Method and system for adaptive dynamic allocation of computing resources of intelligent converged terminals

Abstract

The application discloses a method and system for adaptive dynamic allocation of computing resources of intelligent fusion terminals, and belongs to the field of resource allocation. The method comprises the following steps: constructing a multi-dimensional tensor potential field for each intelligent fusion terminal; exchanging the multi-dimensional tensor potential field with one-hop neighbor terminals according to a preset period; calculating the potential field gradient between each one-hop neighbor terminal; obtaining a local task; matching a penetration coefficient vector according to the task type; calculating the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient; selecting the terminal with the maximum matching degree as a candidate target terminal; calculating the benefit and cost of migrating the local task to the candidate target terminal; and triggering the migration decision when the benefit is greater than the sum of the cost and a delay threshold, and then judging and executing the task granularity segmentation. The application solves the technical problem that the prior art lacks effective prediction of the dynamic change trend of the terminal running state, thereby intensifying the thermal safety risk of the terminal.

Description

Technical Field

[0001] This invention relates to the field of resource allocation, and more specifically to a method and system for adaptive dynamic allocation of computing resources for intelligent fusion terminals. Background Technology

[0002] With the rapid development of edge computing technology, intelligent fusion terminals integrating multiple types of heterogeneous computing units have been widely used in scenarios such as industrial edge control and vehicle-road collaboration. Multi-terminal collaborative computing has become the core technical path to improve the overall computing power efficiency of the edge side. However, existing intelligent fusion terminals only focus on the instantaneous computing power load status of the terminal and do not predict the dynamic change trend of the terminal's operating status. This results in obvious lag and blindness in task migration decisions, which can easily lead to the ping-pong effect of frequent migration and return of tasks between terminals. Ultimately, this leads to low utilization of heterogeneous computing power resources and increased risk of terminal thermal overload, making it difficult to meet the computing needs of high real-time requirements and complex and diverse task types in edge scenarios. Summary of the Invention

[0003] This application provides a method and system for adaptive dynamic allocation of computing resources for intelligent converged terminals, aiming to solve the technical problem that the lack of effective prediction of the dynamic change trend of terminal operating status in the existing technology leads to an increase in terminal thermal security risks.

[0004] In view of the above problems, this application provides a method and system for adaptive dynamic allocation of computing resources for intelligent fusion terminals.

[0005] The first aspect disclosed in this application provides a method for adaptive dynamic allocation of computing resources in an intelligent fusion terminal, the method comprising:

[0006] A multidimensional tensor potential field is constructed for each intelligent fusion terminal in the terminal collaboration network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with its one-hop neighbor terminal according to a preset period, and calculates the potential field gradient between itself and each one-hop neighbor terminal. The local task of each intelligent fusion terminal is obtained, and a penetration coefficient vector is matched according to the task type. The migration matching degree between the local task and each one-hop neighbor terminal is calculated in combination with the potential field gradient. The neighbor terminal with the highest matching degree is selected as the candidate target terminal. The benefit value and cost value of migrating the local task to the candidate target terminal are calculated. When the benefit value is greater than the sum of the cost value and the hysteresis threshold, a migration decision is triggered, and the task granularity is judged and executed according to the task structure.

[0007] Another aspect of this application discloses an adaptive dynamic allocation system for computing resources in an intelligent fusion terminal, the system comprising: The calculation module is used to construct a multidimensional tensor potential field for each intelligent fusion terminal in the terminal collaborative network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with its one-hop neighbor terminal according to a preset period and calculates the potential field gradient between itself and each one-hop neighbor terminal. The matching module is used to obtain the local task of each intelligent fusion terminal, match the penetration coefficient vector according to the task type, and calculate the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient. The neighbor terminal with the highest matching degree is selected as the candidate target terminal. The execution module is used to calculate the benefit value and cost value of migrating the local task to the candidate target terminal. When the benefit value is greater than the sum of the cost value and the hysteresis threshold, the migration decision is triggered and the task granularity is judged and executed according to the task structure.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: By constructing a multidimensional tensor potential field that integrates computing power load potential, thermal potential, and task affinity potential, a unified quantitative representation of the terminal's full-dimensional operating status is achieved. Combined with potential field gradient calculation with potential field change rate correction, accurate matching of task migration targets is completed. Furthermore, a benefit and cost hysteresis verification mechanism and a granular segmentation strategy based on task directed dependency graphs are introduced to effectively solve the technical problem that existing technologies lack effective prediction of the dynamic change trend of terminal operating status, which leads to an increase in terminal thermal security risks. This improves the overall resource utilization and task execution efficiency of the terminal cluster, reduces the average task execution latency and terminal thermal security risks, and ensures the efficiency and stability of edge multi-terminal collaborative computing.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the adaptive dynamic allocation method for computing resources of an intelligent fusion terminal is provided for embodiments of this application.

[0011] Figure 2 This application provides a schematic diagram of the structure of an adaptive dynamic allocation system for computing resources of an intelligent fusion terminal.

[0012] Explanation of reference numerals in the attached diagram: Calculation module 11, Matching module 12, Execution module 13. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] The overall concept of the technical solution provided in this application is as follows: This application provides a method and system for adaptive dynamic allocation of computing resources in intelligent fusion terminals. By constructing a multidimensional tensor potential field that integrates computing power load potential, thermal potential, and task affinity potential components to uniformly represent the terminal's full-dimensional operating status, and employing potential field gradient calculation corrected for the potential field change rate, it simultaneously considers static resource differences and dynamic trend prediction. Combined with penetration coefficient vector matching task type, it quantifies personalized task requirements and achieves precise task granularity segmentation matching the computing power ratio at both ends based on the identification of strongly connected components in a directed dependency graph. This improves the overall utilization rate of heterogeneous computing resources and reduces terminal thermal security risks.

[0015] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0016] Example 1, as Figure 1 As shown in the embodiments of this application, an adaptive dynamic allocation method for computing resources of an intelligent fusion terminal is provided, the method comprising: S100: Constructs a multidimensional tensor potential field for each intelligent fusion terminal in the terminal collaboration network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with its one-hop neighbor terminal according to a preset period and calculates the potential field gradient between itself and each one-hop neighbor terminal.

[0017] Specifically, a multidimensional tensor potential field is first constructed for each intelligent converged terminal in the terminal collaboration network. The terminal collaboration network is a distributed network cluster composed of multiple interconnected intelligent converged terminals that can collaboratively complete computing tasks. This multidimensional tensor potential field is a multidimensional structured data carrier used to quantify the terminal's own operating status and task adaptability. For each intelligent converged terminal, which is an edge intelligent device that integrates multiple types of heterogeneous computing units such as CPU, GPU, and NPU and has the capabilities of local data processing, wireless / wired communication, and cross-terminal task collaboration, the task queue length and remaining computing power of the three core computing units of local CPU, GPU, and NPU are collected in real time to calculate the computing power load potential. The computing power load potential is a scalar of the terminal's current computing resource load level. The higher the load and the less remaining computing power, the higher the computing power load potential value.

[0018] Subsequently, the real-time operating temperature of the terminal chip is collected synchronously and compared with the preset chip safe operating temperature threshold to calculate the thermal potential. The thermal potential is a scalar of the terminal chip's thermal safety margin. The closer the real-time chip temperature is to the safety threshold, the higher the thermal potential value. Then, offline benchmark testing is used to obtain the processing speedup ratio of the intelligent fusion terminal for different types of computing tasks. Based on the speedup ratio, multiple task affinity components for different task types are calculated. The task affinity component is a vector component of the terminal's adaptability to specific types of tasks. The higher the processing speedup ratio of the terminal for a certain type of task, the lower the value of the corresponding task affinity component, which means that the terminal has a more significant efficiency advantage in performing this type of task.

[0019] Finally, the computing power load potential, thermal potential, and all task affinity potential components are dimensionally aligned and combined to form the multidimensional tensor potential field of the intelligent fusion terminal at the current moment, thus completing the potential field construction of a single terminal.

