Multi-task processing method and system based on large model

Abstract

This invention relates to the field of task scheduling technology, specifically to a multi-task processing method and system based on a large model, comprising the following steps: extracting resource parameters using a large model to generate a conflict intensity matrix, screening resource-overlapping node pairs, determining structural conflicts, constructing a dependency graph to calculate complexity scores and prune dense regions, extracting high-frequency paths to divide scheduling units, and generating a task priority sequence. In this invention, by constructing a conflict intensity matrix and combining node clustering features to identify task resource conflict relationships, the sensitivity and accuracy of task partitioning to resource distribution are improved. Nodes are divided into upper and lower bounds based on median deviation, and structural optimization and reorganization are achieved by combining resource type dissimilarity rates. Path depth and branch factors constitute the basis for complexity scoring. Weights are dynamically adjusted based on occupancy time to identify pruning dense regions. The most frequently called paths constitute the main path set. Non-main path tasks are divided according to node coverage relationships, enhancing the clarity of scheduling boundaries and resource isolation capabilities.

Description

Technical Field

[0001] This invention relates to the field of task scheduling technology, and in particular to a multi-task processing method and system based on a large model. Background Technology

[0002] The field of task scheduling technology encompasses the systematic research and implementation of the execution order of computing tasks, resource allocation, and optimization of the execution process. The core of this technology lies in designing a reasonable task allocation mechanism to address issues such as resource conflicts, increased latency, and decreased system efficiency caused by the parallel execution of multiple tasks. It mainly involves the task assignment strategies of operating systems, dynamic allocation methods of computing resources, and concurrency control in multiprocessor systems. Task scheduling is widely used in multi-core processor environments, distributed computing architectures, and cloud computing platforms. By setting the priority of different tasks, controlling the execution timing, and selecting scheduling algorithms, it effectively supports the stable and efficient operation of computing systems.

[0003] Among them, the multi-task processing method refers to the task organization and scheduling strategy designed to improve the efficiency of computing resource utilization and task execution in a multi-core processing or distributed system environment. This patent topic addresses the problems of resource competition, processing priority determination and dynamic load balancing among multiple tasks. It adopts a task partitioning mechanism based on a preset strategy, constructs a task dependency graph to clarify the execution order, and uses a classification and discrimination model to prioritize tasks based on historical task execution data. During task execution, it dynamically monitors resource usage and adjusts the scheduling strategy to achieve reasonable task allocation among different computing nodes and real-time management during the execution process.

[0004] Traditional multitasking techniques handle resource conflicts through fixed task allocation strategies, rely on static data for task priority division, lack real-time awareness of task running status and behavior, cannot accurately depict resource overlap relationships, and have task structure decomposition based on logical flow rather than resource characteristics. The distribution of nodes has a high degree of stacking, which increases the probability of conflicts. The dependency graph only reflects the static order between tasks and does not have complexity dimension analysis, making it difficult to implement fine-grained optimization for densely structured areas of the path. The frequency of calls is not taken into account in task path identification, causing the main scheduling path to deviate from the actual execution hotspot, resulting in resource scheduling offsets and path scheduling conflicts, and a significant increase in task processing chain latency fluctuations. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a multi-task processing method and system based on a large model. The technical solution is as follows:

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-task processing method based on a large model, comprising the following steps:

[0007] S1: Use a large model to obtain the resource type encoding, access frequency and occupation duration of task nodes, calculate the Euclidean distance of resource demand vectors to obtain the conflict intensity matrix, use a preset threshold to filter resource overlapping node pairs, judge structural conflicts based on node clustering characteristics, call density clustering algorithm to evaluate density, and generate recombinable task node structures.

[0008] S2: Based on the reconfigurable task node structure, divide the upper and lower bound nodes using the conflict intensity median deviation parameter, extract the upper bound nodes to construct independent units, match the lower bound nodes and merge them according to the resource type dissimilarity rate to generate a primary task scheduling unit set.

[0009] S3: Based on the primary task scheduling unit set, extract the resource access channel number and scheduling trigger order, construct a dependency graph and calculate the node path depth and branch factor to obtain a complexity score, adjust the weights in combination with the occupied time, and perform subgraph and node aggregation in dense areas according to the score to generate a dependency subgraph set;

[0010] S4: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the path with the highest call frequency to form the main path set, divide the remaining nodes to form scheduling groups, and generate the task structure partitioning result.

[0011] As a further embodiment of the present invention, the reconfigurable task node structure includes node clustering labels, resource conflict labels, and structure adjustment indication information; the primary task scheduling unit set includes fusion node index, independent unit identifier, and unit structure configuration parameters; the dependency subgraph set includes node dependency relationship graph, task path hierarchy structure, and aggregation subgraph identifier; and the task structure partitioning result includes main path node index, group boundary list, and structure level number.

[0012] As a further aspect of the present invention, the steps of obtaining the resource type code, access frequency, and occupation duration of task nodes using a large model, calculating the Euclidean distance of the resource demand vector to obtain the conflict intensity matrix, filtering resource-overlapping node pairs using a preset threshold, judging structural conflicts based on node clustering characteristics, and evaluating density using a density clustering algorithm to generate a recombinable task node structure are as follows:

[0013] S101: Use a large model to obtain the resource type encoding, access frequency and duration of task nodes, construct and obtain vector representations according to resource dimensions, calculate the distance values ​​of multiple node pairs on each resource dimension, construct a symmetric conflict intensity matrix by combining the distances of each dimension, and generate resource distance intensity matrix values.

[0014] S102: Based on the resource distance strength matrix value, call the resource overlap judgment threshold, filter the task node combinations in the node pair whose resource distance is less than the resource overlap judgment threshold, and perform aggregation and classification according to the resource type encoding information mapped by the node, identify the number of aggregated nodes under each type of resource encoding, calculate the frequency distribution of each group of resource combinations in the task set, and obtain the resource overlap node distribution rate.

[0015] The resource overlap determination threshold is the mean distance of all nodes to the resource minus the standard deviation;

[0016] S103: Call the resource overlap node distribution rate, and based on the access frequency and occupancy time data of each node in the resource combination, call the DBSCAN density clustering algorithm to calculate the local resource access density value, mark the high-density resource structure, construct the connectivity graph and classify and aggregate the high-density structure to obtain the recombinable task node structure.

[0017] The DBSCAN density clustering algorithm calculates the Euclidean distance between two-dimensional vectors composed of the resource access frequency and occupation duration of task nodes, determines the search radius based on the inflection point method of the k-th nearest neighbor distance curve, and sets the minimum number of neighboring nodes by combining the dimension multiple method and the task size adjustment method. It calculates the local density between nodes, identifies high-density resource structures, and aggregates them into recombinable task node structures.

[0018] The local resource access density value is obtained by constructing a two-dimensional vector based on the access frequency and duration of task nodes, setting a search radius and a minimum number of neighboring nodes in the vector space, counting the number of nodes in the neighborhood and analyzing the local density.

[0019] As a further aspect of the present invention, the steps of dividing upper and lower bound nodes according to the reconfigurable task node structure using the conflict intensity median deviation parameter, extracting upper bound nodes to construct independent units, matching lower bound nodes and merging them according to the resource type dissimilarity rate to generate a primary task scheduling unit set are as follows:

[0020] S201: Based on the resource conflict intensity value corresponding to each group of nodes in the reconfigurable task node structure, set the median value of the conflict intensity according to the median deviation parameter, calculate the offset difference between the resource conflict value and the median value of each node, identify the offset direction, extract nodes with positive offset direction as upper bound nodes, extract nodes with negative offset direction as lower bound nodes, record the number of upper and lower bound nodes and the corresponding structural position information, and generate the distribution of upper and lower bound nodes.

[0021] The median deviation parameter is obtained by extracting the resource conflict intensity value of each node, calculating the median, extracting the difference between each node's value and the median, obtaining the offset difference, dividing the offset direction, and identifying the upper and lower bound nodes.

[0022] S202: Based on the distribution of the upper and lower bound nodes, extract the set of upper bound nodes, identify the resource type code and original position information in the structure of each upper bound node, establish independent index relationships between nodes and mark the corresponding resource attribute items in the index path, call the resource usage frequency difference between nodes, calculate the resource isolation ratio, construct the resource independent partition structure of the upper bound nodes according to the isolation ratio, and generate the resource utilization separation index.

[0023] The resource isolation ratio is calculated by taking the difference in access frequency of each node in the upper bound node set for the same resource type, and normalizing the difference to the ratio of the average access frequency. This ratio is used to quantify the heterogeneity of resource usage and the construction of an independent resource partitioning structure.

[0024] S203: Call the index information and node resource code content corresponding to each upper bound node in the resource utilization separation index, identify the lower bound nodes that have direct dependency connections with the upper bound nodes and continuous structure numbers according to the structure number and connected edge relationship in the index path, extract the resource type code of the corresponding lower bound node, calculate the resource type dissimilarity rate according to the number of resource type differences and the total number of resource items, and perform fusion matching and node reorganization on the upper and lower bound nodes according to the dissimilarity rate value to generate a primary task scheduling unit set;

[0025] The resource utilization separation index is obtained by extracting the resource type code and resource usage frequency of the upper bound node, calculating the difference in access frequency between nodes on the same resource type and normalizing it to the isolation ratio, constructing an independent resource partitioning structure, and obtaining the resource utilization separation index.

[0026] The resource type dissimilarity rate is obtained by identifying the proportion of non-overlapping resource types between upper and lower bound nodes to the total number of resource types, and by pairing and merging them according to the node resource code and structure number.

