantage of this formula lies in penalizing unstable paths in the network while evaluating the computing power-to-latency ratio through the jitter correction term in the denominator, ensuring the reliability of the selected path in continuous transmission. Substituting specific values ​​into the calculation, the denominator term is... The numerator is Finally obtained Calculate the baseline value for the preferred path This benchmark is based on the average raw path performance of the 50 most recent successful task unloadings. Based on the current network congestion coefficient Through formula The calculation shows that, Perform comparison operations If the result is true, it indicates that the real-time performance of the current path Node_3 is better than the original baseline, meeting the high-performance unloading requirements. The paths that meet the conditions will then be indexed. and its performance value Store the data in a temporary cache. If the calculation result is lower than the baseline value, mark the path as "suboptimal" and record the deviation value to obtain the path performance quantification value.

[0039] S403: Call the path performance quantification value, compare the path performance quantification value with the policy construction benchmark value, and if it is greater than the policy construction benchmark value, perform a combination and arrangement of computing power capacity parameters, latency index parameters and path index to generate a task execution policy scheme. Call path performance quantization value The task type identifier of the current interactive task is read as "Augmented Reality Assisted Repair". The strategy configuration database is queried to obtain the policy construction baseline value for this type of task in the current weak network environment. This baseline value is a dynamic threshold used to determine whether to initiate a complex full uninstallation strategy or downgrade to a lightweight uninstallation, and to execute comparison logic. ,Right now If the condition is met, proceed to the high-priority strategy generation process and check the computing capacity parameters. Latency parameters and path index Execute combination permutation, call resource allocation template, based on Allocate cloud container resources proportionally and set the number of CPU cores. Core, memory allocation ,in accordance with Configure transport protocol parameters and set the TCP congestion window size. A complete instruction set containing target nodes, resource quotas, and transmission protocol configurations is constructed, as shown in Table 4. Table 4 details the differentiated execution strategies generated for different performance quantification values. Finally, it is confirmed that the current "full cloud rendering" mode is adopted, and the strategy object is serialized into JSON format control instructions to generate a task execution strategy scheme.

[0040] Table 4: Comparison Table of Execution Strategies for Smart Terminal Interaction Tasks As shown in Table 4, by comparing the performance quantification value of real-time computing with the dynamic benchmark value, it was determined that Node_3 meets the high standard requirement of full unloading, thus automatically matching strategy S_001. This result shows that the performance of Node_4, which has average performance (although it meets the basic benchmark of 0.7125, it does not reach the advanced benchmark of 0.78 constructed by the strategy), was successfully avoided, ensuring that high computing power tasks obtain the optimal resource matching.

[0041] Please see Figure 6 The specific steps of S5 are as follows: S501: Based on the task execution strategy, call the computing power parameter set and bandwidth parameter set of the execution path to obtain the occupancy value, filter items higher than the benchmark value based on the resource allocation benchmark value, perform sorting and aggregation operations, and obtain the computing power and bandwidth occupancy matrix. First, the generated "full cloud rendering" strategy instruction (strategy ID: S_001) is parsed, and the current execution path is locked as Node_3 (central cloud node). The underlying performance monitoring probe interface is called to read the computing power parameter set and bandwidth parameter set associated with this node in real time. The computing power parameters include the number of GPU stream processors used. and video memory usage Bandwidth parameters include uplink real-time throughput. Assuming the current data collection time The monitoring value is indivual, GB, Mbps, based on the resource allocation benchmark value, filters items that exceed the benchmark value. The formula for calculating the resource allocation benchmark value is defined as follows: ,in This represents the average resource usage over the past 10 minutes. Estimate the requirements for the current task. The historical smoothing coefficient is set to 0.6. This formula dynamically sets the water level through weighted moving average logic to prevent interference from instantaneous fluctuations. Taking video memory resources as an example, it is set... GB, GB, substitute into the calculation to obtain the benchmark value GB, execute the filtering and comparison logic to determine the current occupied value. If this is established, it indicates that the video memory resources are in a high-load, active state; similarly, the bandwidth benchmark is calculated. Mbps, judgment The system is established. Resource items that pass the screening and exceed the baseline value are extracted, sorted, and aggregated. They are arranged from high to low according to their utilization rate (current value / total capacity). Assuming a total GPU capacity of 2048 cores, 8GB of video memory, and 100Mbps bandwidth, the calculated utilization rates are as follows: , , The sorted results are [video memory, GPU, bandwidth]. A vector containing resource type identifiers and standardized usage values ​​is constructed to obtain the computing power bandwidth usage matrix.

