A cross-domain computing power resource dynamic scheduling method and system and a medium

Abstract

The application discloses a kind of cross-domain computing resource dynamic scheduling method, system and medium, comprising: collecting to be scheduled task data, candidate computing resource data and cross-domain link data, constructs cross-domain scheduling data;Calculate data migration cost, computing execution cost and link occupation cost, establish collaborative cost matrix;Quantitative processing is carried out to collaborative cost matrix, and obtain quantitative collaborative cost matrix and quantitative residual matrix;Establish task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix and residual correction weight matrix, and execute improved Sinkhorn algorithm to obtain task resource soft matching matrix;According to resource and link constraint, generate task resource hard matching table, establish cross-domain communication connection, allocate computing power quota and send task data in segments, obtain cross-domain computing resource dynamic scheduling result.The application introduces quantitative residual correction and cross-domain collaborative cost into task resource matching process, and forms dynamic scheduling path.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to a method, system and medium for dynamic scheduling of cross-domain computing resources. Background Technology

[0002] With the continuous integration of industrial internet, cloud computing, and edge computing infrastructure, computing resources are gradually exhibiting characteristics of cross-regional, cross-network domain, and heterogeneous distribution. Computing tasks for data analysis, model training, real-time control, and online services require task selection, resource allocation, data transmission, and execution coordination across multiple computing nodes. Existing cross-domain computing scheduling primarily employs centralized scheduling platforms that collect data on task size, computing power requirements, node load, resource idle rate, task queue length, network latency, and bandwidth usage status, and select candidate computing resources based on rule-based sorting, weighted scoring, or load balancing strategies.

[0003] Existing technologies typically map task requirements to node resource status using several evaluation indicators, then calculate a comprehensive score according to preset weights, or perform tiered screening based on the node's current load and link status. Some solutions assess data transmission latency and available bandwidth before task distribution and switch to other nodes when resources are insufficient. While these methods can complete task allocation in a single-domain environment or under relatively stable network conditions, the evaluation indicators often participate in the scoring independently, and the coupling relationship between task data location, task communication dependencies, resource supply capacity, and cross-domain link status is not uniformly characterized.

[0004] During cross-domain execution, task data migration, computing power consumption, and link contention all simultaneously affect the scheduling results. Static scoring based solely on resource idle rate, load rate, network latency, bandwidth utilization, and queue length fails to reflect the collaborative costs incurred after task allocation. Different metrics have varying dimensions, numerical ranges, and rates of change; fixed weights struggle to adapt to link fluctuations and resource status changes in a timely manner. This can easily lead to high-scoring candidate resources incurring significant data migration overhead, link congestion, or resource supply imbalances during actual execution. Furthermore, the lack of unified handling of soft matching relationships between candidate resources, total supply and demand constraints, and the implementation of scheduling results impacts the continuity and executability of cross-domain task allocation.

[0005] Therefore, how to provide a method, system, and medium for dynamic scheduling of cross-domain computing resources is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method, system, and medium for dynamic scheduling of cross-domain computing resources. This invention achieves coordinated scheduling of tasks, resources, and link status through collaborative cost quantification and improved sinkhorn matching.

[0007] A cross-domain computing resource dynamic scheduling method according to an embodiment of the present invention includes:

[0008] Collect data on tasks to be scheduled, candidate computing resources, and cross-domain link data, perform normalization processing and corresponding correlation to obtain cross-domain scheduling data;

[0009] Calculate the data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, and establish a collaborative cost matrix;

[0010] The collaborative cost matrix is ​​quantized to obtain the quantized collaborative cost matrix and the quantized residual matrix;

[0011] Based on the task data to be scheduled, candidate computing resource data, quantized collaborative cost matrix and quantized residual matrix, establish the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix and residual correction weight matrix, and complete the total balance;

[0012] The improved Sinkhorn algorithm is executed based on the task requirement marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix to obtain the task resource soft matching matrix.