[0027] As a further aspect of the present invention, the specific formula for calculating the resource type dissimilarity rate based on the number of resource type differences and the total number of resource items is as follows:

[0028] ;

[0029] Calculate the resource type dissimilarity rate;

[0030] in, The resource type distinctness rate is expressed in dimensionless form. The total number of resource types defined in the system. For the index number of the resource type, For the index number of the resource type, This is the index number of the lower bound node. This represents the total number of lower bound nodes extracted so far. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first The lower bound node of the nth node The proportion or importance weight of a resource type within the node's resource structure, expressed in dimensionless units. For the first The lower bound node of the nth node The proportion of a resource type in the node's resource structure or the importance weight of a resource type, expressed in dimensionless units.

[0031] As a further aspect of the present invention, the steps of extracting resource access channel numbers and scheduling triggering orders based on the primary task scheduling unit set, constructing a dependency graph and calculating node path depth and branch factor to obtain a complexity score, adjusting weights based on occupied time, and pruning dense regions to execute subgraphs and node aggregation according to the score to generate a dependency subgraph set are as follows:

[0032] S301: Extract the resource access channel number and scheduling trigger order based on the primary task scheduling unit set, extract the directed connection path between nodes according to the trigger sequence, construct the position index matrix of each pair of nodes in the path graph, and set the weight and connection direction of the edges between nodes to generate a task node dependency graph.

[0033] S302: Call the path index and edge connection relationship in the task node dependency graph, identify the number of nodes in each path and use it as the path depth, identify the number of subsequent nodes connected to each node and use it as the branch factor, call the path depth and branch factor to calculate the original complexity score, combine the occupied time information of each node recorded in the scheduling table, and set the score adjustment coefficient according to the occupied time to adjust the original complexity score and obtain the node scheduling complexity score value.

[0034] S303: Based on the node scheduling complexity score, filter the node score data, extract the node indexes with scores greater than the complexity score percentile threshold, and identify the corresponding aggregation distribution segments in the dependency graph. Based on the score difference between adjacent nodes, construct and trim continuous aggregation path intervals in the graph structure. Each node in the aggregation segment is a subgraph merging object, generating a dependency subgraph set.

[0035] The complexity score percentile threshold is constructed by creating a sorted array based on the scheduling complexity scores of all nodes, and then using a grid search method to compare the scheduling performance of each percentile value on multiple task sets, selecting the score value at the 80th percentile as the threshold for filtering dense path nodes.

[0036] As a further aspect of the present invention, the steps of extracting the most frequently invoked paths to form a main path set, dividing the remaining nodes into scheduling groups, and generating the task structure partitioning result based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set are as follows:

[0037] S401: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the node number included in each path, collect the trigger frequency of the corresponding node in the scheduling log, identify the total number of calls for each node in each path, extract the path number with the highest trigger frequency in each path and construct a dataset to generate a high-frequency path trigger sequence.

[0038] The high-frequency path trigger sequence is a set of frequent paths obtained by identifying the total trigger frequency of each node in the scheduling log of each path and selecting the set of path numbers with the highest trigger frequency.

[0039] S402: Call the path number in the high-frequency path trigger sequence and obtain the corresponding all node numbers. Remove the main path node number from the node number set of the primary task scheduling unit set and extract the remaining node number set. Perform edge traversal on the resource dependency relationship between nodes within the remaining node number set. Perform hierarchical clustering analysis on the remaining nodes according to the connection directionality and resource channel number to generate a node scheduling partition structure.

[0040] S403: Construct a path attribution index table based on the remaining nodes in the node scheduling partition structure, construct a group number mapping table using the set of main path nodes in the high-frequency path trigger sequence, merge the path attribution index table and the group number mapping table into a unified number index structure, and map it to the scheduling structure diagram to generate the task structure partitioning result;

[0041] The node scheduling partitioning structure is formed by traversing the resource dependencies between the remaining nodes after removing the main path nodes, and then performing dependency clustering based on the resource channel number and connection directionality to form a grouping structure for task nodes.

[0042] As a further aspect of the present invention, the method further includes step S5:

[0043] S5: Based on the task structure division results, collect the resource utilization rate and processing time of each task unit, construct a task processing behavior vector, evaluate the resource change rate, and adjust and sort the priority of each task unit to obtain a task execution priority sequence.

[0044] The task execution priority sequence specifically consists of task unit number, sorting weight coefficient, and priority sorting index.

[0045] As a further aspect of the present invention, based on the task structure partitioning results, the resource utilization rate and processing time of each task unit are collected, a task processing behavior vector is constructed, the resource change rate is evaluated, and the priority of each task unit is adjusted and sorted to obtain the task execution priority sequence. The specific steps are as follows:

[0046] S501: Based on the task structure division results, collect the resource utilization rate and processing time of each task unit, perform interval normalization processing and construct a joint data structure, combine the resource dimension and time dimension, form a two-dimensional behavior vector, and generate a task scheduling feature vector set.

[0047] The task scheduling feature vector set is constructed by performing interval normalization on the resource utilization rate and processing time of each task unit and combining them into a two-dimensional behavior vector.

[0048] S502: Call the resource dimension data in the task scheduling feature vector set, use the sliding window difference method to calculate the resource utilization difference value of consecutive task units in the same resource dimension, extract the absolute value sequence of the difference value as the fluctuation amplitude, calculate the fluctuation mean of each resource dimension, remove resource items with fluctuation values ​​less than the average offset value, calculate the resource dimension variation coefficient, summarize it into a stability index, and generate a resource utilization variation coefficient group.

[0049] The sliding window difference method uses a window size of 3 and a sliding step size of 1 to perform stepwise difference calculations on the resource utilization rate sequence, extract the difference between adjacent task units and form a fluctuation amplitude sequence.

[0050] The specific formula for calculating the coefficient of variation of the resource dimension is as follows:

[0051] ;

[0052] Calculate the coefficient of variation for the resource dimension;

[0053] in, Let be the coefficient of variation of the j-th resource dimension. Let the resource utilization rate of the j-th resource dimension be the resource utilization rate in the k-th task unit. Let be the arithmetic mean of the resource utilization rate of the j-th resource dimension across all n task units. This represents the total number of consecutive task units participating in resource utilization statistics. A minimal positive real perturbation factor, existing in constant form, is used in the denominator to prevent division by zero errors. This serves as the index identifier for the resource dimension. This is the index identifier for the j-th resource dimension. Time series index identifier for task unit;

[0054] The coefficient of variation for the resource dimension is obtained by statistically analyzing the usage rate sequence of each type of resource in each task unit and calculating the ratio of the standard deviation to the arithmetic mean of the usage rate sequence, which is used to reflect the degree of fluctuation in resource usage.

[0055] The resource utilization coefficient of variation set is a numerical set constructed by extracting the coefficient of variation of the utilization rate for each resource dimension and is used to reflect the stability and fluctuation range of resource utilization.

[0056] S503: Based on the resource utilization variation coefficient group and the normalized value of processing time in the task scheduling feature vector set, calculate the time proportion weight, and use the resource fluctuation value and processing time value to adjust and sort the priority of each task unit to generate a task execution priority sequence.

[0057] On the other hand, a multi-task processing system based on a large model is provided. This system is applied to multi-task processing methods based on large models and includes:

[0058] The resource conflict analysis module uses a large model to obtain the resource type code, access frequency and occupation duration of task nodes, calculates the Euclidean distance of resource demand vectors and constructs a conflict intensity matrix, combines a set threshold to filter node pairs, calls a density clustering algorithm to evaluate density and judge structural conflicts, and generates a recombinable task node structure to be passed to the scheduling structure generation module.

[0059] The scheduling structure generation module divides the upper and lower bound nodes according to the recombinable task node structure, calls the conflict intensity median deviation parameter to extract the upper bound node to construct an independent unit, matches the lower bound node and merges them according to the resource type dissimilarity rate, and generates a primary task scheduling unit set which is then passed to the dependency graph construction module.

[0060] The dependency graph construction module extracts the resource access channel number and scheduling trigger order based on the primary task scheduling unit set, constructs the dependency graph, calls path depth and branch factor to calculate the complexity score, adjusts the score weight in combination with the occupied time, prunes dense areas and performs subgraph and node aggregation, generates a dependency subgraph set and passes it to the structure partitioning module.

[0061] The structure partitioning module identifies the path call frequency and extracts the most frequent path based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, constructs the main path set, partitions the remaining nodes and forms scheduling groups, and generates the task structure partitioning result, which is then passed to the priority sorting module.

[0062] The priority sorting module collects the resource utilization rate and processing time of each task unit based on the task structure division result to construct a task processing behavior vector, evaluates the resource change rate, calls the priority adjustment function to sort, and generates a task execution priority sequence.

[0063] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0064] The large-scale multi-task processing system considers resource usage differences and structural distribution characteristics during the task partitioning phase, enhancing the resource independence and structural boundary identification capabilities of the scheduling unit and avoiding resource contention caused by high overlap between nodes. During scheduling path construction, it balances structural complexity and execution overhead, accurately identifying densely scheduled areas and pruning redundant nodes to improve path execution efficiency and controllability. Task partitioning results are driven by call frequency to categorize primary and secondary paths, and non-primary path tasks are grouped based on node coverage relationships, resulting in a clear partitioning structure and reduced interference between tasks. The task sorting mechanism simultaneously considers resource fluctuations and processing time, with the dynamic scheduling order reflecting the current resource status, thus improving overall scheduling accuracy and concurrent execution efficiency. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0067] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0068] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0069] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0070] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0071] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0072] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0073] Please see Figure 1 This invention provides a technical solution, a multi-task processing method based on a large model, comprising the following steps:

[0074] S1: Use a large model to obtain the resource type encoding, access frequency and occupation duration of task nodes, calculate the Euclidean distance of resource demand vectors to obtain the conflict intensity matrix, use a preset threshold to filter resource overlapping node pairs, judge structural conflicts based on node clustering characteristics, call density clustering algorithm to evaluate density, and generate recombinable task node structures.