[0042] S502: Based on the computing power and bandwidth occupancy matrix, perform weighted operations on the computing power occupancy column and the bandwidth occupancy column, monitor the task execution latency value in real time, normalize the task execution latency value, perform differential calculation with the weighted result, and provide feedback to adjust the adaptation weight parameters to generate an adaptation offset parameter set. According to the computing power bandwidth usage matrix Extract the computing power consumption column representing the computing dimension (the average of the first two items). ) and the bandwidth usage column representing the transmission dimension ( The computing power usage column and bandwidth usage column are weighted and a dynamic load aggregation formula is set. ,in and Substitute the resource sensitivity coefficient of the current rendering task into the numerical calculation. Real-time monitoring of task execution latency values, and obtaining the total processing loop latency of the current frame using the timestamp difference method. The latency is calculated in milliseconds (ms), and a difference is calculated based on the task execution latency value and the weighted result. First, the latency value is normalized, referring to... ms, calculate delay saturation The differential feedback formula is defined as follows: The advantage of this formula lies in its ability to accurately identify efficiency losses in resource scheduling (such as low resource utilization but high latency) by quantifying the deviation between actual latency saturation and theoretical resource load. A positive value indicates the existence of non-resource bottlenecks or scheduling lags. Feedback is used to adjust the fitness weight parameters, applying the weight correction formula. ,in To adjust the rate, based on cloud weights Adjust downwards. The difference between the original weights and the adjusted values Encapsulate and generate a set of adaptation offset parameters.

[0043] S503: Call the adaptation offset parameter set, compare the offset parameter with the preset adjustment threshold, filter the indexes that are lower than the adjustment threshold, perform encoding mapping and organize the results to obtain the resource scheduling instruction; Call the fit offset parameter set Read the current operating status parameters, compare the offset parameter with the preset adjustment threshold, and define the dynamic setting formula for the adjustment threshold as follows: ,in As the baseline tolerance, The overall load factor (taken as 0.6) is calculated as follows. Execute comparison logic to determine the offset value. Note that here you should filter indices "below" the adjustment threshold according to the step description, assuming there is an alternative path with an offset parameter of... ,but If the screening criteria are met, it indicates that the path performance is stable and does not require significant adjustments. For the selected stable index items, encoding mapping is performed and the results are organized. The pre-set resource scheduling instruction mapping table is queried (as shown in Table 5). Based on the node ID, resource allocation level, and stability status, the corresponding hexadecimal opcode is generated. For example, Node_3 is mapped to the segment start 0xA, and resource level Level_1 is mapped to 0x01. The basic instructions are then combined, as shown in Table 5. Table 5 details the scheduling instruction encoding rules and meanings under different resource states. This result shows that the final machine code instruction containing precise resource locking and priority configuration is generated, which can be directly parsed and executed by the underlying scheduler to obtain the resource scheduling instruction.

[0044] Table 5: Resource Scheduling Instruction Encoding Mapping Table As shown in Table 5, by verifying and mapping the offset parameters, this embodiment generates a definite instruction code (such as 0xA10F40) for resource items that meet the stability requirements (offset value is less than the threshold). This result shows that the scheduling filters out paths with excessive fluctuations or that require complex recalculation, and directly issues deterministic execution instructions to the underlying hardware, thereby completing the physical allocation of resources on a microsecond time scale.