[0020] Subsequently, each intelligent fusion terminal will complete the bidirectional exchange of its multidimensional tensor potential field with its one-hop neighbor terminals according to a preset fixed period. The one-hop neighbor terminal refers to the adjacent terminal in the terminal collaboration network that can directly communicate wirelessly or wired with the current terminal without going through other terminals for forwarding, thereby limiting the communication range. Each intelligent fusion terminal broadcasts the latest generated multidimensional tensor potential field to all its one-hop neighbor terminals, and at the same time synchronously receives the corresponding potential field vector broadcast by all one-hop neighbor terminals, thus completing the synchronization of the operating status of all neighbors.

[0021] Subsequently, each intelligent fusion terminal retrieves its own multidimensional tensor potential field vector from the previous cycle stored locally, calculates the potential field change rate between the current time and the previous time, where the potential field change rate is a numerical value representing the degree of drastic change in the terminal's own operating state over time, used to predict the changing trend of the terminal's load and operating state. Based on this, the component level difference between the current terminal and each one-hop neighbor terminal at the current time is calculated, and this difference is used as the basic gradient of the potential field, where the potential field gradient is an indicator representing the difference in operating state and task migration potential between the current terminal and neighbor terminals. The larger the gradient value, the greater the state difference between the two terminals, and the higher the potential benefit space for task cross-terminal migration.

[0022] Finally, the difference between the potential field change rate of the one-hop neighbor terminal and the potential field change rate of the current terminal is calculated. Based on this difference, the aforementioned basic gradient is corrected, and the potential field gradient between the current terminal and each one-hop neighbor terminal is finally obtained.

[0023] S200: Obtain the local task of each intelligent fusion terminal, match the penetration coefficient vector according to the task type, calculate the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient, and select the neighbor terminal with the highest matching degree as the candidate target terminal.

[0024] Specifically, the local tasks of each intelligent converged terminal are first acquired in real time. These local tasks are business tasks received by the intelligent converged terminal during business operation and need to be computed and executed within the terminal collaborative network. They are the objects of computing resource scheduling and have built-in attributes such as a unique task type identifier, clear computing requirements, execution latency constraints, and data dependencies.

[0025] Subsequently, parameter matching is performed based on the task type of the local task. The task type is categorized according to the task's computational characteristics, dependence on heterogeneous computing units, and real-time requirements. These categories include: CPU-intensive general logic computation tasks, GPU / NPU-intensive AI model inference tasks, and real-time control tasks with low latency requirements. Different task types have varying sensitivities to different components of the terminal's multidimensional tensor potential field. Specifically, the matching process involves querying a penetration coefficient table pre-generated through offline calibration using a standard test task set, and obtaining the corresponding penetration coefficient for each task type. The permeability coefficient vector, also known as the sensitivity weight vector, is a multidimensional weight vector whose dimensions are perfectly aligned with the components of the multidimensional tensor potential field. The value of each element in the vector represents the sensitivity of the task to the corresponding potential field component. The higher the value, the more sensitive the task migration decision is to changes in the potential field component. This vector is generated through a standardized offline calibration process. Specifically, for each type of standard test task, the changes in the potential field of the terminal are measured during its execution. Based on the data analysis, the sensitivity of the task to each multidimensional potential field component is determined, and the corresponding permeability coefficient is obtained and summarized into a permeability coefficient table that can be directly queried.

[0026] Subsequently, the adaptation quantification calculation is completed by combining the potential field gradient. The potential field gradient is a multi-dimensional potential field gradient with potential field change rate correction between the current intelligent fusion terminal and each one-hop neighbor terminal. It quantifies the differences in the full-dimensional operating state between the current terminal and the corresponding neighbor terminal and the potential potential energy of task cross-terminal migration. For each one-hop neighbor terminal, the penetration coefficient vector obtained by matching is summed element-wise with the potential field gradient corresponding to that neighbor terminal to obtain the comprehensive migration benefit value of that one-hop neighbor terminal for this local task. Then, the comprehensive migration benefit values ​​of all one-hop neighbor terminals are normalized and mapped to the interval of 0-1 to finally obtain the migration matching degree. The migration matching degree is a normalized value that quantifies the degree of adaptation and comprehensive optimization benefit of the local task migration to the corresponding one-hop neighbor terminal. The higher the value, the stronger the adaptability of the neighbor terminal to the local task and the more significant the comprehensive resource optimization effect after the task migration.

[0027] Finally, all migration matching scores are sorted in descending order, and the neighboring terminal with the highest matching score is selected as the candidate target terminal. The candidate target terminal is the one-hop neighboring terminal with the best ability to undertake the current local task after full screening of the adaptability. It is the target object for task migration benefit and cost verification, migration decision triggering and task granularity segmentation execution. The task migration operation will only be officially triggered after the terminal passes the subsequent benefit and cost threshold verification.

[0028] S300: Calculate the benefit value and cost value of migrating the local task to the candidate target terminal. When the benefit value is greater than the sum of the cost value and the hysteresis threshold, trigger the migration decision and judge and execute the task granularity segmentation according to the task structure.

[0029] Specifically, the local task is first identified as the business computing task to be executed by the current intelligent fusion terminal, which has clear computing requirements, task type identification, and data dependencies. The candidate target terminal is selected through a full-scale sorting of migration matching degree and is the one-hop neighbor terminal with the best adaptability to this local task. The benefit value and cost value corresponding to this task migration behavior are calculated for each. The benefit value is an indicator that quantifies the reduction in the total multidimensional tensor potential field of the current local terminal after the local task is migrated to the candidate target terminal. The specific value is the comprehensive decrease in the computing power load potential and thermal potential of the local terminal after the task is unloaded. It directly reflects the comprehensive positive benefits brought by the task migration to the local terminal, such as the release of computing resources, the relief of load pressure, and the improvement of thermal security margin. The cost value is an indicator that quantifies the additional resource consumption generated during the task migration process. Specifically, it is the sum of communication energy consumption and data packaging and parsing computational overhead generated during the transmission of the execution code and dependent data of the local task from the current local terminal to the candidate target terminal.

[0030] Subsequently, a hysteresis threshold is introduced to verify the rationality of the migration. The hysteresis threshold is a pre-calibrated fixed value buffer threshold, which is used to avoid the ping-pong effect of frequent task migration and relocation caused by small real-time fluctuations in the terminal potential field. It sets a safety verification boundary for the migration decision. The migration decision is officially triggered only when the calculated benefit value is greater than the sum of the cost value and the hysteresis threshold. The migration decision is a scheduling instruction that approves the local task to be unloaded and executed on the candidate target terminal. If the verification condition is not met, the local task will continue to be executed locally on the current terminal and will not initiate any cross-terminal migration operation.

[0031] After the migration decision is officially triggered, the process of granular task segmentation based on the task structure is immediately initiated. The task structure comprises the computational logic splitting rules within the local task and the data dependencies between subtasks, fully represented by a directed dependency graph. Granular task segmentation, based on the task's internal logical structure, divides the local task into independently executable sub-blocks and identifies and filters scheduling operations for sub-blocks that can be migrated across terminals. The specific execution follows the task's dependency logic and the matching rules of the computing capabilities of both ends. First, a directed dependency graph for the local task is constructed, where nodes represent the split task sub-blocks, and directed edges represent the data dependencies between sub-blocks. Then, strongly connected components in the directed dependency graph are identified, and each strongly connected component is marked as an indivisible atomic sub-block—the smallest task execution unit with strong cyclic data dependencies that cannot be split across different terminals and nodes for execution. Splitting such a component would lead to errors in the task execution logic. The overhead of cross-node data interaction may far outweigh the benefits of migration, so it must be executed entirely within a single terminal. Next, based on the ratio of computing power between the current local terminal and the candidate target terminal, the migration ratio is determined. This is the target ratio of the amount of computing power to be unloaded for this task migration to the total computing power of the local task. This is used to match the difference in computing resources between the two ends and maximize the resource utilization efficiency of cross-terminal collaboration. Then, from ordinary sub-blocks other than atomic sub-blocks, several sub-blocks are selected in descending order of sub-block computing power to ensure that the total computing power of the selected sub-blocks is closest to the preset migration ratio of the total computing power of the local task. Finally, these selected sub-blocks are sent to the candidate target terminal as migrateable sub-blocks to complete the unloading and execution. At the same time, an exception judgment rule is set for indivisible tasks: if the proportion of the computing power of the atomic sub-block in the total computing power of the local task exceeds a preset threshold, the local task is marked as an indivisible task, and the overall task migration operation is directly executed.