[0075] S2: Based on the reconfigurable task node structure, divide the upper and lower bound nodes using the median deviation parameter of conflict intensity, extract the upper bound nodes to construct independent units, match the lower bound nodes and merge them according to the resource type dissimilarity rate to generate a primary task scheduling unit set.

[0076] S3: Extract resource access channel numbers and scheduling trigger order based on the primary task scheduling unit set, construct a dependency graph and calculate node path depth and branch factor to obtain complexity score, adjust weights based on occupied time, and prune dense areas to execute subgraphs and node aggregations according to the score to generate a dependency subgraph set;

[0077] S4: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the most frequently called path to form the main path set, divide the remaining nodes to form scheduling groups, and generate the task structure partitioning result;

[0078] S5: Based on the task structure partitioning results, collect the resource utilization rate and processing time of each task unit, construct the task processing behavior vector, evaluate the resource change rate, and adjust and sort the priority of each task unit to obtain the task execution priority sequence.

[0079] The reconfigurable task node structure includes node clustering labels, resource conflict labels, and structure adjustment indication information. The primary task scheduling unit set includes fusion node index, independent unit identifier, and unit structure configuration parameters. The dependency subgraph set includes node dependency relationship graph, task path hierarchy structure, and aggregation subgraph identifier. The task structure partitioning result includes main path node index, group boundary list, and structure level number. The task execution priority sequence specifically includes task unit number, sorting weight coefficient, and priority arrangement index.

[0080] Please see Figure 1 The specific steps for generating a recombinable task node structure include: using a large model to obtain the resource type encoding, access frequency, and occupancy duration of task nodes; calculating the Euclidean distance of resource demand vectors to obtain the conflict intensity matrix; using a preset threshold to filter resource-overlapping node pairs; judging structural conflicts based on node clustering characteristics; calling a density clustering algorithm to evaluate density; and generating a recombinant task node structure.

[0081] S101: Use a large model to obtain the resource type encoding, access frequency and duration of task nodes, construct and obtain vector representations according to resource dimensions, calculate the distance values ​​of multiple node pairs on each resource dimension, construct a symmetric conflict intensity matrix by combining the distances of each dimension, and generate resource distance intensity matrix values.

[0082] To obtain the resource type encoding, access frequency, and duration of task nodes using a large model, the task scheduling system first needs to call each task node in the task list. For each task node, the corresponding resource call information is extracted from the system resource log, recording the resource encoding (such as the type identifier code of CPU, GPU, memory, disk I / O, etc.), the number of times the resource is called per unit time (i.e., access frequency, in times / second), and the average duration of resource occupation by the task (in seconds). These three types of data are mapped into a vector according to the task node number and resource type. For example, if node A has an access frequency of 5 times / second on the CPU, an average occupation of 0.2 seconds per access, and a resource type of encoding 01, then its vector form is [01, 5, 0.2]. After organizing the data structure for each task node as described above, a task-resource triplet set can be obtained. Based on this, for any two nodes (such as nodes A and B), the Euclidean distance is calculated for their access frequency and duration of occupation under the dimension of consistent resource encoding. That is, let the access frequency and duration of occupation of node A be respectively... Node B is Then the distance on this resource dimension is:

[0083] ;

[0084] Repeat this process, calculating the distance values ​​for all resource dimensions, then combining the distances of each dimension into a vector of length equal to the total number of resources according to the resource encoding order. Finally, perform the corresponding operations on all task node pairs to construct a symmetric matrix D, where each element... This represents the combined distance between task node i and node j across all resource dimensions, used to express the degree of difference between their resource access characteristics. The following is an example data illustration:

[0085] Table 1. Task Node Resource Call Characteristics

[0086]

[0087] As shown in Table 1, the distance between nodes A and B under CPU resources is:

[0088] ;

[0089] The distance under GPU resources is:

[0090] ;

[0091] The combined resource distance vector of node pair AB is [2.002, 1.044]. This form of vector matrix is ​​constructed based on multiple node pairs, and the resource distance vector of each pair of nodes is regularized and stored. Finally, the resource distance strength value matrix between nodes is output in matrix form. The diagonal elements in the matrix are always 0, and the remaining symmetrically distributed elements represent the combined distance values ​​between any two nodes in all resource dimensions.

[0092] S102: Based on the resource distance strength matrix value, call the resource overlap judgment threshold, filter the task node combinations in the node pair whose resource distance is less than the resource overlap judgment threshold, and perform aggregation and classification according to the resource type coding information mapped by the node, identify the number of aggregated nodes under each type of resource coding, calculate the frequency distribution of each resource combination in the task set, and obtain the resource overlap node distribution rate.

[0093] Based on the resource distance strength matrix values, the resource distance combination values ​​of each task node pair in the matrix are read row by row. The specific resource distance of each node pair in each dimension is extracted. Let the combined distance of a task node pair AB in the CPU and GPU dimensions be [2.002, 1.044]. The resource overlap judgment threshold set by the system needs to be set based on the average distance and standard deviation between resources. The resource overlap threshold is set as: the mean of the resource distances of all node pairs minus a standard deviation. This standard deviation is calculated by taking the square root of the average of the squared differences of all distance values. If the average distance of all node pairs in the CPU dimension is set to 2.3 and the standard deviation is set to 0.5, then the threshold is set to 1.8. The judgment condition is: if the distance of a node pair in any resource dimension is lower than the threshold of that dimension, it is considered that the resources may overlap. During the screening, a judgment action is performed for each resource dimension, that is, a comparison is performed. Node pairs that meet the criteria will be extracted to form candidate node groups. The resource type encoding information of each candidate node group will be read, and the number of node pairs appearing under the same encoding will be counted. This count requires traversing all candidate node groups. A counter will be set up for each resource encoding group. If the CPU encoding is 01, and there are 8 pairs of nodes involving this resource in all candidate node groups, then the aggregate node count for 01 will be set to 8. Simultaneously, the node distribution frequency under this encoding will be calculated. The frequency is represented by the ratio of the number of times this resource encoding appears in the candidate nodes to the total number of node pairs. For example, if there are 50 candidate node pairs, and encoding 01 involves 8 pairs, then the frequency is... Frequency statistics need to be performed sequentially on all resource codes to form a frequency array, such as resource code 01 having a frequency of 0.16, code 02 having a frequency of 0.22, etc. Based on this, the distribution rate of overlapping resource nodes is obtained. The following is sample data:

[0094] Table 2 Resource Coding Frequency Statistics

[0095]

[0096] As shown in Table 2, the resource code 02 (GPU) has the highest distribution frequency, indicating that the GPU resources have a relatively high overlap degree in the current task nodes. This distribution frequency will be used as the initial input item in density clustering later, which will directly affect the density evaluation of the node structure. A reasonable range should be defined for the "higher frequency" in the above statistical steps. If the mean of all resource distribution frequencies is set to 0.15 and the standard deviation is 0.05, then the high-frequency resource threshold is set above 0.2, and the low-frequency resource threshold is set below 0.1, which is used as the basis for subsequent density estimation judgment. This process finally outputs the resource overlap node distribution rate under each resource type code, which is used to guide the density evaluation of high-overlap resources and the structural body aggregation operation;

[0097] The resource overlap determination threshold is the mean of the resource distances of all nodes minus the standard deviation.

[0098] S103: Invoke the resource overlap node distribution rate. Based on the access frequency and occupation duration data of each node in the resource combination, invoke the DBSCAN density clustering algorithm to calculate the local resource access density value, mark the high-density resource structure, construct a connectivity graph and classify and aggregate the high-density structures to obtain a reconfigurable task node structure;

[0099] Invoke the resource overlap node distribution rate. First, read the distribution frequency corresponding to each resource code in Table 2, and perform aggregation analysis on the resource types with frequencies greater than the set threshold value. For example, if the frequency corresponding to the GPU resource code 02 is 0.22, which has exceeded the preset high-density threshold of 0.2, then all task nodes involving this resource are extracted from the task set, and a resource access record table is established according to the node numbers, recording the access frequency and average occupation duration of each node to this resource. For example, the access frequency of node A is 2 times per second and the occupation duration is 1.5 seconds, and that of node B is 1 time per second and the occupation duration is 1.2 seconds. The system constructs a two-dimensional feature vector space according to the node resource combination, and represents each task node as a two-dimensional vector point , forming a coordinate set for density calculation. The density evaluation adopts the fixed radius method, and the search radius is set to 0.5, and the minimum number of neighborhood nodes MinPts is set to 3. Perform neighborhood search on any node in this two-dimensional space, and count the number of nodes falling within the radius centered on it. If the number ≥ MinPts, then this node is defined as a density core point. If the number of nodes falling within is < MinPts, but it itself is within the neighborhood of a certain core point, then it is a boundary point. If neither condition is met, then it is a noise point. The specific search action is: calculate the Euclidean distance between the current node and all other nodes in turn. Whenever the distance is less than The node is added to the neighborhood set. If the number of elements in the set reaches a threshold, it is marked as a core point. For example, node A has coordinates (2, 1.5), and the other nodes have coordinates (1.5, 1.4), (2.2, 1.6), and (3.0, 1.1). The distances to these nodes are calculated as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] Where d1 and d2 are less than 0.5, they fall into the neighborhood, and the number of nodes is 2, which does not satisfy MinPts=3. Therefore, A is not a core point. If A falls into the neighborhood of other core points, it is a boundary point. By traversing all nodes and building a neighborhood graph, multiple connected subgraphs are finally formed. Each subgraph consists of density-related nodes. The system classifies all node identifiers in the connected graph into the same type of resource structure, forming a recombinable task node structure, and stores it with labels. Each node in the structure still retains the original resource call parameters for subsequent reconstruction and optimization of task orchestration operations. This process uses the resource overlap frequency as the input condition and the local density value as the filtering basis to realize the structural mapping and labeling of resource-dense areas, complete the classification operation of connected subgraphs, and output the task node structure identifier set.