[0045] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for offloading interactive tasks and scheduling resources in intelligent terminals based on edge-side state awareness, characterized in that, Includes the following steps: S1: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, a sliding window is used for smoothing to form a set of end-side status indicators; S2: Based on the set of terminal state indicators, collect the maximum tolerable latency and data volume of the interactive task, and use a weighted adaptation calculation model to normalize and jointly map the task characteristics and terminal state, and output the adaptation index. S3: Call the adaptation index, and according to the preset mapping rules of computing power utilization and communication latency, perform adaptation scoring on the terminal state, and perform weighted fusion with the adaptation score and sort in descending order to generate a priority sorting table; S4: Call the priority sorting table, select the execution path at the top, judge the resource matching conditions, analyze the computing power capacity and latency indicators of the path, and construct a task execution strategy scheme. S5: Based on the task execution strategy, call the computing power and bandwidth resources of the corresponding execution path, dynamically adjust the resource allocation, monitor the task execution latency in real time, input the closed-loop feedback adjustment algorithm, adjust the adaptation weight parameters, and generate resource scheduling instructions.

2. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 1, characterized in that, The set of edge-side status indicators includes processor utilization statistics, graphics processor utilization statistics, and communication latency stability indicators. The adaptability indicators include task latency sensitivity, resource utilization matching degree, and task data volume weight. The priority ranking table includes path score value, path ranking number, and scheduling factor value. The task execution strategy scheme includes path selection rule set, resource allocation threshold, and latency requirement range. The resource scheduling instructions include resource allocation instruction set, execution monitoring indicators, and feedback adjustment parameters.

3. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: By collecting the CPU utilization, GPU utilization, end-to-edge node communication latency, and end-to-cloud communication latency of the smart terminal, the continuous sampled values ​​of the data frame are processed by sliding window smoothing, compared with the smoothing threshold, and sampled values ​​less than the threshold are aggregated to generate CPU and GPU smoothing sequence sets. S102: Call the CPU and GPU smooth sequence set, retrieve the sampled values ​​based on the two types of communication delay data frames, compare them with the communication delay filtering threshold, filter the sampled values ​​below the threshold, perform serialization aggregation, and align them with the CPU and GPU smooth sequence set to obtain the multi-source synchronization index sequence; S103: Based on the multi-source synchronization index sequence, for the CPU utilization, GPU utilization and two types of communication delay sample values ​​in the sequence, perform a comparison with the synchronization-related benchmark value according to the sample value distribution, and aggregate the sample values ​​that are less than the benchmark value to establish a set of end-side status indicators.

4. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 3, characterized in that, The smoothing threshold is the upper limit of the load corresponding to the thermal design power consumption of the equipment under non-full load conditions. It is determined by adding the fluctuation margin to the original average utilization rate of the equipment when it is running under standard test load at 25℃±2℃. The communication latency filtering threshold is set based on the service quality requirements of real-time interactive applications. The synchronization and association benchmark value is dynamically determined based on the segmented intervals of the terminal's current remaining battery power.

5. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Based on the set of terminal status indicators, collect the maximum tolerable latency and data volume of the interactive task, compare the maximum tolerable latency with the latency-related sample values ​​in the set of terminal status indicators, and perform a normalization operation based on the difference sequence to obtain the task normalized load vector. S202: Call the normalized load vector of the task, and based on the performance parameter set extracted from the end-side status index set, match the normalized delay term with the corresponding sample value of the performance parameter set item by item, perform point-to-point operation and mapping to obtain the characteristic mapping parameter set; S203: Call the feature mapping parameter set, perform weighted operation on the feature mapping parameter set and the corresponding performance sample value in the end-side state index set based on the weighted adaptation calculation model, and perform weighted summation based on the preset weight configuration set to obtain the adaptation index.