[0013] The task resource matching relationship is determined based on the task resource soft matching matrix, and the task resource hard matching table is obtained.

[0014] Establish a cross-domain communication connection based on the task resource hard matching table, allocate computing power quotas and send data of tasks to be scheduled, trigger candidate computing power resources to execute the tasks to be scheduled, and obtain the dynamic scheduling results of cross-domain computing power resources.

[0015] Optionally, the collection of task data to be scheduled, candidate computing resource data, and cross-domain link data is normalized and correlated accordingly, specifically as follows:

[0016] Data on tasks to be scheduled, candidate computing resources, and cross-domain links are collected separately according to the collection time. The data on tasks to be scheduled includes the amount of task data, the amount of computing power required by the task, and the communication dependency of the task. The data on candidate computing resources includes available computing power, load rate, and task queue capacity. The data on cross-domain links includes network latency, available bandwidth of the link, percentage of available bandwidth of the link, and bandwidth utilization rate of the link.

[0017] Based on the data of the tasks to be scheduled, the candidate computing resources data, and the cross-domain link data, the collection time correction is performed, and the associated time width is calculated to obtain the resource correction data, link correction data, and associated time width;

[0018] Based on the data of tasks to be scheduled, resource correction data, link correction data, and associated time span, candidate associated data are obtained by filtering the data.

[0019] Validate the data items in the candidate related data and handle missing data to obtain complete related data;

[0020] Normalize the data items in the complete associated data to obtain normalized associated data;

[0021] The normalized correlation data is arranged in the order of the tasks to be scheduled, candidate computing resources, and cross-domain links to obtain cross-domain scheduling data.

[0022] Optionally, the step of calculating the data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, and establishing a collaborative cost matrix, specifically involves:

[0023] Based on the correspondence between the tasks to be scheduled, candidate computing resources, and cross-domain links, normalized task data volume, task computing power requirement, task communication dependency, available computing power, load rate, network latency, available link bandwidth, percentage of available link bandwidth, and link bandwidth occupancy rate are extracted from the cross-domain scheduling data to obtain candidate cost data.

[0024] Calculate the data migration cost, computing power execution cost, and link occupancy cost based on the candidate cost data;

[0025] The migration modulation cost and link modulation cost are calculated based on the data migration cost, link occupancy cost, normalized task communication dependency, and normalized link bandwidth occupancy.

[0026] Calculate the migration weight value, computing power weight value, and link weight value based on the migration modulation cost, computing power execution cost, link modulation cost, and candidate cost data.

[0027] Calculate the corresponding cost weight based on the migration weight value, computing power weight value, and link weight value;

[0028] The collaborative cost value is calculated based on the migration modulation cost, computing power execution cost, link modulation cost and corresponding cost weights, and the collaborative cost values ​​are arranged in the order of the tasks to be scheduled and the candidate computing power resources to obtain the collaborative cost matrix.

[0029] Optionally, the quantization process performed on the collaborative cost matrix specifically includes:

[0030] Extract the collaborative cost value and the corresponding link available bandwidth percentage from the collaborative cost matrix, and sort them to obtain the cost sorting sequence and bandwidth percentage sequence.

[0031] Input the cost sorting sequence and bandwidth percentage sequence into the Lloyd-Max algorithm to divide the continuous subsequence and calculate the initial quantization representative value and the initial quantization interval boundary;

[0032] The boundary correction amount is calculated based on the bandwidth ratio sequence, the initial quantization representative value, and the initial quantization interval boundary. The initial quantization interval boundary is then corrected to obtain the corrected quantization interval boundary.

[0033] The cost sorting sequence is redefined based on the corrected quantization interval boundaries, and the quantization representative value of each quantization interval is updated.

[0034] The quantization interval boundary is recalculated and corrected based on the updated quantization representative value until the difference between two adjacent quantization representative values ​​meets the quantization convergence threshold, thus obtaining the target quantization representative value and the target quantization interval boundary.