[0032] Furthermore, in the method provided in the application embodiments, a multidimensional tensor potential field is constructed for each intelligent fusion terminal, including: real-time acquisition of the task queue length and remaining computing power of the local CPU, GPU, and NPU of each intelligent fusion terminal, and calculation of computing power load potential; acquisition of chip temperature of each intelligent fusion terminal and comparison with a preset safety threshold to calculate thermal potential; acquisition of the processing speedup ratio of each intelligent fusion terminal for various tasks through offline benchmark testing, and calculation of multiple task affinity components for different task types; and combination of the computing power load potential, the thermal potential, and the task affinity components into the multidimensional tensor potential field.

[0033] Specifically, the first step is to calculate the computing load potential. The computing load potential is a scalar indicator that quantifies the real-time supply and demand balance of computing power across the three heterogeneous computing units (CPU, GPU, and NPU) on the terminal, as well as the degree of task backlog. Its value is positively correlated with the terminal's computing load pressure, and its range is normalized to the [0,1] interval. A value closer to 1 indicates a more saturated terminal computing load and a weaker ability to handle new tasks. Higher loads and less remaining computing power result in a higher computing load potential value. During the calculation, each intelligent fusion terminal collects the task queue length and remaining computing power of its local CPU, GPU, and NPU in real-time according to a preset millisecond-level collection cycle. The task queue length refers to... The total number of tasks submitted and queued for execution in the real-time task scheduling system of the corresponding heterogeneous computing unit represents the total number of tasks waiting to be processed by the unit. A larger value indicates more tasks to be processed and a tighter computing resource. Remaining computing power refers to the peak floating-point computing capacity of the corresponding heterogeneous computing unit that is not occupied and can be used to undertake new computing tasks under the current operating conditions. It is usually measured in trillions of operations per second and is equal to the rated peak computing power of the computing unit minus the currently occupied computing power. It directly represents the unit's idle computing power reserve. A larger value indicates more sufficient computing power redundancy to undertake new tasks. Then, for each type of heterogeneous computing unit, the unit-level load rate is calculated using the formula: Unit load rate = Task queue length / Maximum queue length that the unit can support × Weight 1 + Occupied computing power / Rated peak computing power × Weight 2; Weight 1 and Weight 2 are preset weighting coefficients, and Weight 1 + Weight 2 = 1. They can be calibrated offline according to the terminal application scenario. Then, the unit load rates of CPU, GPU and NPU are weighted and averaged to obtain the overall load rate of the terminal. Finally, the computing power load potential is standardized and calculated by the formula computing power load potential = 1 - overall load rate, and the scalar value of computing power load potential in the range of [0,1] is obtained.

[0034] Subsequently, thermal potential calculations are performed. Thermal potential is a boundary scalar indicator that quantifies the thermal safety status of the terminal's core computing chip and constrains the terminal's task-bearing capacity. The value range is also normalized to the [0,1] interval. The closer the value is to 1, the closer the chip's operating temperature is to the safe upper limit, and the higher the risk of overheating, frequency throttling, or system crash. During the calculation process, the terminal collects the junction temperature data of the CPU, GPU, and NPU core chips in real time through built-in multi-point temperature sensors, and takes the highest value as the terminal's real-time core temperature T. At the same time, the pre-calibrated chip safety temperature threshold is retrieved. This threshold is the maximum junction temperature limit specified by the chip manufacturer for long-term stable operation, and the thermal potential is then calculated using a preset exponential mapping formula: ; Where k is a preset thermal sensitivity coefficient, which can be calibrated through offline testing. = When the thermal potential is 1, far below When the thermal potential approaches 0, a nonlinear quantitative characterization of the terminal's thermal safety status is achieved, adapting to the physical characteristic that the thermal risk increases exponentially as the chip temperature approaches the threshold.

[0035] Next, the task affinity components are calculated. Task affinity is a multi-dimensional vector index that quantifies a terminal's ability to adapt to specific types of computing tasks and its efficiency in accelerating processing. Each component corresponds to a standard task type, and its value range is normalized to the [0,1] interval. The closer the component value is to 1, the higher the terminal's processing efficiency and the stronger its adaptability for that type of task. Before the calculation, an offline benchmark test calibration is completed. That is, a standardized test task set is constructed for all types of standard computing tasks covered in the terminal collaborative network, including AI inference, video encoding and decoding, data encryption, and general numerical calculations. The test task set is run in full on each intelligent fusion terminal, and the average execution latency of each type of task on the corresponding terminal is calculated. At the same time, the average execution latency of this type of task on a benchmark general-purpose processor was calculated. This is used to calculate the terminal's processing speedup ratio for this type of task. The processing speedup ratio refers to the factor by which the terminal's efficiency in executing a specific type of task is improved compared to a benchmark general-purpose processor. The calculation formula is as follows: ; A higher value indicates a stronger acceleration capability for this type of task on the terminal. Then, the processing speedup ratio for all task types was subjected to maximum-min normalization, using the following formula: ; in This represents the maximum speedup ratio for all task types on this terminal. To obtain the minimum value, the task affinity potential components corresponding to each task type and taking values ​​in the range [0,1] are obtained. All components are arranged in the order of preset task types to form the multidimensional vector of task affinity potential for the terminal. Finally, the terminal performs dimensional splicing and tensor encapsulation on the standardized computing load potential scalar, thermal potential scalar, and multidimensional vector of task affinity potential arranged in a fixed dimension order to form a multidimensional tensor potential field with fixed dimensions, all values ​​normalized to the range [0,1], and capable of direct mathematical operations and cross-terminal comparison. Each dimension of this tensor potential field corresponds to a core operating attribute or task adaptation capability of the terminal, realizing a unified tensor representation of the terminal's full-dimensional operating status.

[0036] Furthermore, in the method provided in the application embodiments, each intelligent fusion terminal exchanges its own multidimensional tensor potential field with a one-hop neighbor terminal according to a preset period, and calculates the potential field gradient between itself and each one-hop neighbor terminal. This includes: each intelligent fusion terminal broadcasting its own multidimensional tensor potential field at the current moment to each one-hop neighbor terminal, and receiving the potential field vector of each neighbor; obtaining the tensor potential field vector of itself stored in the previous moment of each intelligent fusion terminal, and calculating the potential field change rate between the current moment and the previous moment; calculating the potential field difference between each intelligent fusion terminal and each one-hop neighbor terminal at the current moment as the basic gradient, and then correcting the basic gradient according to the potential field change rate of itself and its neighbors to obtain the potential field gradient between each intelligent fusion terminal and each one-hop neighbor terminal.

[0037] Specifically, each intelligent fusion terminal first performs bidirectional exchange of its multidimensional tensor potential fields according to a preset period, which is a fixed time interval for potential field information synchronization between terminals in the terminal collaboration network, usually set to the millisecond level. This interval can be optimized offline based on network communication load and terminal computing power fluctuation frequency to balance the real-time performance of information synchronization with network communication overhead. It then communicates directly with one-hop neighbor terminals, which are adjacent terminals in the terminal collaboration network that can communicate directly with the current intelligent fusion terminal without going through other terminals. These are the interaction objects with the lowest information interaction latency and the highest transmission reliability in distributed collaboration, avoiding the transmission latency and data timing distortion caused by multi-hop forwarding. Specifically, at the beginning of each preset period, each intelligent fusion terminal first completes the final calculation and data encapsulation of its own multidimensional tensor potential field at the current time t, which is a full-dimensional standardized tensor vector composed of computing power load potential, thermal potential, and multi-task affinity potential components. This fully represents the terminal's computing power resource status, thermal security boundary, and task adaptability.