6. The method for intelligent terminal interaction task offloading and resource scheduling based on end-side state awareness according to claim 5, characterized in that, The specific steps for S3 are as follows: S301: Call the fitness index, perform interval retrieval and match mapping entries for the computing power utilization parameter, perform boundary pruning and discretization processing on the matching results, aggregate and normalize the discretized data by time slice to form a sequence, and generate a computing power quantized sequence. S302: Call the computational quantization sequence, perform segmented retrieval and align mapping entries according to the communication delay parameters, construct a communication delay quantization sequence, perform item-by-item proportional conversion on the two types of quantization sequences according to the weight matrix, and merge the corresponding items after conversion into vectors to obtain a weighted fusion sequence; S303: Based on the weighted fusion sequence, perform value range verification and unify the units for the quantized values, remove abnormal indexes from the verified data and calculate the sorting key, re-index the sorting key and arrange it in structured descending order to generate a priority sorting table.

7. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 6, characterized in that, The specific steps of S4 are as follows: S401: Based on the priority sorting table, retrieve the computing power capacity parameter carried by the first path and compare it with the computing capacity threshold. If it is greater than the threshold, compare the latency index parameter with the transmission latency threshold. If the latency index parameter is less than the threshold, aggregate the path parameters to generate a computing power latency parameter set. S402: Call the computing power latency parameter set, perform a ratio calculation on the computing power capacity parameter and the latency index parameter, compare the ratio with the path optimization benchmark value, record the path index that is greater than the path optimization benchmark value, and obtain the path performance quantification value. S403: Call the path performance quantification value, compare the path performance quantification value with the strategy construction benchmark value, and if it is greater than the strategy construction benchmark value, perform a combination and arrangement of the computing power capacity parameter, latency index parameter and path index to generate a task execution strategy scheme.

8. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 7, characterized in that, The computational capacity threshold is statically calibrated based on the number of floating-point operations required for the current interactive task and the device's processing capability. The transmission delay threshold is a value dynamically set according to the network environment where the path is located. If the delay index parameter is less than the transmission delay threshold, the path parameters are aggregated to generate a set of computing power delay parameters that includes computing power capacity parameters and delay index parameters. The path optimization benchmark value is calculated based on the average original path performance of successfully unloaded tasks. When the ratio between the computing power capacity parameter and the latency index parameter carried by the path is greater than the path optimization benchmark value, the path index is recorded and the corresponding path performance quantification value is generated. The strategy construction benchmark value is a dynamic threshold set according to the task type and network conditions. When the path performance quantification value is greater than the strategy construction benchmark value, the computing power capacity parameter and the latency index parameter are combined and arranged to generate a task execution strategy scheme.

9. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 7, characterized in that, The specific steps of S5 are as follows: S501: According to the task execution strategy scheme, call the computing power parameter set and bandwidth parameter set of the execution path, obtain the occupancy value, filter items higher than the benchmark value based on the resource allocation benchmark value, perform sorting and aggregation operations, and obtain the computing power and bandwidth occupancy matrix. S502: Based on the computing power and bandwidth occupancy matrix, perform weighted operations on the computing power occupancy column and the bandwidth occupancy column, monitor the task execution latency value in real time, normalize the task execution latency value, perform differential calculation with the weighted result, and feed back to adjust the adaptation weight parameters to generate an adaptation offset parameter set. S503: Call the adaptation offset parameter set, compare the offset parameter with the preset adjustment threshold, filter the indexes that are lower than the adjustment threshold, perform encoding mapping and organize the results to obtain the resource scheduling instruction.

10. The method for intelligent terminal interaction task offloading and resource scheduling based on terminal-side state awareness according to claim 9, characterized in that, The resource allocation benchmark value is calculated by weighted average of the original computing power and bandwidth usage records and real-time task demand data. The preset adjustment threshold is a threshold that is dynamically adjusted based on resource usage and task execution status.

Images