[0035] The quantization collaborative cost matrix and quantization residual matrix are generated based on the target quantization representative value, the target quantization interval boundary, and the collaborative cost matrix.

[0036] Optionally, the establishment of the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix specifically involves:

[0037] Calculate the task requirement value corresponding to each task to be scheduled based on the task data to be scheduled, and calculate the resource supply value corresponding to each candidate computing resource based on the candidate computing resource data.

[0038] Arrange the task demand values ​​and resource supply values ​​according to the order of the tasks to be scheduled and the candidate computing resources, respectively, to obtain the task demand marginal vector and resource supply marginal vector, and calculate the total demand and total supply respectively.

[0039] Perform total balance based on total demand and total supply, and update the task demand marginal vector, resource supply marginal vector, quantified collaborative cost matrix, and quantified residual matrix;

[0040] Calculate the entropy regularization parameter based on the updated quantized residual matrix and cross-domain link data;

[0041] Calculate the initial transmission kernel matrix based on the updated quantized cooperative cost matrix and entropy regularization parameter;

[0042] Calculate the residual correction weight matrix based on the updated quantization residual matrix.

[0043] Optionally, the improved Sinkhorn algorithm is specifically as follows:

[0044] Calculate the residual corrected transfer kernel matrix based on the initial transfer kernel matrix and the residual corrected weight matrix;

[0045] Calculate the task residual weight vector and the resource residual weight vector based on the residual correction weight matrix;

[0046] Calculate the task correction bias and resource correction bias based on the residual correction transmission kernel matrix, task requirement marginal vector, resource supply marginal vector, task residual weight vector, and resource residual weight vector;

[0047] Compare the task correction deviation and the resource correction deviation. When the task correction deviation is greater than the resource correction deviation, generate a row scaling factor. When the resource correction deviation is not less than the task correction deviation, generate a column scaling factor.

[0048] When generating row scaling factors, the residual correction transmission kernel matrix is ​​updated based on the row scaling factors. When generating column scaling factors, the residual correction transmission kernel matrix is ​​updated based on the column scaling factors to obtain the updated transmission kernel matrix.

[0049] The updated transmission kernel matrix is ​​used as the residual corrected transmission kernel matrix to recalculate the task correction bias and resource correction bias. When both the task correction bias and resource correction bias meet the convergence threshold, the updated transmission kernel matrix is ​​determined as the task-resource soft matching matrix.

[0050] Optionally, determining the task resource matching relationship based on the task resource soft matching matrix specifically involves:

[0051] Generate a task processing sequence based on the task resource soft matching matrix, the task requirement marginal vector, and the task data to be scheduled.

[0052] Generate a candidate sequence of task resources based on the task processing sequence and the task resource soft matching matrix;

[0053] Calculate the link occupancy requirement value based on the data of the tasks to be scheduled;

[0054] Candidate computing resources are verified based on the candidate sequence of task resources, candidate computing resource data, cross-domain link data, and link occupancy requirements, and the matching relationship of task resources is determined.

[0055] When a candidate computing resource fails the verification, the task resource candidate sequence is updated;

[0056] When the number of remaining candidate computing resources is 0, the unmatched status of the corresponding task to be scheduled is recorded. When the task resource matching relationship is determined, the candidate computing resource data and cross-domain link data are updated, and the task resource matching relationship and unmatched status are collected according to the task processing sequence to obtain the task resource hard matching table.

[0057] Optionally, the step of establishing a cross-domain communication connection based on the task resource hard matching table, allocating computing power quotas and sending data of tasks to be scheduled, and triggering candidate computing power resources to execute tasks to be scheduled, specifically involves:

[0058] Generate a task execution sequence based on the task resource hard match table;

[0059] Establish cross-domain communication connections between the tasks to be scheduled and the candidate computing resources based on the task execution sequence, and calculate the connection waiting time.