[0104] The DBSCAN density clustering algorithm calculates the Euclidean distance between two-dimensional vectors composed of the resource access frequency and occupation duration of task nodes, determines the search radius based on the inflection point method of the k-th nearest neighbor distance curve, and sets the minimum number of neighboring nodes by combining the dimension multiple method and the task size adjustment method. It calculates the local density between nodes, identifies high-density resource structures, and aggregates them into recombinable task node structures.

[0105] The local resource access density value is obtained by constructing a two-dimensional vector based on the access frequency and duration of task nodes, setting a search radius and a minimum number of neighboring nodes in the vector space, counting the number of nodes in the neighborhood and analyzing the local density.

[0106] Please see Figure 1 The specific steps for generating a primary task scheduling unit set, based on the reconfigurable task node structure, using the median deviation parameter of conflict intensity to divide nodes into upper and lower bounds, extracting upper bound nodes to construct independent units, matching lower bound nodes, and merging them according to the resource type dissimilarity rate, include:

[0107] S201: Based on the resource conflict intensity value corresponding to each group of nodes in the reconfigurable task node structure, set the median value of the conflict intensity according to the median deviation parameter, calculate the offset difference between the resource conflict value and the median value of each node, identify the offset direction, extract nodes with positive offset direction as upper bound nodes, extract nodes with negative offset direction as lower bound nodes, record the number of upper and lower bound nodes and their corresponding structural position information, and generate the distribution of upper and lower bound nodes.

[0108] Based on the resource conflict intensity value corresponding to each group of nodes in the reconfigurable task node structure, the conflict intensity values ​​of all nodes in the node set are first extracted, and a one-dimensional array set is constructed to represent the node conflict index sequence. Then, the median is calculated. Let the conflict intensity values ​​of the five nodes be 2.3, 1.7, 2.9, 1.5, and 2.1, respectively. The corresponding sorted sequence is [1.5, 1.7, 2.1, 2.3, 2.9], and the median value is 2.1. Let this median value be the reference benchmark. Then, the node set is traversed, and the median value is subtracted from the conflict intensity value of each node to obtain the offset value. Let the value of node A be 2.3, then the offset value is... If the offset is positive and node B is 1.7, then the offset value is... The offset is negative. The judgment criteria are: offset values ​​greater than 0 are classified as upper bound nodes, and offset values ​​less than 0 are classified as lower bound nodes. In this embodiment, the offset values ​​of node A and node C (2.9) are 0.2 and 0.8 respectively, and are classified as upper bound nodes. The offset values ​​of node B and node D (1.5) are -0.4 and -0.6 respectively, and are classified as lower bound nodes. The value of node E is 2.1, the offset value is 0, and it is not included in either category. Record the node number after classification and its position in the structure. Assume that the original position of node C is the 1st node of the 2nd group of the 3rd layer, then the position identifier is (3-2-1). The structural position information needs to be labeled in two levels according to the level and group to form a position array. At the same time, the number of upper and lower bound nodes is counted. Here, there are 2 upper bound nodes and 2 lower bound nodes. The output is the node identifier set and the distribution of upper and lower bound nodes.

[0109] Table 3. Task Node Resource Conflict Intensity Table

[0110]

[0111] As shown in Table 3, the offset direction is determined by the difference between the conflict intensity value and the median value, thereby realizing the extraction of upper and lower bound nodes and the recording of location information, and thus obtaining the distribution of upper and lower bound nodes.

[0112] The median deviation parameter is obtained by extracting the resource conflict intensity value of each node, calculating the median, extracting the difference between each node's value and the median, obtaining the offset difference, dividing the offset direction, and identifying the upper and lower bound nodes.

[0113] S202: Based on the distribution of upper and lower bound nodes, extract the set of upper bound nodes, identify the resource type code and original position information in the structure of each upper bound node, establish independent index relationships between nodes and mark the corresponding resource attribute items in the index path, call the difference in resource usage frequency between nodes, calculate the resource isolation ratio, construct the resource independent partition structure of upper bound nodes according to the isolation ratio, and generate resource utilization separation index.

[0114] Based on the distribution of upper and lower bound nodes, the upper bound node set is first extracted. Nodes A and C, identified in Table 3, are positive offset nodes, with their resource type codes set to 01 (CPU) and 03 (memory), respectively. Their layer and group information are then marked according to their original structural positions within the structure; for example, node A's original structural position is 1-1-1, and node C's is 3-2-1. Based on this, independent index paths are constructed. This involves generating a triplet of node number, resource code, and original position for each upper bound node to form an index structure, such as (A, 01, 1-1-1) and (C, 03, 3-2-1). Each index path needs to identify the involved nodes. Resource attribute items are extracted, which are the specific resource items corresponding to the resource type codes. For example, 01 represents CPU and 03 represents memory. Further, the resource usage frequency between nodes is read. Nodes belonging to the same upper bound set are paired according to their resource codes. For each paired node pair under each resource type, the difference between access frequency values ​​is calculated. Assuming nodes A and C have access frequencies of 4 times / second and 2 times / second respectively, the difference is 2 times / second. The average difference ratio is used to evaluate the resource isolation level, which is the ratio of the sum of the access frequency differences of all paired nodes for that resource type to the average access frequency. Let the sum of the access frequency differences be 4 and the average access frequency be 3, then the isolation ratio is... The same calculation operation is performed on the node set to calculate the isolation ratio under all resource codes and form independent partition indicators. Resources with isolation ratios higher than a set threshold are considered independent resources. If the isolation threshold is set to 1.2, then the above ratio of 1.33 meets the independence condition. The node pair is classified as an independent resource unit in the resource utilization structure. The above process is repeated for all node pairs under all resource codes. The values ​​of the resource isolation ratios are formed into an array, let the array be [1.33, 0.95, 1.5], which represent the isolation strength of the three types of resources respectively. Compared with the threshold array [1.2, 1.2, 1.2], the first and third items meet the independence condition. Therefore, the resource independent partition structure of the upper bound node is generated. At the same time, a separation index matrix is ​​generated with the resource code as the key value to reflect the degree of independence under different resources.

[0115] This process ultimately forms a resource partitioning structure with independent upper bound nodes, identifying the resource utilization differences between each node, and thus constituting a set of resource utilization separation indicators. This indicator set will subsequently be used as a key reference factor in the lower bound node matching and fusion structure construction logic.

[0116] The resource isolation ratio is calculated by taking the difference in access frequency of each node in the upper bound node set for the same resource type, and normalizing the difference to the average access frequency. It is used to quantify the degree of heterogeneity in resource usage and to construct an independent resource partitioning structure.

[0117] S203: Call the index information and node resource code content corresponding to each upper bound node in the resource utilization separation index. Based on the structure number and connected edge relationship in the index path, identify the lower bound nodes that have direct dependency connection with the upper bound node and continuous structure number. Extract the resource type code of the corresponding lower bound node. Calculate the resource type dissimilarity rate based on the number of resource type differences and the total number of resource items. Based on the dissimilarity rate value, perform fusion matching and node reorganization on the upper and lower bound nodes to generate a primary task scheduling unit set.

[0118] The specific formula for calculating the resource type difference rate based on the number of resource type differences and the total number of resource items is as follows:

[0119] ;

[0120] Calculate the resource type dissimilarity rate;

[0121] in, The resource type distinctness rate is expressed in dimensionless form. The total number of resource types defined in the system. For the index number of the resource type, For the index number of the resource type, This is the index number of the lower bound node. This represents the total number of lower bound nodes extracted so far. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first The lower bound node of the nth node The proportion or importance weight of a resource type within the node's resource structure, expressed in dimensionless units. For the first The lower bound node of the nth node The proportion of a resource type in the node's resource structure or the importance weight of a resource type, in dimensionless units;

[0122] formula:

[0123] ;

[0124] Detailed explanation of the formula and the derivation process of the formula calculation;

[0125] The formula is used to calculate the resource type dissimilarity rate, and the result is used to measure the degree of difference in resource type distribution among lower bound nodes;

[0126] Parameter meanings and settings:

[0127] The resource type dissimilarity rate reflects the degree of difference in resource type distribution among lower bound nodes;

[0128] The total number of resource types is set to 4;

[0129] The total number of lower bound nodes is set to 3;

[0130] For the first Does the nth lower bound node have the nth lower bound node? The status values ​​of the class resource;

[0131] For the first The lower bound node of the nth node The proportion of a resource type within the node's resource structure, or the importance weight of a resource type, is set as follows:

[0132] For resource type 1, set it to 0.4;

[0133] For resource type 2, set it to 0.3;

[0134] For resource type 3, set it to 0.2;

[0135] For resource type 4, set it to 0.1;

[0136] Substitute the parameters into the formula to calculate:

[0137] The relationship between lower bound nodes and resource types is defined as follows:

[0138] Node 1: Resource Type 1 ( , Resource type 2 () , );

[0139] Node 2: Resource Type 2 ( , Resource type 3 () , );

[0140] Node 3: Resource Type 3 ( , Resource type 4 () , );

[0141] Calculate the weighted sum for each resource type:

[0142] Resource type 1: ;

[0143] Resource type 2: ;

[0144] Resource type 3: ;

[0145] Resource type 4: ;

[0146] Calculate the weighted sum of all resource types:

[0147] ;

[0148] Calculate the weighted average:

[0149] ;

[0150] Calculate the numerator:

[0151] ;

[0152] Calculate the denominator:

[0153] ;

[0154] The final calculation of the resource type dissimilarity rate is as follows:

[0155] ;

[0156] A result of 0.0292 indicates that the distribution of resource types among the lower bound nodes is relatively uniform, with low variation. This result can be used to evaluate the rationality of resource allocation and guide subsequent resource optimization and scheduling strategies.