[0038] Subsequently, through an adapted wireless or wired communication module, it broadcasts its own multidimensional tensor potential vector at the current time t to all its one-hop neighbor terminals, denoted as . Simultaneously, the current terminal synchronously receives the multidimensional tensor potential field vectors broadcast by all one-hop neighbor terminals within the same preset period, corresponding to the current time t. For each one-hop neighbor terminal with index n, its broadcast potential field vector is denoted as... This ensures that the potential field information of the current terminal and all one-hop neighbor terminals is fully synchronized at the same time, ensuring that the timestamps of the interacting potential field data are completely consistent, and avoiding subsequent gradient calculation errors caused by time sequence misalignment.

[0039] After synchronizing the potential field information, the current terminal immediately calculates the potential field change rate. The potential field change rate is the amplitude and trend of the terminal's own multidimensional tensor potential field during two adjacent synchronization cycles. It can characterize the dynamic change speed of the terminal's operating state and predict the potential field trend of the terminal in the future. The specific process is as follows: The current terminal first retrieves the self-multidimensional tensor potential field vector of the previous time t-τ, where τ is the duration of the preset period, which is pre-saved in the local non-volatile storage. For each dimension component i in the multidimensional tensor potential field, corresponding to the computing power load potential, thermal potential, and the first to Nth task affinity potential components, the formula is used. ; We obtain the rate of change of the potential field in each dimension of the current terminal, and concatenate all the rate of change of the potential field in a fixed dimensional order to form the vector of the rate of change of the potential field in all dimensions of the current terminal. Similarly, for each one-hop neighbor terminal n, the current terminal retrieves the potential field vector broadcast by that neighbor terminal at the previous time t-τ, which is stored locally. Through formula ; The single-dimensional potential field change rate of each dimension of the neighboring terminal is calculated, and then concatenated to form the full-dimensional potential field change rate vector of the neighboring terminal. The positive or negative value of the rate of change represents the trend of the potential field in the corresponding dimension. A positive value indicates that the potential field is rising, the resource tension of the corresponding terminal is increasing, or the task carrying capacity is decreasing. A negative value indicates that the potential field is falling, and the available resource redundancy of the corresponding terminal is increasing. The absolute value of the rate of change represents the speed at which the terminal's operating status changes.

[0040] After calculating the potential field change rate, the current terminal first calculates the basic gradient with each one-hop neighbor terminal. This is the difference vector between the current terminal and each one-hop neighbor terminal at the same time, representing the difference vectors in each dimension of the multidimensional tensor potential field. This directly reflects the static resource state difference between the two terminals at the current time and is the core computational basis for the task migration potential. Specifically, for each one-hop neighbor terminal n, the current terminal calculates the element-wise difference to obtain its own basic gradient vector with that neighbor terminal at the current time t. The formula for calculating the fundamental gradient in a single dimension is: ; The full-dimensional basic gradient vector is formed by concatenating all single-dimensional basic gradients in a fixed order. When the single-dimensional basic gradient is positive, it means that the potential field of the current terminal in that dimension is higher than that of the corresponding neighbor terminal. In other words, the resource status of the current terminal in that dimension is more strained. Migrating the task to the neighbor terminal can improve the running status of that dimension.

[0041] Subsequently, the current terminal corrects the basic gradient based on its own and its neighboring terminals' potential field change rates, obtaining the final potential field gradient. This final gradient vector, after dynamic trend correction, simultaneously reflects the static resource differences between terminals and the future operational status change trends. Its magnitude and direction directly determine the potential benefits and optimal direction of task migration. The larger the absolute value of the gradient, the greater the difference in resource endowments between the two terminals, and the higher the potential benefits of cross-terminal task migration. The correction process is as follows: First, for each one-hop neighboring terminal n, calculate the potential field change rate difference vector between the neighboring terminal and the current terminal. The formula for calculating the difference in the rate of change of a single dimension is: ; This difference vector accurately reflects the difference in the operational status trends between neighboring terminals and the current terminal. If the difference in the rate of change in a single dimension is negative, it means that the potential field of the corresponding neighboring terminal decreases faster than that of the current terminal in that dimension, and its resource advantage in that dimension will further expand in the future. Conversely, it means that the resource advantage of the neighboring terminal will gradually narrow. Subsequently, the basic gradient is corrected based on this difference vector, resulting in the final potential field gradient vector. The calculation formula is: ; Where α is a preset trend correction coefficient with a value range of [0,1], which can be calibrated through offline scenario testing and is used to balance the weight of static resource differences and dynamic trend prediction in gradient calculation. τ is a preset period duration, which is used to convert the difference in the rate of change per unit time into the predicted value of potential field change within a complete synchronization period. Through this correction process, the static resource differences between terminals at the current moment are fully preserved, and the prediction of the future trend of terminal operation status is also incorporated.

[0042] Furthermore, in the method provided in the application embodiment, the basic gradient is further corrected based on the potential field change rate of the terminal itself and its neighbors to obtain the potential field gradient between each intelligent fusion terminal and each one-hop neighbor terminal, including: calculating the difference between the potential field change rate of the one-hop neighbor terminal and the potential field change rate of each intelligent fusion terminal itself; and correcting the basic gradient based on the difference to obtain the potential field gradient.

[0043] Specifically, the difference in potential field change rate between the one-hop neighbor terminal and the local intelligent fusion terminal is calculated first. The potential field change rate is a standardized vector representing the unit-time variation amplitude and trend of each dimension component of the multidimensional tensor potential field of a single intelligent fusion terminal during the synchronization cycle of two adjacent potential fields. Each element in the vector corresponds to the unit-time variation of the computing power load potential, thermal potential, and various task-specific task affinity components in the multidimensional tensor potential field. The sign indicates the direction of change in the corresponding dimension's potential field; a positive value indicates an increase in the corresponding dimension's potential field, increased resource strain in that dimension, and a decrease in the terminal's ability to handle new tasks; a negative value indicates a decrease in the corresponding dimension's potential field, increased redundancy of available resources in that dimension, and enhanced ability to handle new tasks. The absolute value directly reflects the speed of change in the terminal's operating status. A one-hop neighbor terminal refers to... The local intelligent fusion terminal currently performing gradient calculation can communicate directly point-to-point with neighboring terminals without needing to be forwarded by other terminals. It is also a low-latency, high-reliability interactive object defined by the solution, which can avoid the transmission latency and data timing distortion caused by multi-hop forwarding. Its own potential field change rate is the full-dimensional change rate vector calculated by the local terminal based on the current synchronization time t and the previous synchronization time t-τ, where τ is the fixed period of the potential field information synchronization between terminals preset by the solution. It is a quantitative representation of the dynamic change trend of the local terminal's own operating state. The potential field change rate of the one-hop neighbor terminal is the full-dimensional change rate vector calculated by each one-hop neighbor terminal within the same synchronization period t, based on its own multi-dimensional tensor potential field broadcast at time t and time t-τ. This data is synchronized to the local terminal by the neighbor terminal in the periodic potential field broadcast.

[0044] Based on this, the difference in potential field change rate is calculated, which is a standardized vector representing the difference in the multidimensional potential field change trends between a one-hop neighbor terminal and the local terminal within the same synchronization period. It is also the basis for trend correction of the basic gradient. It is obtained by calculating the element-wise difference between the potential field change rate of the one-hop neighbor terminal and the potential field change rate of the local terminal. The sign and absolute value of the value can accurately predict the future trend of the resource endowment difference between terminals: When the difference is negative, it means that the potential field change rate of the neighbor terminal in the corresponding dimension is lower than that of the local terminal. If the potential field of the local terminal in this dimension is on the rise and resources are continuously scarce, while the potential field of the neighbor terminal in this dimension is on the fall and resource redundancy is continuously expanding, the absolute value of the difference will increase synchronously. This means that the resource advantage of the neighbor terminal in this dimension will be further expanded in the next synchronization period than at the current time, and the potential benefit of task migration is higher. When the difference is positive, it means that the potential field of the neighbor terminal in the corresponding dimension is rising faster than that of the local terminal. Its current resource advantage will narrow rapidly in the future or even turn into a resource disadvantage. The basic gradient needs to be corrected downward to avoid incorrect migration decisions.