[0060] Resource status processing is performed based on the establishment status of the cross-domain communication connection. When the cross-domain communication connection is established within the connection waiting time, the computing power quota is allocated according to the computing power requirement of the task. When the cross-domain communication connection is not established within the connection waiting time, the available computing power and task queue capacity of the corresponding candidate computing power resources are restored, the available bandwidth of the corresponding cross-domain link is restored, and the corresponding task to be scheduled is recorded as unmatched.

[0061] The basic transmission volume is calculated based on the data of the tasks to be scheduled and the cross-domain link data, and the data of the tasks to be scheduled is sent in segments according to the basic transmission volume.

[0062] After each data transmission is completed, update the cross-domain link data and recalculate the base transmission volume;

[0063] The tasks to be scheduled are executed based on the segmented data of the tasks to be scheduled and the computing power quota, and the task execution results are obtained.

[0064] By collecting task execution results and unmatched states according to the task execution sequence, the dynamic scheduling results of cross-domain computing resources are obtained.

[0065] A cross-domain computing resource dynamic scheduling system includes:

[0066] The scheduling data processing module collects data on tasks to be scheduled, candidate computing resources, and cross-domain link data, and processes them to obtain cross-domain scheduling data.

[0067] The collaborative cost construction module calculates data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, generates corresponding cost weights, and establishes a collaborative cost matrix.

[0068] The quantization processing module performs quantization processing on the collaborative cost matrix to obtain the quantized collaborative cost matrix and the quantized residual matrix.

[0069] The transmission parameter construction module establishes the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix based on the task data to be scheduled, candidate computing power resource data, quantized collaborative cost matrix, and quantized residual matrix, and performs total balance.

[0070] The soft matching solution module executes the improved Sinkhorn algorithm based on the task requirement margin vector, resource supply margin vector, initial transmission kernel matrix, and residual correction weight matrix to obtain the task resource soft matching matrix.

[0071] The hard matching generation module verifies candidate computing resources and cross-domain links based on the task resource soft matching matrix, determines the task resource matching relationship, and generates a task resource hard matching table.

[0072] The scheduling and execution module establishes a cross-domain communication connection based on the task resource hard matching table, allocates computing power quotas, sends the data of the tasks to be scheduled in segments, and triggers the candidate computing power resources to execute the tasks to be scheduled, thereby obtaining the dynamic scheduling results of cross-domain computing power resources.

[0073] A computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements a cross-domain computing resource dynamic scheduling method.

[0074] The beneficial effects of this invention are:

[0075] (1) This invention associates task data volume, task computing power requirement, task communication dependency, available computing power, load rate, network latency, available bandwidth ratio of links and bandwidth occupancy rate, and constructs a data migration cost, computing power execution cost, link occupancy cost and coordination cost matrix accordingly, thereby realizing a unified representation of task demand, resource supply and cross-domain link status, thereby reducing the matching deviation caused by scheduling based on a single resource indicator.

[0076] (2) This invention generates a quantized collaborative cost matrix and a quantized residual matrix by performing quantization processing on the collaborative cost matrix and using the proportion of available bandwidth of the link to correct the boundary of the quantization interval. This achieves the synchronous preservation of the collaborative cost distribution and quantization error, thereby providing input data that takes into account both cost level and local deviation for subsequent task resource matching.

[0077] (3) The improved Sinkhorn algorithm in this invention combines the initial transmission kernel matrix with the residual correction weight matrix element by element, and selects row scaling or column scaling according to the task correction deviation and resource correction deviation. At the same time, resource verification, state rollback and dynamic segmented transmission are introduced in the hard matching and scheduling execution stages, realizing task resource matching and cross-domain execution connection under the supply and demand marginal constraints, thereby improving the executability and stability of scheduling results under cross-domain link fluctuation conditions. Attached Figure Description

[0078] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0079] Figure 1 This is a flowchart of a cross-domain computing resource dynamic scheduling method proposed in this invention;

[0080] Figure 2This is a system architecture diagram of a cross-domain computing resource dynamic scheduling system proposed in this invention;

[0081] Figure 3 The flowchart of the improved Sinkhorn algorithm for a cross-domain computing resource dynamic scheduling method proposed in this invention is shown below.