[0157] The resource utilization separation index is obtained by extracting the resource type code and resource usage frequency of the upper bound node, calculating the difference in access frequency between nodes on the same resource type and normalizing it to the isolation ratio, constructing an independent resource partitioning structure, and obtaining the resource utilization separation index.

[0158] The resource type dissimilarity rate is obtained by identifying the proportion of non-overlapping resource types between upper and lower bound nodes to the total number of resource types, and by pairing and merging nodes based on their resource codes and structure numbers.

[0159] Please see Figure 1 The specific steps for generating a dependency subgraph set include: extracting resource access channel numbers and scheduling trigger orders from the primary task scheduling unit set; constructing a dependency graph and calculating node path depth and branch factor to obtain a complexity score; adjusting weights based on the duration of use; and pruning dense regions into subgraphs and aggregating nodes according to the score.

[0160] S301: Extract resource access channel numbers and scheduling triggering order based on the primary task scheduling unit set, extract directed connection paths between nodes according to the triggering sequence, construct the position index matrix of each pair of nodes in the path graph, and set the weight and connection direction of the edges between nodes to generate a task node dependency graph.

[0161] Based on the extraction of resource access channel numbers and scheduling trigger orders from the primary task scheduling unit set, the process first traverses all scheduling unit node sets, reading the channel number and corresponding trigger order of each task node. These are then sorted in ascending order to construct a scheduling execution sequence. In the sample data, the trigger orders of nodes A to E are 1 to 5, forming the node trigger sequence [A, B, C, D, E]. For adjacent nodes in this sequence, directed connection paths are defined sequentially, such as A→B, B→C, C→D, D→E. A position index matrix is ​​generated for each pair of connected nodes, where the element in the i-th row and j-th column is marked as 1 to indicate the existence of a path A(i)→B(j). A 5×5 path matrix is ​​constructed for all node pairs and filled with directed connections. Simultaneously, the starting and target node channel numbers of each node in each path are recorded, and connection weights are set based on the difference in node trigger orders. Let the connection weight be the difference in trigger orders. The weight is calculated in reverse order. For example, if the triggering difference between A and B is 1, then the weight is 1. If the difference between C and D is 1, then the weight is also 1. If the triggering interval between two nodes exceeds 1, then the weight is reduced accordingly. For example, if the difference between A and C is 2, then the weight is 0.5. In this path, the connection direction is from the lower-order node to the higher-order node. Weight and direction annotations are added to all connection edges to finally construct the task node dependency graph. The graph structure is represented by an adjacency list. Each node stores the number of the subsequent node it points to and the corresponding edge weight, forming a complete dependency directed graph structure.

[0162] Table 4. Scheduling Unit Node Triggering and Channel Information Table

[0163]

[0164] As shown in Table 4, node-dependent directed paths can be constructed and edge weights and connection directions can be assigned according to the triggering order and channel number, ultimately generating a task node dependency graph.

[0165] S302: Call the path index and edge connection relationship in the task node dependency graph, identify the number of nodes in each path and use it as the path depth, identify the number of subsequent nodes connected to each node and use it as the branch factor, call the path depth and branch factor to calculate the original complexity score, combine the occupied time information of each node recorded in the scheduling table, and set the score adjustment coefficient according to the occupied time to adjust the original complexity score and obtain the node scheduling complexity score value.

[0166] The process involves calling the path indices and edge connections in the task dependency graph. First, the completed task dependency graph structure is read, and the number of nodes in each path is counted as the path depth. For each complete path from the source node to the terminal node, a node counting operation is performed. For example, in the path A→B→C→D→E, if the number of nodes is 5, then the path depth is 5. Simultaneously, outgoing edge analysis is performed on each node, counting the number of its directly connected successor nodes, which is used as the branch factor for that node. For example, if node A only points to node B, then A's branch factor is 1. If a node points to multiple successor nodes, its branch factor increases accordingly. For example, if node F points to nodes G and H, then its branch factor is 2. The original complexity score is expressed as the product of the path depth and the branch factor. Let the path depth be... The node branch factor is The complexity score is For each path, the average score of all nodes is taken as the path complexity score. For example, if the branch factors of the 5 nodes in path A→E are 1, 1, 1, 1, and 0, with an average of 0.8, then the original score is... Based on the resource occupancy duration recorded for each node in the scheduling table, the score is weighted and adjusted. Let node A's occupancy duration be 0.5 seconds, B's be 1.0 seconds, C's be 0.8 seconds, D's be 1.2 seconds, and E's be 0.5 seconds. The average occupancy duration is calculated to be 0.8 seconds. The score adjustment coefficient is calculated as the ratio of occupancy duration to the average. Let node C's occupancy duration be 0.8 seconds, and the coefficient be 1.0; node D's occupancy duration be 1.2 seconds, and the coefficient be 1.5. The final adjusted score is the original score multiplied by the weight coefficient to form the final scheduling complexity score. If the original score is 4.0 and the average coefficient is 1.1, then the final score is... This operation is performed on all node paths to form a complete node scheduling complexity score table, providing data support for subsequent screening and structure pruning.

[0167] S303: Based on the node scheduling complexity score, filter the node score data, extract the node indexes with scores greater than the complexity score percentile threshold, and identify the corresponding aggregation distribution segments in the dependency graph. Based on the score difference between adjacent nodes, construct and trim continuous aggregation path intervals in the graph structure. Each node in the aggregation segment is a subgraph merging object, generating a dependency subgraph set.

[0168] Based on the node scheduling complexity scores, the score data for all nodes is filtered. First, all calculated scheduling complexity scores are extracted to form a one-dimensional score array. For example, if the score array is [2.3, 4.4, 3.1, 5.0, 3.6], the statistical quantile values ​​of the score distribution are calculated based on this array. The complexity score quantile threshold is set to the 80th percentile, meaning that the score values ​​located at the 80th percentile after sorting are extracted as the filtering threshold. If the sorted array is [2.3, 3.1, 3.6, 4.4, ... [5.0], the 80th position is the 4th value 4.4. Let this value be the score percentile threshold. Compare all scores one by one and filter out the nodes with scores greater than 4.4. The node number is E (5.0). Extract the topological structure position of the node in the dependency graph. According to its connected path index, obtain the path segment where node E is located. Record the connection status and score difference between it and the nodes before and after it. For example, the score of node D is 4.4 and the score of node E is 5.0, with a difference of 0.6. If we assume that the score difference between adjacent nodes is less than 1.0 as a continuous aggregation segment, then D and E constitute a continuous path segment. Aggregate this path segment. The aggregation process involves merging all node labels and edge relationships within a segment to form a subgraph object. If nodes E and F also satisfy the condition of continuous scores and a difference of less than 1.0, the extended path segment becomes D→E→F, generating a complete continuous path interval. This interval is then extracted as an independent subgraph in the dependency graph, and all node numbers and edge sets in this subgraph are recorded. After this filtering and aggregation process, all aggregated node substructures with prominent scores are extracted, ultimately constructing multiple aggregated subgraph sets. Each subgraph represents the structural expression of a score-dense segment, used for structural fusion and scheduling compression in subsequent optimization stages. This process implements the logic for pruning the dependency graph structure and generating subgraphs under score-driven conditions, and the output is a set of dependency subgraphs and their structural list.

[0169] The complexity score percentile threshold is determined by constructing a sorted array based on the scheduling complexity scores of all nodes, and then using a grid search method to compare the scheduling performance of each percentile value on multiple task sets, selecting the score value at the 80th percentile as the threshold for filtering dense path nodes.

[0170] Please see Figure 1 Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, the most frequently invoked paths are extracted to form the main path set, and the remaining nodes are divided into scheduling groups to generate the task structure partitioning result. The specific steps include:

[0171] S401: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the node number included in each path, collect the trigger frequency of the corresponding node in the scheduling log, identify the total number of calls for each node in each path, extract the path number with the highest trigger frequency in each path and construct a dataset to generate a high-frequency path trigger sequence.

[0172] Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, each dependency subgraph path is first numbered, and all node numbers contained within it are extracted. In the example, path P1 contains nodes A and B, P2 contains nodes C and D, and P3 contains nodes E and F. Then, the scheduling log is accessed to extract the trigger frequency of each node. Let the trigger frequency of node A be 15 times, B be 20 times, C be 12 times, D be 10 times, E be 25 times, and F be 18 times. Next, the frequencies of all nodes within each path are summed. The total frequency of path P1 is 15+20=35, P2 is 22, and P3 is 43. A path frequency mapping table is formed by combining all path numbers with their corresponding frequency values. The maximum value extraction operation is performed, that is, the path with the highest frequency value is selected from all paths. The main path number is determined to be P3. The main path has the highest trigger frequency and is set as a high-frequency path. A trigger sequence is constructed according to the path number. If multiple paths have the same frequency, they are processed into the main path sequence in ascending order of number. This sequence is used to represent the dominant execution line in the subsequent task scheduling logic. Finally, a set of high-frequency path trigger sequences is generated.