[0045] Subsequently, the base gradient is corrected based on this difference to obtain the final potential field gradient. The base gradient is the static difference vector of the multidimensional tensor potential field between the local terminal and the one-hop neighbor terminal at the current time t. It directly reflects the static resource difference between the two terminals at the current time and is the benchmark value for gradient correction. When the base gradient is positive, it means that the potential field of the local terminal in this dimension is higher than that of the neighbor terminal. Migrating the task to the neighbor terminal can improve the operating status in this dimension. In the correction process, a trend correction coefficient is also introduced. This is an adjustable coefficient that is pre-calibrated through offline scenarios and used to balance the weight of static resource difference and dynamic trend prediction in the final potential field gradient. The value range is fixed at [0,1] and can be flexibly adjusted. In industrial edge scenarios where the terminal load fluctuates drastically, a higher coefficient value can be set to strengthen the trend prediction weight. In smart scenarios with stable load, a lower coefficient value can be set to prioritize the accuracy of static resource difference. Finally, the corrected potential field gradient is obtained by element-wise superimposing the base gradient and the difference in the potential field change rate after calibration by the trend correction coefficient. This gradient is a full-dimensional standardized vector that simultaneously includes the static resource difference between the terminals at the current time and the future operating status change trend.

[0046] Furthermore, in the method provided in the application embodiment, the migration matching degree between the local task and each one-hop neighbor terminal is calculated by matching the penetration coefficient vector according to the task type and combining the potential field gradient. This includes: querying a preset penetration coefficient table according to the task type to obtain the sensitivity weight vector of the local task to the multi-dimensional potential field components; for each one-hop neighbor terminal, performing element-wise weighted summation of the sensitivity weight vector and the corresponding potential field gradient to obtain the comprehensive migration benefit value of each one-hop neighbor terminal; and normalizing and mapping the comprehensive migration benefit value to obtain the migration matching degree between the local task and each one-hop neighbor terminal.

[0047] Specifically, firstly, the local terminal identifies the task type of the local task to be processed. Task type refers to a standardized classification based on the task's business attributes and computing resource requirements, including but not limited to AI model inference, high-definition video encoding / decoding, data encryption / decryption, general numerical computation, and log aggregation analysis. Different types of tasks have fundamentally different requirements for the terminal's heterogeneous computing power, operating temperature, and specialized processing capabilities. Subsequently, based on the identified task type, the terminal queries a pre-stored penetration coefficient table. The penetration coefficient table is a standardized data table that maps various tasks to the sensitivity weights of corresponding multi-dimensional potential field components, generated offline by running standard test task sets. Its calibration logic is to perform a calibration for each type of standard test task. The test task measures the changes in each potential field component when it is executed on the terminal, analyzes the sensitivity of the task to each multidimensional potential field component, and finally generates a standardized weight set corresponding to each type of task. After querying, the penetration coefficient vector corresponding to the current local task type can be obtained, which is the sensitivity weight vector of the local task to the multidimensional potential field component. The dimension of this vector is completely consistent with the dimensions of the aforementioned multidimensional tensor potential field and potential field gradient. Each weight element in the vector corresponds one-to-one with the computing power load potential, thermal potential, and affinity potential components of each type of task in the multidimensional potential field. The weight value ranges from [0,1]. The higher the weight value, the more sensitive the current local task is to the change of the potential field component, and the higher the priority of the terminal state difference in this dimension in the migration decision.

[0048] After obtaining the sensitivity weight vector of the current task, for each one-hop neighbor terminal, the sensitivity weight vector is summed element-wise with the dynamically corrected potential gradient corresponding to that terminal to obtain the comprehensive migration benefit value corresponding to that one-hop neighbor terminal. The potential gradient is a full-dimensional standardized vector calculated in the previous steps, which simultaneously covers the static resource differences between the local terminal and the neighbor terminal at the current moment and the future operating state change trend. The value of each dimension directly reflects the state improvement that the local terminal can obtain in that dimension after migrating the task to the neighbor terminal. The element-wise weighted summation means that for each corresponding dimension of the vector, the weight value of the sensitivity weight vector is multiplied by the corresponding dimension value of the potential gradient, and then the product results of all dimensions are added together to finally obtain a scalar comprehensive migration benefit value. The physical meaning of this value is the comprehensive improvement benefit that can be obtained by migrating the task to the one-hop neighbor terminal in combination with the personalized sensitivity requirements of the current task. The larger the value, the higher the potential comprehensive benefit of migration.

[0049] After obtaining the comprehensive migration benefit value of all one-hop neighbor terminals, all benefit values ​​are normalized and mapped. That is, through the max-min normalization algorithm, all comprehensive migration benefit values ​​are linearly mapped to the standardized interval [0,1]. Finally, the migration matching degree between the current local task and each one-hop neighbor terminal is obtained. The migration matching degree is a standardized scalar value in the interval 0-1. The higher the value, the higher the degree of matching between the resource requirements and adaptation characteristics of the one-hop neighbor terminal and the current local task, and the stronger the comprehensive benefit and rationality of migrating the task to the terminal.

[0050] Furthermore, in the method provided in the application embodiment, the permeability coefficient table is generated by offline calibration through running a standard test task set: for each type of test task, the change data of each potential field during task execution is measured; based on the change data of each potential field, the sensitivity of each type of test task to the multidimensional potential field components is analyzed to obtain the permeability coefficient corresponding to each type of test task, and the permeability coefficient table is established.

[0051] Specifically, the first step is to construct a standard test task set. This set is a standardized and reproducible collection of test tasks that covers all typical task types and is pre-constructed based on the full-scenario application services of intelligent converged terminals. Each task category strictly corresponds to the high-frequency processing business types in the actual operation of the terminal, including but not limited to AI model inference, high-definition video encoding and decoding, data encryption and decryption, general numerical calculation, and edge data aggregation and analysis. The test tasks under each category have fixed computational load, data throughput, execution process, and operating parameters to ensure that only the core variable of task type is retained during the test, eliminating the interference of task fluctuations on the test results. At the same time, clear performance evaluation indicators are set for each type of test task, including single task execution latency, throughput, execution success rate, and operational stability, providing a quantitative evaluation benchmark for subsequent sensitivity analysis.

[0052] Subsequently, in a hardware environment completely identical to the actual terminal deployment, including the target intelligent fusion terminal's CPU, GPU, NPU full heterogeneous computing units, temperature sensing module, and supporting hardware; and a software environment including the corresponding version of the operating system, chip driver, task scheduling framework, and runtime environment, multiple rounds of repeated testing were conducted for each type of test task. The potential field change data during task execution was measured. The potential field change data refers to the real-time change data of all components corresponding one-to-one with the multi-dimensional tensor potential field dimension of the core solution. Specifically, it includes the real-time dynamic change values ​​of the terminal's computing power load potential, thermal potential, and various task affinity potential components throughout the entire task execution cycle. During the test, two types of data were collected synchronously at a preset high-frequency sampling period. One type is the potential field increment data brought about by task execution, that is, the real-time change of each potential field component compared to the idle baseline state when the terminal executes the test task. The idle baseline state is the stable value of each potential field component when the terminal is not running services. The other type is the test task execution effect index data at the corresponding sampling time. Each type of test task was repeated for no less than a preset number of rounds. After removing abnormal data, the average value was taken as the final effective potential field change data.