[0082] Figure 4 This is a schematic diagram of the task resource matching results of a cross-domain computing power resource dynamic scheduling method proposed in this invention. Detailed Implementation

[0083] Example 1: Reference Figures 1-3 A method for dynamic scheduling of cross-domain computing resources, comprising:

[0084] Collect data on tasks to be scheduled, candidate computing power resources, and cross-domain link data, perform normalization processing and corresponding correlation to obtain cross-domain scheduling data. The data on tasks to be scheduled includes task data volume, task computing power requirement, and task communication dependency. The candidate computing power resources data includes available computing power, load rate, and task queue capacity. The cross-domain link data includes network latency, available link bandwidth, percentage of available link bandwidth, and link bandwidth utilization rate.

[0085] Based on cross-domain scheduling data, the data migration cost, computing power execution cost, and link occupancy cost are calculated respectively, and a collaborative cost matrix is ​​established. The data migration cost is obtained by dividing the normalized task data volume by the normalized available link bandwidth and adding it to the normalized network latency. The computing power execution cost is obtained by dividing the normalized task computing power requirement by the normalized available resource computing power and adding it to the normalized load rate. The link occupancy cost is obtained by dividing the normalized task communication dependency by the normalized available link bandwidth ratio. The collaborative cost matrix is ​​obtained by multiplying the data migration cost by the normalized task communication dependency, multiplying the link occupancy cost by the normalized link bandwidth occupancy rate, and then performing a weighted summation of the two products with the computing power execution cost according to the corresponding cost weights.

[0086] The collaborative cost matrix is ​​subjected to Lloyd-Max quantization to obtain the quantized collaborative cost matrix and the quantized residual matrix. The Lloyd-Max quantization process corrects the quantization interval boundary according to the proportion of available bandwidth of the link.

[0087] Based on the task data to be scheduled, candidate computing resource data, quantized collaborative cost matrix, and quantized residual matrix, a task demand marginal vector, a resource supply marginal vector, an initial transmission kernel matrix, and a residual correction weight matrix are established, and total balance is completed. The total balance adds virtual task nodes when the total supply is greater than the total demand, and adds virtual resource nodes when the total demand is greater than the total supply. The initial transmission kernel matrix is ​​obtained by multiplying the quantized collaborative cost matrix with the reciprocal of the entropy regularization parameter and then performing negative exponential mapping. The residual correction weight matrix is ​​obtained by normalizing the absolute value of the quantized residual matrix and then performing negative exponential mapping.

[0088] Based on the task requirement marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix, the improved Sinkhorn algorithm is executed to obtain the task resource soft matching matrix. The improved Sinkhorn algorithm multiplies the initial transmission kernel matrix and the residual correction weight matrix element by element, and alternately performs row scaling and column scaling until the marginal deviation meets the convergence threshold.

[0089] The task resource matching relationship is determined based on the task resource soft matching matrix to obtain the task resource hard matching table. The task resource matching relationship is verified in descending order of soft matching coefficient to check the available computing power, task queue capacity and link available bandwidth, and the available computing power and link available bandwidth corresponding to the matched resources are deducted.

[0090] Establish a cross-domain communication connection based on the task resource hard matching table, allocate computing power quotas and send data of tasks to be scheduled, trigger candidate computing power resources to execute the tasks to be scheduled, and obtain the dynamic scheduling results of cross-domain computing power resources.

[0091] In this embodiment, the collection of data on tasks to be scheduled, candidate computing resources, and cross-domain link data, followed by normalization processing and corresponding association, specifically involves:

[0092] Data on tasks to be scheduled, candidate computing resources, and cross-domain links are collected separately according to the collection time. The data on tasks to be scheduled includes the amount of task data, the amount of computing power required by the task, and the communication dependency of the task. The data on candidate computing resources includes available computing power, load rate, and task queue capacity. The data on cross-domain links includes network latency, available bandwidth of the link, percentage of available bandwidth of the link, and bandwidth utilization rate of the link.