[0173] Table 5. Frequency of Scheduling Path Triggering

[0174]

[0175] As shown in Table 5, path P3 is identified as the main path because its node set has the highest total call frequency, and the main path trigger sequence thus forms a frequent path set.

[0176] The high-frequency path trigger sequence is a set of frequent paths obtained by identifying the total trigger frequency of each node in the scheduling log of each path and selecting the set of path numbers with the highest trigger frequency.

[0177] S402: Call the path number in the high-frequency path trigger sequence to obtain all the corresponding node numbers, remove the main path node numbers from the node number set of the primary task scheduling unit set and extract the remaining node number set, and perform edge traversal on the resource dependency relationship between the nodes within the remaining node number set. Based on the connection directionality and resource channel number, perform hierarchical clustering analysis on the remaining nodes to generate a node scheduling partition structure.

[0178] The process retrieves all corresponding node numbers from the path number in the high-frequency path trigger sequence. In the example, the main path is P3, and the node numbers are E and F. The set of all task node numbers is extracted from the primary task scheduling unit, denoted as [A, B, C, D, E, F]. The main path nodes E and F are removed, leaving the set of remaining nodes as [A, B, C, D]. Then, resource dependency extraction is performed within this remaining node set. For any two nodes, a comparison analysis of resource channel numbers and connection directions is performed. Let node A's channel be CH1 and node B's channel be CH2. If A is earlier than B in the scheduling trigger order, and A and B share resources or are indirectly connected through resource transfer, then A→B constitutes a dependency edge. This process is repeated for all node pairs, and dependency edges are marked, ultimately forming a dependency subgraph structure among the remaining nodes. Subsequently, dependency clustering is performed on this subgraph. The clustering process follows this logic: First, starting with any node, all directly connected subsequent nodes are traversed, and records are generated. The resource channel number is used to group multiple nodes into the same group if they form a chain path and use the same channel number. For example, if nodes A→B→C all use channel CH1, then A, B, and C are clustered into one group. If a node uses a different channel, it is split at the boundary and assigned to another group. For example, if node D uses channel CH3, then it forms an independent partition. This process is repeated to complete the dependency grouping of all nodes, and finally, multiple node scheduling partition structures are generated. Each partition represents a set of task nodes whose resource channels and scheduling sequence logic are logically coupled. This structure serves as an important intermediate product for task scheduling structure partitioning and is used for subsequent task number mapping.

[0179] S403: Construct a path attribution index table based on the remaining nodes in the node scheduling partition structure, construct a group number mapping table using the set of main path nodes in the high-frequency path trigger sequence, merge the path attribution index table and the group number mapping table into a unified number index structure, and map it to the scheduling structure diagram to generate the task structure partitioning result;

[0180] Based on the remaining nodes in the node scheduling partition structure, a path attribution index table is constructed. First, a unique group number is assigned to each scheduling partition. Let nodes A, B, and C be assigned to partition 1, and node D to partition 2. Then, the path attribution index table format is {A: G1, B: G1, C: G1, D: G2}. This index structure indicates the scheduling group number to which each node belongs in the structural partition. Next, the main path node sets E and F extracted from the high-frequency path trigger sequence are used to construct the main path mapping number. Let the main path number be M1, then the mapping relationship is {E: M1, F: M1}. The path attribution index table and the main path mapping table are merged to generate a unified number index structure. This structure assigns a unique number label to all nodes, where the main path nodes are the main path number, and non-main path nodes are the scheduling partition number. The unified structure is represented as {A: G1, B: G1, C: G1, D: G2, E: M1, F: Next, the numbering structure is mapped to the scheduling structure graph. Number replacement is performed on the nodes in the original graph, making each node display its assigned number in the structure graph. Finally, the entire task graph is structurally partitioned based on the assigned numbers, grouping nodes with the same number into a substructure. A structural partitioning graph is generated in the graph, with each subgraph represented as a structural block. Nodes within a structural block retain connections based on their original dependencies, while structural blocks are treated as logical module connections. The final output is a complete task structure partitioning result. This result possesses complete numbering attribution, scheduling relationships, and resource logic, facilitating the scheduling system to perform structural-level scheduling planning.

[0181] The node scheduling partitioning structure is formed by traversing the resource dependencies between the remaining nodes after removing the main path nodes, and then performing dependency clustering based on the resource channel number and connection directionality to form a grouping structure for task nodes.

[0182] Please see Figure 1 Based on the task structure partitioning results, the resource utilization rate and processing time of each task unit are collected to construct a task processing behavior vector. The resource change rate is evaluated, and the priority of each task unit is adjusted and sorted to obtain the task execution priority sequence. The specific steps include:

[0183] S501: Based on the task structure partitioning results, collect the resource utilization rate and processing time of each task unit, perform interval normalization processing and construct a joint data structure, combine the resource dimension and time dimension, form a two-dimensional behavior vector, and generate a task scheduling feature vector set.

[0184] Based on the task structure partitioning results, the resource utilization rate and corresponding processing time of each task unit are first extracted. A column vector of resource utilization rate and a column vector of processing time are constructed and normalized. The normalization method is linear interval normalization to the range [0,1]. The resource utilization rate normalization is performed according to the formula. Assuming the original resource utilization rate values ​​are [80,65,90,50,70], the maximum value is 90, the minimum value is 50, and after normalization, they are [(80-50) / (90-50)=0.75,(65-50) / 40=0.375,(90-50) / 40=1.0,(50-50) / 40=0.0,(70-50) / 40=0.5]. The original processing time values ​​are [6.0,4.5,8.0,3.0,5.5], the maximum value is 8.0, the minimum value is 3.0, and after normalization... [(6.0-3.0) / 5=0.6,(4.5-3.0) / 5=0.3,(8.0-3.0) / 5=1.0,(3.0-3.0) / 5=0.0,(5.5-3.0) / 5=0.5], combine the normalized resource values ​​with the time values ​​to form a two-dimensional vector pair. Each task unit can be represented as (normalized resource usage value, normalized duration value), such as task T1 as (0.75,0.6) and T3 as (1.0,1.0). The two-dimensional vectors of all task units constitute the scheduling feature vector set, with the structure {T1:[0.75,0.6],T2:[0.375,0.3],T3:[1.0,1.0],T4:[0.0,0.0],T5:[0.5,0.5]}, which is used as the basic data for subsequent resource change rate analysis and priority ranking.

[0185] Table 6. Normalized Vector Table of Task Unit Resources and Duration

[0186]

[0187] As shown in Table 6, each task unit forms a two-dimensional scheduling behavior vector by normalizing resource and duration data, which is used as the basic data set in subsequent priority adjustment calculations.

[0188] The task scheduling feature vector set is constructed by normalizing the resource utilization and processing time of each task unit within a range and combining them into a two-dimensional behavior vector.

[0189] S502: Call the resource dimension data in the task scheduling feature vector set, use the sliding window difference method to calculate the resource utilization difference value of consecutive task units in the same resource dimension, extract the absolute value sequence of the difference value as the fluctuation amplitude, calculate the fluctuation mean of each resource dimension, remove resource items with fluctuation values ​​less than the average offset value, calculate the resource dimension variation coefficient, summarize it into a stability index, and generate a resource utilization variation coefficient group.

[0190] The sliding window difference method uses a window size of 3 and a sliding step size of 1 to perform stepwise difference calculations on the resource utilization rate sequence, extract the difference between adjacent task units and form a fluctuation amplitude sequence.

[0191] The specific formula for calculating the coefficient of variation of the resource dimension is as follows:

[0192] ;

[0193] Calculate the coefficient of variation for the resource dimension;

[0194] in, Let be the coefficient of variation of the j-th resource dimension. Let the resource utilization rate of the j-th resource dimension be the resource utilization rate in the k-th task unit. Let be the arithmetic mean of the resource utilization rate of the j-th resource dimension across all n task units. This represents the total number of consecutive task units participating in resource utilization statistics. A minimal positive real perturbation factor, existing in constant form, is used in the denominator to prevent division by zero errors. This serves as the index identifier for the resource dimension. This is the index identifier for the j-th resource dimension. Time series index identifier for task unit;

[0195] formula:

[0196] ;

[0197] Detailed explanation of the formula and its calculation derivation:

[0198] The formula is used to calculate the dispersion of the utilization rate of each type of resource dimension during task scheduling. The results are used to identify the fluctuation stability characteristics of resource utilization and support the subsequent stability induction of resource structure.

[0199] Parameter meanings and settings:

[0200] Let be the coefficient of variation for the j-th resource dimension. The sampling results recorded in the task scheduling log are set to: 62%, 65%, 58%, 60%, and 55%, i.e. , , , , ;

[0201] The arithmetic mean of the usage rates of the five task units in this resource dimension is calculated as follows:

[0202] ;

[0203] The number of continuous sampling task units is set to 5;

[0204] The smallest positive disturbance term is set to 0.001 to prevent the denominator from being zero;

[0205] Substitute the parameters into the formula to calculate:

[0206] Calculate the difference of squares:

[0207] ;

[0208] Calculate the average:

[0209] ;

[0210] Calculate the square root:

[0211] ;

[0212] Substitute into the formula to calculate the coefficient of variation:

[0213] ;

[0214] The result of 0.05667 indicates that the utilization rate of the j-th resource dimension has a low degree of variation in the current task scheduling process, with a fluctuation range of less than 6%. This value is used to screen the dimension items with significant fluctuations in the resource utilization variation coefficient matrix, further summarize the stability index, and improve the coordination of resource scheduling distribution.

[0215] The coefficient of variation for resource dimensions is calculated by statistically analyzing the usage rate sequence of each type of resource in each task unit and then calculating the ratio of the standard deviation of the usage rate sequence to the arithmetic mean. This coefficient reflects the degree of fluctuation in resource usage.