[0053] Subsequently, based on the collected potential field change data, the sensitivity of each type of test task to the multidimensional potential field components is analyzed. The sensitivity of multidimensional potential field components refers to the degree to which the execution effect of a single type of test task responds to the change of a specific component in the multidimensional tensor potential field. It is used to quantify the dependence of this type of task on the terminal's operating state in a certain dimension. The higher the sensitivity value, the more significant the effect of a small change in the potential field component will be on the execution effect of this type of task. In other words, the task has higher requirements for the terminal's operating state in that dimension. The analysis process adopts a strict single-variable control method. For each type of test task, all components in the multidimensional tensor potential field except for the target analysis component are fixed to the empty baseline value. Only the value of the target analysis component is adjusted by a single gradient. The test task is repeatedly executed under different gradient component values. The change range of the task execution effect is statistically analyzed. By calculating the ratio of the change rate of the task execution effect to the change rate of the target potential field component, the sensitivity coefficient of this type of test task to the target component is obtained. This process is repeated to complete the calculation of the sensitivity coefficients of all components of the multidimensional tensor potential field for this type of task, forming a full-dimensional sensitivity coefficient set corresponding to this type of task.

[0054] Subsequently, based on the full-dimensional sensitivity coefficient set, the penetration coefficient corresponding to each type of test task is obtained. The penetration coefficient is a sensitivity coefficient after standardization and normalization. Specifically, through the max-min normalization algorithm, the sensitivity coefficients of all dimensions of this type of task are linearly mapped to the standardized interval [0,1] to eliminate the calculation bias caused by the difference in the dimensions of different dimensions. After normalization, the penetration coefficients of all dimensions are arranged in a fixed dimensional order according to the multidimensional tensor potential field, which forms the penetration coefficient vector corresponding to this type of test task, that is, the sensitivity weight vector of the local task to the multidimensional potential field components. The closer the penetration coefficient value is to 1, the higher the sensitivity of the task to the potential field component of that dimension, and the greater the weight of that dimension in the subsequent migration matching degree calculation.

[0055] Finally, all test task categories are mapped one-to-one with their corresponding standardized penetration coefficient vectors to construct a structured, quickly searchable penetration coefficient table. This table will be pre-stored in the local non-volatile storage of each intelligent fusion terminal. When the terminal is running online, it only needs to identify the type of local task to quickly obtain the corresponding penetration coefficient vector by looking up the table, without having to perform complex sensitivity analysis calculations online.

[0056] Furthermore, in the method provided in the application embodiment, the benefit value is the reduction in local potential field that can be achieved by migrating the local task to the candidate target terminal; the cost value is the migration energy consumption of migrating the local task to the candidate target terminal.

[0057] Specifically, firstly, the benefit value is the reduction in local potential field by migrating the local task to the candidate target terminal. The candidate target terminal is the one-hop neighbor terminal with the highest matching degree with the current local task, selected by migration matching degree calculation. It is the sole target evaluation object for task migration. The local potential field is weighted by combining the current local task penetration coefficient vector. The penetration coefficient vector is the comprehensive scalar value of the sensitivity weight of the task to the multidimensional potential field components and the local terminal's multidimensional tensor potential field, obtained by querying the penetration coefficient table. It is used to characterize the comprehensive resource pressure and operational risk borne by the local terminal when executing the task. The higher the value, the greater the comprehensive pressure on the local terminal when executing the task. In the calculation process, based on the offline calibrated baseline data of potential field changes of this type of task, the expected increase of computing power load potential, thermal potential, and each task affinity potential component in the local terminal's multidimensional tensor potential field is calculated when the local task is left to be fully executed on the local terminal. Then, the expected increase of each component is summed element-wise by weighted summation using the penetration coefficient vector to obtain the expected comprehensive potential field increment of the local terminal when executing the task.

[0058] Subsequently, the actual increase in the comprehensive potential field of the local terminal is calculated when the local task is migrated to the candidate target terminal for execution. This increase only includes the minimal increase in potential field generated by the local terminal to complete task splitting, data transmission, and result reception and merging. Finally, the difference between the expected increase in comprehensive potential field and the actual increase in comprehensive potential field is determined as the amount of local potential field that can be reduced by this task migration, which is also the benefit value of this migration. This benefit value is a standardized scalar value. The larger the value, the higher the comprehensive resource pressure and operational risk that this migration alleviates for the local terminal.

[0059] Subsequently, the cost value is the migration energy consumption for migrating the local task to the candidate target terminal. Migration energy consumption refers to the total additional energy consumption incurred by the local terminal and the candidate target terminal during the entire task migration lifecycle, compared to the full execution of the task locally. Specifically, migration energy consumption is broken down into three quantifiable core modules and calculated separately. The first module is data transmission energy consumption, i.e., the energy consumption generated by the communication module during the process of the local terminal transmitting the execution code, input data, and dependency parameters of the task to be migrated to the candidate target terminal, and the candidate target terminal transmitting the task execution results back to the local terminal. This energy consumption is based on the offline calibrated unit data volume transmission energy consumption benchmark value, combined with the estimated total transmission data volume and the real-time signal-to-noise ratio of the communication link. The first module is for precise calculation of transmission rate. The second module is for task processing energy consumption, which is the computing power energy consumption generated by the local terminal performing granular segmentation, data encapsulation, verification and encryption of tasks, and by the candidate target terminal receiving, parsing and verifying task data packets. This energy consumption is calculated based on the offline calibrated unit computing power operation energy consumption benchmark value, combined with the estimated computing power operation volume required for this task splitting and encapsulation. The third module is for collaborative scheduling energy consumption, which is the small amount of communication and computing power energy consumption generated by the local terminal and the candidate target terminal completing migration handshake, state synchronization and execution progress interaction. After summing the energy consumption values ​​of the three modules, they are converted into a scalar value with the same dimension as the benefit value and can be directly compared through a standardized mapping rule unified with the benefit value. This is the cost value of this migration, and finally the standardization of benefit value and cost value is completed.

[0060] Furthermore, in the method provided in the application embodiment, the judgment and execution of triggering migration decision and performing task granular segmentation according to the task structure include: constructing a directed dependency graph of the local task, where nodes are task sub-blocks and directed edges represent data dependencies; identifying strongly connected components in the directed dependency graph and marking each strongly connected component as an indivisible atomic sub-block; determining a migration ratio based on the ratio of computing power between the local terminal and the candidate target terminal; selecting several sub-blocks from ordinary sub-blocks other than atomic sub-blocks in descending order of sub-block computing power, wherein the ratio of the total computing power of the selected sub-blocks to the total task computing power is closest to the migration ratio, and sending the selected sub-blocks as migrateable sub-blocks to the candidate target terminal.

[0061] Specifically, the first step is to construct a directed dependency graph for the local task to be migrated. A directed dependency graph is a structured directed graph data structure used to fully represent the task execution logic and the data dependencies between sub-blocks. The construction process first involves static code analysis and execution flow decomposition to break down the complete local task into several smallest execution units with independent inputs and outputs, namely task sub-blocks. At the same time, the computational amount, input data amount, and output data amount of each task sub-block are accurately counted. Then, the execution order and data dependencies between all task sub-blocks are sorted out. If the execution of a task sub-block depends on the output of another task sub-block, a directed edge is drawn between the two sub-blocks from the preceding sub-block to the following sub-block. Finally, a complete directed dependency graph is formed with task sub-blocks as nodes and data dependencies as directed edges.

[0062] Subsequently, a classic graph theory algorithm for identifying strongly connected components is used to identify strongly connected components in a directed dependency graph. A strongly connected component is a set of subgraphs in a directed dependency graph where any two nodes can be reached from each other via directed edges. All task sub-blocks within this set have bidirectional cyclic data dependencies and must be executed in a closed loop within the same terminal. Forcibly splitting them to different terminals would cause problems such as cyclic data interaction, execution timing disorder, or even task logic collapse. Therefore, each identified strongly connected component is packaged as a whole and marked as an indivisible atomic sub-block. At the same time, the total computation of each atomic sub-block and the proportion of the sum of the computation of all atomic sub-blocks in the total computation of the local task are calculated. If the proportion of the computation of the atomic sub-block exceeds the preset threshold of the scheme, the local task is directly marked as an indivisible task, and the entire task is migrated.