[0093] Based on the data of the tasks to be scheduled, the candidate computing resources data, and the cross-domain link data, the collection time correction is performed, and the associated time width is calculated to obtain the resource correction data, link correction data, and associated time width;

[0094] In Example 1, the network latency corresponding to the candidate computing resources is obtained based on the correspondence between candidate computing resources and cross-domain links. The resource correction time is obtained by subtracting the corresponding network latency from the acquisition time of the candidate computing resources data, and the candidate computing resources data is associated with the resource correction time to form resource correction data. The link correction time is obtained by subtracting the corresponding network latency from the acquisition time of the cross-domain link data, and the cross-domain link data is associated with the link correction time to form link correction data. The resource correction data and link correction data are arranged according to the resource correction time and link correction time to complete the acquisition time correction. The associated bandwidth is obtained by adding the network latency to the product of the network latency and the link bandwidth occupancy rate.

[0095] Data is filtered based on the task data to be scheduled, resource correction data, link correction data, and associated time width to obtain candidate associated data. The data filtering is based on the collection time of the task data to be scheduled, and the resource correction data and link correction data with the smallest correction time difference are selected within the associated time width.

[0096] In Example 1, the correction time difference is obtained by subtracting the correction time corresponding to each resource correction data and each link correction data within the associated time span from the collection time of the task data to be scheduled, and taking the absolute value of the difference.

[0097] The data items in the candidate associated data are verified and the missing data is processed to obtain complete associated data. When the percentage of available bandwidth of the link is missing, the link correction data with the second smallest correction time difference is selected for replacement. When other data items are missing, the data change rate is calculated based on the two valid values ​​before the missing time. The data change rate, the time difference between the missing time and the next valid value, and the percentage of available bandwidth of the link are multiplied together, and the product is added to the next valid value. Then the result of the addition is used as the replacement value for other missing items.

[0098] Normalization is performed on the data items in the complete associated data to obtain normalized associated data. In the normalization process, when the maximum value of a data item is greater than the minimum value, the difference between the data item and the minimum value is divided by the difference between the maximum value and the minimum value and then 1 is added. When the maximum value is equal to the minimum value, the data item is set to 1.

[0099] In Example 1, all complete associated data formed within the same scheduling batch are combined into a normalized statistical set; the corresponding data values ​​in the normalized statistical set are collected according to the data item name to obtain the data value sequence of each data item; the maximum and minimum values ​​are determined from each data value sequence.

[0100] The normalized correlation data is arranged in the order of the tasks to be scheduled, candidate computing resources, and cross-domain links to obtain cross-domain scheduling data.

[0101] In this embodiment, the step of calculating the data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, and establishing a collaborative cost matrix, specifically involves:

[0102] Based on the correspondence between the tasks to be scheduled, candidate computing resources, and cross-domain links, normalized task data volume, task computing power requirement, task communication dependency, available computing power, load rate, network latency, available link bandwidth, percentage of available link bandwidth, and link bandwidth occupancy rate are extracted from the cross-domain scheduling data to obtain candidate cost data.

[0103] Calculate the data migration cost, computing power execution cost, and link occupancy cost based on the candidate cost data;

[0104] In Example 1, the data migration cost is obtained by dividing the normalized task data volume by the normalized available link bandwidth and adding it to the normalized network latency; the computing power execution cost is obtained by dividing the normalized task computing power requirement by the normalized available resource computing power and adding it to the normalized load rate; and the link occupancy cost is obtained by dividing the normalized task communication dependency by the normalized proportion of available link bandwidth.

[0105] The migration modulation cost and link modulation cost are calculated based on the data migration cost, link occupancy cost, normalized task communication dependency, and normalized link bandwidth occupancy rate. The migration modulation cost is obtained by multiplying the data migration cost by the normalized task communication dependency, and the link modulation cost is obtained by multiplying the link occupancy cost by the normalized link bandwidth occupancy rate.