[0216] The resource utilization coefficient of variation set is a numerical set constructed by extracting the coefficient of variation of utilization rate for each resource dimension, and is used to reflect the stability and fluctuation range of resource utilization.

[0217] S503: Based on the resource utilization variation coefficient group and the normalized value of processing time in the task scheduling feature vector set, calculate the time proportion weight, and use the resource fluctuation value and processing time value to adjust and sort the priority of each task unit to generate a task execution priority sequence.

[0218] Based on the resource utilization coefficient of variation set and the normalized processing time value in the task scheduling feature vector set, the variability of the normalized resource utilization value of all task units needs to be evaluated first. Let the formula for the coefficient of variation be... ,in Standard deviation, To calculate the mean, we perform standard deviation calculation on the resource normalized value array [0.75, 0.375, 1.0, 0.0, 0.5]. The mean is (0.75 + 0.375 + 1.0 + 0.0 + 0.5) / 5 = 0.525, and the squared difference is [(0.75 - 0.525)² + (0.375 - 0.525)² + (1.0 - 0.525)² + (0.0 - 0.525)² + (0.5 - 0.525)²] / 5 = (0.0506 + 0.0225 + 0.2256 + 0.2756 + 0.0006) / 5 = 0.5749 / 5 = 0.1150. The standard deviation is... coefficient of variation Then, the time percentage weight of each task unit is calculated. Let the weight be the proportion of the normalized processing time of that task unit to the sum of the normalized processing times of all tasks. The total normalized time value is 0.6 + 0.3 + 1.0 + 0.0 + 0.5 = 2.4. The time weight of T1 is 0.6 / 2.4 = 0.25, and for T3 it is 1.0 / 2.4 ≈ 0.417, and so on. The resource fluctuation value is set as the absolute value of the difference between the normalized resource value of the task and the mean of 0.525. For example, the fluctuation value of T3 is |1.0 - 0.525| = 0.475. This value is combined with the processing time value to calculate the priority score. Let the score be the weighted sum of resource fluctuation and time weight, with weighting coefficients α = 0.6 (resource fluctuation) and β = 0.4 (time weight). Then the priority score for T3 is... The same calculation is performed on all task units to obtain a score, and finally, the tasks are sorted from largest to smallest score to form a task execution priority sequence, such as [T3, T1, T5, T2, T4]. This sequence is used by the scheduler to determine the task start order during the execution phase. This sorting result is constructed based on the intensity of resource fluctuations and the proportion of time, and implements the quantitative arrangement logic of task priorities.

[0219] Please see Figure 2 A multi-task processing system based on a large model, used to execute the aforementioned multi-task processing method based on a large model, includes:

[0220] The resource conflict analysis module uses a large model to obtain the resource type code, access frequency and occupation duration of task nodes, calculates the Euclidean distance of resource demand vectors and constructs a conflict intensity matrix, combines a set threshold to filter node pairs, calls a density clustering algorithm to evaluate density and judge structural conflicts, and generates a recombinable task node structure to be passed to the scheduling structure generation module.

[0221] The scheduling structure generation module divides the upper and lower bound nodes according to the recombinable task node structure, calls the conflict intensity median deviation parameter to extract the upper bound node to build an independent unit, matches the lower bound node and merges them according to the resource type dissimilarity rate, and generates a primary task scheduling unit set which is then passed to the dependency graph construction module.

[0222] The dependency graph construction module extracts resource access channel numbers and scheduling trigger order based on the primary task scheduling unit set, constructs a dependency graph, calls path depth and branch factor to calculate complexity score, adjusts score weight based on occupied time, prunes dense areas and performs subgraph and node aggregation, generates a dependency subgraph set and passes it to the structure partitioning module.

[0223] The structure partitioning module identifies the path call frequency and extracts the most frequent path based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, constructs the main path set, partitions the remaining nodes and forms scheduling groups, and generates the task structure partitioning result, which is then passed to the priority sorting module.

[0224] The priority sorting module collects the resource utilization and processing time of each task unit based on the task structure division results to construct a task processing behavior vector, evaluates the resource change rate, calls the priority adjustment function to sort, and generates a task execution priority sequence.

[0225] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0226] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0227] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0228] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0229] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0230] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0231] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0232] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0233] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0234] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0235] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-task processing method based on a large model, characterized in that, The method includes: S1: Use a large model to obtain the resource type encoding, access frequency and occupation duration of task nodes, calculate the Euclidean distance of resource demand vectors to obtain the conflict intensity matrix, use a preset threshold to filter resource overlapping node pairs, judge structural conflicts based on node clustering characteristics, call density clustering algorithm to evaluate density, and generate recombinable task node structures. S2: Based on the reconfigurable task node structure, divide the upper and lower bound nodes using the conflict intensity median deviation parameter, extract the upper bound nodes to construct independent units, match the lower bound nodes and merge them according to the resource type dissimilarity rate to generate a primary task scheduling unit set. S3: Based on the primary task scheduling unit set, extract the resource access channel number and scheduling trigger order, construct a dependency graph and calculate the node path depth and branch factor to obtain a complexity score, adjust the weights in combination with the occupied time, and perform subgraph and node aggregation in dense areas according to the score to generate a dependency subgraph set; S4: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the path with the highest call frequency to form the main path set, divide the remaining nodes to form scheduling groups, and generate the task structure partitioning result.

2. The multi-task processing method based on a large model according to claim 1, characterized in that, The reconfigurable task node structure includes node clustering labels, resource conflict labels, and structure adjustment indication information. The primary task scheduling unit set includes fusion node index, independent unit identifier, and unit structure configuration parameters. The dependency subgraph set includes node dependency relationship graph, task path hierarchy structure, and aggregation subgraph identifier. The task structure partitioning result includes main path node index, group boundary list, and structure level number.

3. The multi-task processing method based on a large model according to claim 1, characterized in that, The specific steps for generating a recombinable task node structure are as follows: Utilizing a large model to obtain the resource type encoding, access frequency, and duration of task nodes; calculating the Euclidean distance of resource demand vectors to obtain the conflict intensity matrix; using a preset threshold to filter resource-overlapping node pairs; judging structural conflicts based on node clustering characteristics; and calling a density clustering algorithm to evaluate density. S101: Use a large model to obtain the resource type encoding, access frequency and duration of task nodes, construct and obtain vector representations according to resource dimensions, calculate the distance values ​​of multiple node pairs on each resource dimension, construct a symmetric conflict intensity matrix by combining the distances of each dimension, and generate resource distance intensity matrix values. S102: Based on the resource distance strength matrix value, call the resource overlap judgment threshold, filter the task node combinations in the node pair whose resource distance is less than the resource overlap judgment threshold, and perform aggregation and classification according to the resource type encoding information mapped by the node, identify the number of aggregated nodes under each type of resource encoding, calculate the frequency distribution of each group of resource combinations in the task set, and obtain the resource overlap node distribution rate. The resource overlap determination threshold is the mean of the distances between all nodes and the resource, minus the standard deviation. S103: Call the resource overlap node distribution rate, and based on the access frequency and occupancy time data of each node in the resource combination, call the DBSCAN density clustering algorithm to calculate the local resource access density value, mark the high-density resource structure, construct the connectivity graph and classify and aggregate the high-density resource structure to obtain the recombinable task node structure. The DBSCAN density clustering algorithm calculates the Euclidean distance between two-dimensional vectors composed of the resource access frequency and occupation duration of task nodes, determines the search radius based on the inflection point method of the k-th nearest neighbor distance curve, and sets the minimum number of neighboring nodes by combining the dimension multiple method and the task size adjustment method. It calculates the local density between nodes, identifies high-density resource structures, and aggregates them into recombinable task node structures. The local resource access density value is obtained by constructing a two-dimensional vector based on the access frequency and duration of task nodes, setting a search radius and a minimum number of neighboring nodes in the vector space, counting the number of nodes in the neighborhood and analyzing the local density.

4. The multi-task processing method based on a large model according to claim 3, characterized in that, Based on the reconfigurable task node structure, the steps of dividing upper and lower bound nodes using the conflict intensity median deviation parameter, extracting upper bound nodes to construct independent units, matching lower bound nodes, and merging them according to the resource type dissimilarity rate to generate a primary task scheduling unit set are as follows: S201: Based on the resource conflict intensity value corresponding to each group of nodes in the reconfigurable task node structure, set the median value of the conflict intensity according to the median deviation parameter, calculate the offset difference between the resource conflict value and the median value of each node, identify the offset direction, extract nodes with positive offset direction as upper bound nodes, extract nodes with negative offset direction as lower bound nodes, record the number of upper and lower bound nodes and the corresponding structural position information, and generate the upper and lower bound node distribution. The median deviation parameter is obtained by extracting the resource conflict intensity value of each node, calculating the median, extracting the difference between each node's value and the median, obtaining the offset difference, dividing the offset direction, and identifying the upper and lower bound nodes. S202: Based on the distribution of the upper and lower bound nodes, extract the set of upper bound nodes, identify the resource type code and original position information in the structure of each upper bound node, establish independent index relationships between nodes and mark the corresponding resource attribute items in the index path, call the resource usage frequency difference between nodes, calculate the resource isolation ratio, construct the resource independent partition structure of the upper bound nodes according to the isolation ratio, and generate the resource utilization separation index. The resource isolation ratio is calculated by taking the difference in access frequency of each node in the upper bound node set on the same resource type, and normalizing the difference to the average access frequency ratio. This ratio is used to quantify the heterogeneity of resource usage and the construction of an independent resource partitioning structure. S203: Call the index information and node resource code content corresponding to each upper bound node in the resource utilization separation index, identify the lower bound nodes that have direct dependency connections with the upper bound nodes and continuous structure numbers according to the structure number and connected edge relationship in the index path, extract the resource type code of the corresponding lower bound node, calculate the resource type dissimilarity rate according to the number of resource type differences and the total number of resource items, and perform fusion matching and node reorganization on the upper and lower bound nodes according to the dissimilarity rate value to generate a primary task scheduling unit set; The resource utilization separation index is obtained by extracting the resource type code and resource usage frequency of the upper bound node, calculating the difference in access frequency between nodes on the same resource type and normalizing it to the isolation ratio, constructing an independent resource partitioning structure, and obtaining the resource utilization separation index. The resource type dissimilarity rate is obtained by identifying the proportion of non-overlapping resource types between upper and lower bound nodes to the total number of resource types, and by pairing and merging them according to the node resource code and structure number.