[0063] Subsequently, based on the computing power ratio between the local terminal and the candidate target terminal, the task migration ratio is determined. The candidate target terminal is the one-hop neighbor terminal selected with the highest migration match to the local task and passing the migration benefit-cost threshold verification. The computing power ratio refers to the ratio of the effective available computing power of the local terminal and the candidate target terminal for the current local task type. This computing power value is not simply the chip's peak computing power, but rather a comprehensive calculation combining the remaining computing power in the previously constructed multi-dimensional tensor potential field, task affinity components, and thermal constraints, resulting in effective computing power capable of stably handling the current type of task. The migration ratio is the ratio of the task computing volume that needs to be migrated to the candidate target terminal to the total decomposable computing volume of the local task, calculated using the following formula: Migration ratio = Effective available computing power of candidate target terminals / (Effective available computing power of local terminals + Effective available computing power of candidate target terminals); Finally, from the ordinary sub-blocks outside the atomic sub-blocks that have no strong circular dependencies, they are sorted in descending order of computational cost. A greedy algorithm is used to select the sorted ordinary sub-blocks one by one, continuously accumulating the total computational cost of the selected sub-blocks until the proportion of the total computational cost of the selected sub-blocks to the total computational cost of the local task is closest to the previously determined migration ratio. At this point, the selection stops, and these selected sub-blocks are marked as migrateable sub-blocks. Then, the execution code, input data, dependency information, and execution timing requirements of all migrateable sub-blocks are packaged and encrypted, and sent to the candidate target terminal through a one-hop communication link. The local terminal retains all atomic sub-blocks and unselected ordinary sub-blocks. Subsequently, both ends will strictly follow the execution timing of the directed dependency graph to complete distributed collaborative execution and result aggregation.

[0064] Furthermore, in the method provided in the application embodiment, if the computational amount of the atomic sub-block accounts for more than the proportion of the total computational amount of the local task, the local task is marked as an indivisible task, and the entire task is migrated.

[0065] Specifically, the first step is to perform standardized statistical accounting of the total computational load. This is achieved through static code analysis and execution flow simulation to determine the total computational load of the local task, which represents the total number of floating-point operations required for the full execution of the task. Next, the total computational load of all sub-blocks marked as atomic sub-blocks is calculated. An atomic sub-block is an indivisible minimum execution unit obtained by identifying strongly connected components in the directed dependency graph of the local task. All sub-blocks within an atomic sub-block have bidirectional cyclic data dependencies and must be executed in a closed loop within the same terminal. Distributing them across different terminals would lead to cyclic data interaction, execution timing disorder, or even task logic collapse. The independent computational load of each strongly connected component's corresponding atomic sub-block is summed to obtain the total computational load of the atomic sub-block. Based on this, the following formula is used: Atomic sub-block computation percentage = (Total atomic sub-block computation / Total local task computation) × 100%; The system obtains the percentage of computational cost of atomic sub-blocks and then rigidly compares this percentage with a preset threshold. This preset threshold is a critical value pre-defined through offline business scenarios to determine whether a task is divisible. This threshold is offline-defined based on multiple factors such as edge terminal communication overhead, task execution latency requirements, and distributed collaboration efficiency, and is usually set to 80%. It is the boundary for determining task divisibility. If the comparison result shows that the computational cost of atomic sub-blocks exceeds the preset threshold, it means that most of the computing power overhead of the local task is concentrated in the indivisible, strongly coupled execution unit, and the total computing power of the remaining divisible ordinary sub-blocks without circular dependencies accounts for a very low percentage. There is no economically reasonable space for splitting and migrating. At this time, the system officially marks the local task as an indivisible task. An indivisible task specifically refers to a task with highly coupled core execution logic, a very low percentage of computing power in the divisible independent sub-blocks, and the benefit of reducing computing power load by forcibly splitting and migrating cannot cover the additional overhead of cross-terminal data transmission, timing collaboration, and result merging brought about by splitting.

[0066] Upon completion of the marking, the system immediately terminates the subsequent migration ratio calculation and directly executes the overall task migration. Overall task migration refers to a migration mode where the complete local task is packaged and sent to the candidate target terminal for full execution. Specifically, the complete execution code, all input data, dependency library files, execution parameter configuration, and runtime requirements of the local task are packaged, encrypted, and verified. Through a one-hop direct communication link with the candidate target terminal (the selected neighbor terminal with the highest migration match to the local task and passing the migration benefit-cost threshold verification), the complete task data packet is sent to the candidate target terminal in one go. Simultaneously, a migration handshake and execution status synchronization are completed with the candidate target terminal. After receiving the complete data packet, the candidate target terminal completes verification, parsing, and runtime environment deployment, and executes the task. After execution, the final result is sent back to the local terminal in one go. The local terminal is only responsible for triggering the task migration, monitoring the entire execution status process, and receiving and verifying the final result; it does not participate in the distributed execution process of the task. If the comparison result shows that the computational load of the atomic sub-blocks does not exceed a preset threshold, the task is determined to be divisible, and the system proceeds to the subsequent migration ratio calculation and the segmentation process for selecting migrateable sub-blocks.

[0067] In summary, the adaptive dynamic allocation method for computing resources of the intelligent fusion terminal provided in this application has the following technical effects: By constructing a multidimensional tensor potential field that integrates computing power load potential, thermal potential, and task affinity potential, a unified quantitative representation of the terminal's full-dimensional operating status is achieved. Combined with potential field gradient calculation with potential field change rate correction, accurate matching of task migration targets is completed. Furthermore, a benefit and cost hysteresis verification mechanism and a granular segmentation strategy based on task directed dependency graphs are introduced to effectively solve the technical problem that existing technologies lack effective prediction of the dynamic change trend of terminal operating status, which leads to an increase in terminal thermal security risks. This improves the overall resource utilization and task execution efficiency of the terminal cluster, reduces the average task execution latency and terminal thermal security risks, and ensures the efficiency and stability of edge multi-terminal collaborative computing.

[0068] Example 2, based on the same inventive concept as the adaptive dynamic allocation method of computing resources for the intelligent fusion terminal in the foregoing examples, such as... Figure 2 As shown in the embodiment of this application, an adaptive dynamic allocation system for computing resources of an intelligent fusion terminal is provided. The system includes: The calculation module 11 is used to construct a multidimensional tensor potential field for each intelligent fusion terminal in the terminal collaborative network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with a one-hop neighbor terminal according to a preset period and calculates the potential field gradient between itself and each one-hop neighbor terminal. The matching module 12 is used to obtain the local task of each intelligent fusion terminal, match the penetration coefficient vector according to the task type, calculate the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient, and select the neighbor terminal with the highest matching degree as the candidate target terminal. The execution module 13 is used to calculate the benefit value and cost value of migrating the local task to the candidate target terminal. When the benefit value is greater than the sum of the cost value and the hysteresis threshold, the migration decision is triggered and the task granularity is judged and executed according to the task structure.

[0069] Furthermore, the computing module 11 is also used to perform the following steps: real-time acquisition of the task queue length and remaining computing power of the local CPU, GPU, and NPU of each intelligent fusion terminal, and calculation of computing power load potential; acquisition of chip temperature of each intelligent fusion terminal and comparison with a preset safety threshold, and calculation of thermal potential; acquisition of the processing speedup ratio of each intelligent fusion terminal for various tasks through offline benchmark testing, and calculation of multiple task affinity components for different task types; and combination of the computing power load potential, the thermal potential, and the task affinity components into the multidimensional tensor potential field.

[0070] Furthermore, the calculation module 11 is also used to perform the following steps: each intelligent fusion terminal broadcasts its current multidimensional tensor potential field to each one-hop neighbor terminal and receives the potential field vector of each neighbor; obtains the tensor potential field vector of each intelligent fusion terminal stored in the previous moment and calculates the potential field change rate between the current moment and the previous moment; calculates the potential field difference between each intelligent fusion terminal and each one-hop neighbor terminal at the current moment as the basic gradient, and then corrects the basic gradient according to the potential field change rate of itself and its neighbors to obtain the potential field gradient between each intelligent fusion terminal and each one-hop neighbor terminal.

[0071] Furthermore, the calculation module 11 is also used to perform the following steps: calculate the difference between the potential field change rate of a one-hop neighbor terminal and the potential field change rate of each intelligent fusion terminal itself; and correct the basic gradient according to the difference to obtain the potential field gradient.

[0072] Furthermore, the matching module 12 is also used to perform the following steps: query a preset penetration coefficient table according to the task type to obtain the sensitivity weight vector of the local task to the multidimensional potential field components; for each one-hop neighbor terminal, perform element-wise weighted summation of the sensitivity weight vector and the corresponding potential field gradient to obtain the comprehensive migration benefit value of each one-hop neighbor terminal; normalize and map the comprehensive migration benefit value to obtain the migration matching degree between the local task and each one-hop neighbor terminal.

[0073] Furthermore, the matching module 12 is also used to perform the following steps: for each type of test task, measure the change data of each potential field during task execution; analyze the sensitivity of each type of test task to the multidimensional potential field components based on the change data of each potential field, obtain the permeability coefficient corresponding to each type of test task, and establish the permeability coefficient table.

[0074] Furthermore, the execution module 13 is also used to perform the following steps: the benefit value is the amount of local potential field that can be reduced by migrating the local task to the candidate target terminal; the cost value is the migration energy consumption of migrating the local task to the candidate target terminal.

[0075] Furthermore, the execution module 13 is also used to perform the following steps: constructing a directed dependency graph of the local task, wherein nodes are task sub-blocks and directed edges represent data dependencies; identifying strongly connected components in the directed dependency graph and marking each strongly connected component as an indivisible atomic sub-block; determining a migration ratio based on the ratio of computing power between the local terminal and the candidate target terminal; selecting several sub-blocks from ordinary sub-blocks other than atomic sub-blocks in descending order of sub-block computing power, wherein the ratio of the total computing power of the selected sub-blocks to the total task computing power is closest to the migration ratio, and sending the selected sub-blocks as migrateable sub-blocks to the candidate target terminal.

[0076] Furthermore, the execution module 13 is also used to perform the following steps: if the computational amount of the atomic sub-block accounts for more than the proportion of the total computational amount of the local task, the local task is marked as an indivisible task, and the entire task is migrated.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for adaptive dynamic allocation of computing resources in intelligent converged terminals, characterized in that, include: A multidimensional tensor potential field is constructed for each intelligent fusion terminal in the terminal collaboration network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with a one-hop neighbor terminal according to a preset period and calculates the potential field gradient with each one-hop neighbor terminal. Obtain the local task of each intelligent fusion terminal, match the penetration coefficient vector according to the task type, calculate the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient, and select the neighbor terminal with the highest matching degree as the candidate target terminal. Calculate the benefit and cost of migrating the local task to the candidate target terminal. When the benefit is greater than the sum of the cost and the hysteresis threshold, trigger a migration decision and determine and execute task granularity segmentation based on the task structure. A multidimensional tensor potential field is constructed for each intelligent fusion terminal, including: For each intelligent fusion terminal, the task queue length and remaining computing power of the local CPU, GPU and NPU are collected in real time, and the computing power load potential is calculated. The temperature of the chip in each intelligent fusion terminal is collected and compared with a preset safety threshold to calculate the thermal potential. The speedup ratio of each intelligent fusion terminal for various tasks is obtained through offline benchmark testing, and multiple task affinity components for different task types are calculated. The computing power load potential, the thermal potential, and the task affinity potential components are combined into the multidimensional tensor potential field. Each intelligent fusion terminal exchanges its multidimensional tensor potential field with its one-hop neighbor terminal at a preset period, and calculates the potential field gradient with each one-hop neighbor terminal, including: Each intelligent fusion terminal broadcasts its current multidimensional tensor potential field to each one-hop neighbor terminal and receives the potential field vector of each neighbor. Obtain the tensor potential field vector of each intelligent fusion terminal stored in the previous moment, and calculate the rate of change of the potential field between the current moment and the previous moment; The potential field difference between each intelligent fusion terminal and each one-hop neighbor terminal at the current time is calculated as the basic gradient. Then, the basic gradient is corrected according to the potential field change rate of the terminal itself and its neighbors to obtain the potential field gradient between each intelligent fusion terminal and each one-hop neighbor terminal. The basic gradient is then corrected based on the potential field change rates of the terminal itself and its neighbors to obtain the potential field gradient between each intelligent fusion terminal and each one-hop neighbor terminal, including: Calculate the difference between the potential field change rate of the one-hop neighbor terminal and the potential field change rate of each intelligent fusion terminal itself; The basic gradient is corrected based on the difference to obtain the potential field gradient; Based on the task type, a penetration coefficient vector is matched, and the migration matching degree between the local task and each one-hop neighbor terminal is calculated using the potential field gradient, including: Based on the task type, a preset penetration coefficient table is queried to obtain the sensitivity weight vector of the local task to the multidimensional potential field components. For each one-hop neighbor terminal, the sensitivity weight vector and the corresponding potential field gradient are summed element-wise to obtain the comprehensive migration benefit value of each one-hop neighbor terminal. The comprehensive migration benefit value is normalized and mapped to obtain the migration matching degree between the local task and each one-hop neighbor terminal.

2. The adaptive dynamic allocation method for computing resources of an intelligent fusion terminal as described in claim 1, characterized in that, The permeability coefficient table was generated through offline calibration using a standard test task set. For each type of test task, measure the changes in each potential field during task execution; Based on the analysis of the potential field change data, the sensitivity of each type of test task to the multidimensional potential field components is obtained, the permeability coefficient corresponding to each type of test task is obtained, and the permeability coefficient table is established.

3. The adaptive dynamic allocation method for computing resources of an intelligent fusion terminal as described in claim 1, characterized in that, The benefit value is the amount of local potential that can be reduced by migrating the local task to the candidate target terminal; The cost value is the migration energy consumption of the candidate target terminal for migrating the local task.

4. The adaptive dynamic allocation method for computing resources of an intelligent fusion terminal as described in claim 1, characterized in that, The judgment and execution of triggering migration decisions and task granularity segmentation based on task structure include: Construct a directed dependency graph for the local task, where nodes are task sub-blocks and directed edges represent data dependencies; Identify the strongly connected components in the directed dependency graph and mark each strongly connected component as an indivisible atomic sub-block; The migration ratio is determined based on the ratio of computing power between the local terminal and the candidate target terminal. From ordinary sub-blocks other than atomic sub-blocks, select several sub-blocks in descending order of sub-block computational load. Among them, the proportion of the total computational load of the selected sub-blocks to the total task computational load is closest to the migration ratio. The selected sub-blocks are then sent to the candidate target terminal as migrateable sub-blocks.

5. The adaptive dynamic allocation method for computing resources of an intelligent fusion terminal as described in claim 4, characterized in that, If the computational cost of an atomic sub-block accounts for more than a preset threshold of the total computational cost of the local task, the local task is marked as an indivisible task, and the entire task is migrated.

6. A computing resource adaptive dynamic allocation system for intelligent converged terminals, characterized in that, The system comprises the following steps for implementing the adaptive dynamic allocation method for computing resources of the intelligent fusion terminal according to any one of claims 1 to 5: The calculation module is used to construct a multidimensional tensor potential field for each intelligent fusion terminal in the terminal collaboration network. Each intelligent fusion terminal exchanges its own multidimensional tensor potential field with its one-hop neighbor terminal according to a preset period and calculates the potential field gradient between itself and each one-hop neighbor terminal. The matching module is used to obtain the local task of each intelligent fusion terminal, match the penetration coefficient vector according to the task type, calculate the migration matching degree between the local task and each one-hop neighbor terminal in combination with the potential field gradient, and select the neighbor terminal with the highest matching degree as the candidate target terminal. The execution module is used to calculate the benefit value and cost value of migrating the local task to the candidate target terminal. When the benefit value is greater than the sum of the cost value and the hysteresis threshold, the migration decision is triggered and the task granularity is judged and executed according to the task structure.

Images