5. The multi-task processing method for large models according to claim 4, characterized in that, The specific formula for calculating the resource type difference rate based on the number of resource type differences and the total number of resource items is as follows: ； Calculate the resource type dissimilarity rate; in, The resource type distinctness rate is expressed in dimensionless form. The total number of resource types defined in the system. For the index number of the resource type, For the index number of the resource type, This is the index number of the lower bound node. This represents the total number of lower bound nodes extracted so far. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first Does the nth lower bound node have the nth lower bound node? The status value of the resource class is in dimensionless form. For the first The lower bound node of the nth node The proportion or importance weight of a resource type within the node's resource structure, expressed in dimensionless units. For the first The lower bound node of the nth node The proportion of a resource type in the node's resource structure or the importance weight of a resource type, expressed in dimensionless units.

6. The multi-task processing method based on a large model according to claim 4, characterized in that, The steps for extracting resource access channel numbers and scheduling trigger orders based on the primary task scheduling unit set, constructing a dependency graph and calculating node path depth and branch factor to obtain a complexity score, adjusting weights based on occupied time, and pruning dense regions according to the score to execute subgraphs and node aggregation to generate a dependency subgraph set are as follows: S301: Extract the resource access channel number and scheduling trigger order based on the primary task scheduling unit set, extract the directed connection path between nodes according to the trigger sequence, construct the position index matrix of each pair of nodes in the path graph, and set the weight and connection direction of the edges between nodes to generate a task node dependency graph. S302: Call the path index and edge connection relationship in the task node dependency graph, identify the number of nodes in each path and use it as the path depth, identify the number of subsequent nodes connected to each node and use it as the branch factor, call the path depth and branch factor to calculate the original complexity score, combine the occupied time information of each node recorded in the scheduling table, and set the score adjustment coefficient according to the occupied time to adjust the original complexity score and obtain the node scheduling complexity score value. S303: Based on the node scheduling complexity score, filter the node score data, extract the node indexes with scores greater than the complexity score percentile threshold, and identify the corresponding aggregation distribution segments in the dependency graph. Based on the score difference between adjacent nodes, construct and trim continuous aggregation path intervals in the graph structure. Each node in the aggregation segment is a subgraph merging object, generating a dependency subgraph set. The complexity score percentile threshold is constructed by creating a sorted array based on the scheduling complexity scores of all nodes, and then using a grid search method to compare the scheduling performance of each percentile value on multiple task sets, selecting the score value at the 80th percentile as the threshold for filtering dense path nodes.

7. The multi-task processing method based on a large model according to claim 6, characterized in that, Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, the steps for extracting the most frequently invoked paths to form the main path set, dividing the remaining nodes into scheduling groups, and generating the task structure partitioning result are as follows: S401: Based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, extract the node number included in each path, collect the trigger frequency of the corresponding node in the scheduling log, identify the total number of calls for each node in each path, extract the path number with the highest trigger frequency in each path and construct a dataset to generate a high-frequency path trigger sequence. The high-frequency path trigger sequence is a set of frequent paths obtained by identifying the total trigger frequency of each node in the scheduling log of each path and selecting the set of path numbers with the highest trigger frequency. S402: Call the path number in the high-frequency path trigger sequence and obtain the corresponding all node numbers. Remove the main path node number from the node number set of the primary task scheduling unit set and extract the remaining node number set. Perform edge traversal on the resource dependency relationship between nodes within the remaining node number set. Perform hierarchical clustering analysis on the remaining nodes according to the connection directionality and resource channel number to generate a node scheduling partition structure. S403: Construct a path attribution index table based on the remaining nodes in the node scheduling partition structure, construct a group number mapping table using the set of main path nodes in the high-frequency path trigger sequence, merge the path attribution index table and the group number mapping table into a unified number index structure, and map it to the scheduling structure diagram to generate the task structure partitioning result; The node scheduling partitioning structure is formed by traversing the resource dependencies between the remaining nodes after removing the main path nodes, and then performing dependency clustering based on the resource channel number and connection directionality to form a grouping structure for task nodes.

8. The multi-task processing method based on a large model according to claim 1, characterized in that, The method further includes step S5: S5: Based on the task structure division results, collect the resource utilization rate and processing time of each task unit, construct a task processing behavior vector, evaluate the resource change rate, and adjust and sort the priority of each task unit to obtain a task execution priority sequence. The task execution priority sequence specifically consists of task unit number, sorting weight coefficient, and priority sorting index.

9. The multi-task processing method based on a large model according to claim 8, characterized in that, Based on the task structure partitioning results, the following steps are taken to collect the resource utilization rate and processing time of each task unit, construct a task processing behavior vector, evaluate the resource change rate, and adjust and sort the priority of each task unit to obtain the task execution priority sequence: S501: Based on the task structure division results, collect the resource utilization rate and processing time of each task unit, perform interval normalization processing and construct a joint data structure, combine the resource dimension and time dimension, form a two-dimensional behavior vector, and generate a task scheduling feature vector set. The task scheduling feature vector set is constructed by performing interval normalization on the resource utilization rate and processing time of each task unit and combining them into a two-dimensional behavior vector. S502: Call the resource dimension data in the task scheduling feature vector set, use the sliding window difference method to calculate the resource utilization difference value of consecutive task units in the same resource dimension, extract the absolute value sequence of the difference value as the fluctuation amplitude, calculate the fluctuation mean of each resource dimension, remove resource items with fluctuation values ​​less than the average offset value, calculate the resource dimension variation coefficient, summarize it into a stability index, and generate a resource utilization variation coefficient group. The sliding window difference method uses a window size of 3 and a sliding step size of 1 to perform stepwise difference calculations on the resource utilization rate sequence, extract the difference between adjacent task units and form a fluctuation amplitude sequence. The specific formula for calculating the coefficient of variation of the resource dimension is as follows: ； Calculate the coefficient of variation for the resource dimension; in, Let be the coefficient of variation of the j-th resource dimension. Let the resource utilization rate of the j-th resource dimension be the resource utilization rate in the k-th task unit. Let be the arithmetic mean of the resource utilization rate of the j-th resource dimension across all n task units. This represents the total number of consecutive task units participating in resource utilization statistics. A minimal positive real perturbation factor, existing in constant form, is used in the denominator to prevent division by zero errors. This serves as the index identifier for the resource dimension. This is the index identifier for the j-th resource dimension. Time-series index identifier for task units; The coefficient of variation for the resource dimension is obtained by statistically analyzing the usage rate sequence of each type of resource in each task unit and calculating the ratio of the standard deviation to the arithmetic mean of the usage rate sequence, which is used to reflect the degree of fluctuation in resource usage. The resource utilization coefficient of variation set is a numerical set constructed by extracting the coefficient of variation of the utilization rate for each resource dimension and is used to reflect the stability and fluctuation range of resource utilization. S503: Based on the resource utilization variation coefficient group and the normalized value of processing time in the task scheduling feature vector set, calculate the time proportion weight, and use the resource fluctuation value and processing time value to adjust and sort the priority of each task unit to generate a task execution priority sequence.

10. A multi-task processing system based on a large model, characterized in that, The system is used to implement the multi-task processing method based on a large model as described in any one of claims 1-9, and the system comprises: The resource conflict analysis module uses a large model to obtain the resource type code, access frequency and occupation duration of task nodes, calculates the Euclidean distance of resource demand vectors and constructs a conflict intensity matrix, combines a set threshold to filter node pairs, calls a density clustering algorithm to evaluate density and judge structural conflicts, and generates a recombinable task node structure to be passed to the scheduling structure generation module. The scheduling structure generation module divides the upper and lower bound nodes according to the recombinable task node structure, calls the conflict intensity median deviation parameter to extract the upper bound node to construct an independent unit, matches the lower bound node and merges them according to the resource type dissimilarity rate, and generates a primary task scheduling unit set which is then passed to the dependency graph construction module. The dependency graph construction module extracts the resource access channel number and scheduling trigger order based on the primary task scheduling unit set, constructs the dependency graph, calls path depth and branch factor to calculate the complexity score, adjusts the score weight in combination with the occupied time, prunes dense areas and performs subgraph and node aggregation, generates a dependency subgraph set and passes it to the structure partitioning module. The structure partitioning module identifies the path call frequency and extracts the most frequent path based on the node identifier mapping relationship between the dependency subgraph set and the primary task scheduling unit set, constructs the main path set, partitions the remaining nodes and forms scheduling groups, and generates the task structure partitioning result, which is then passed to the priority sorting module. The priority sorting module collects the resource utilization rate and processing time of each task unit based on the task structure division result to construct a task processing behavior vector, evaluates the resource change rate, calls the priority adjustment function to sort, and generates a task execution priority sequence.

Images