[0106] Calculate the migration weight value, computing power weight value, and link weight value based on the migration modulation cost, computing power execution cost, link modulation cost, and candidate cost data.

[0107] In Example 1, the migration weight value is obtained by multiplying the migration modulation cost by the normalized network latency, the computing power weight value is obtained by multiplying the computing power execution cost by the normalized load rate, and the link weight value is obtained by multiplying the link modulation cost by the normalized link bandwidth occupancy rate and then dividing by the normalized link available bandwidth percentage.

[0108] The corresponding cost weight is calculated based on the migration weight value, computing power weight value, and link weight value. The corresponding cost weight is obtained by dividing the migration weight value, computing power weight value, and link weight value by the sum of the three values ​​respectively.

[0109] The collaborative cost value is calculated based on the migration modulation cost, computing power execution cost, link modulation cost and corresponding cost weights, and the collaborative cost values ​​are arranged in the order of the tasks to be scheduled and the candidate computing power resources to obtain the collaborative cost matrix. The collaborative cost value is obtained by multiplying the migration modulation cost, computing power execution cost and link modulation cost by their corresponding cost weights respectively, and then adding the products.

[0110] In this embodiment, the quantization process performed on the collaborative cost matrix specifically includes:

[0111] Extract the collaborative cost value and the corresponding link available bandwidth percentage from the collaborative cost matrix, and sort them to obtain the cost sorting sequence and bandwidth percentage sequence.

[0112] In Example 1, the collaborative cost value is extracted item by item according to the row and column positions of the collaborative cost matrix. The row position corresponds to the task to be scheduled, and the column position corresponds to the candidate computing power resource. Based on the correspondence between the task to be scheduled, the candidate computing power resource, and the cross-domain link, the available bandwidth ratio of the link corresponding to each collaborative cost value is determined from the cross-domain scheduling data. The collaborative cost value and the corresponding available bandwidth ratio of the link are paired to form a one-to-one data pair. The data pairs are arranged in ascending order of collaborative cost value, and the sorted collaborative cost values ​​are collected in sequence to obtain the cost sorting sequence. The available bandwidth ratio of the link is collected in the same sorting order to obtain the bandwidth ratio sequence.

[0113] Input the cost sorting sequence and bandwidth percentage sequence into the Lloyd-Max algorithm to divide the continuous subsequence and calculate the initial quantization representative value and the initial quantization interval boundary;

[0114] In Example 1, the number of co-cost values ​​and the number of quantization intervals in the cost ranking sequence are statistically analyzed. The number of co-cost values ​​is divided by the number of quantization intervals, and the integer quotient is determined as the basic length of the continuous subsequence. Starting from the first continuous subsequence, one co-cost value is added to each continuous subsequence with the same number of remainders. The co-cost values ​​are continuously divided according to the order of the cost ranking sequence, and the bandwidth proportion sequence is synchronously divided according to the same position to obtain continuous subsequences. The ratio of the sum of co-cost values ​​to the number of co-cost values ​​in each continuous subsequence is calculated to obtain the initial quantization representative value. The initial quantization representative values ​​are arranged in ascending order, and half of the sum of adjacent initial quantization representative values ​​is calculated to obtain the initial quantization interval boundary.

[0115] The boundary correction amount is calculated based on the bandwidth ratio sequence, the initial quantization representative value, and the initial quantization interval boundary. The initial quantization interval boundary is then corrected to obtain the corrected quantization interval boundary.

[0116] In Example 1, the cooperative cost value on the left and right sides of the initial quantization interval is determined according to the boundary, and the available bandwidth ratio of the links on both sides of the boundary is determined from the bandwidth ratio sequence according to the same order. The sum of the available bandwidth ratios of the links on the left and right sides of the boundary is divided by the corresponding quantity to obtain the average bandwidth ratio of the left side and the average bandwidth ratio of the right side. The average bandwidth ratio of the left side is subtracted from the average bandwidth ratio of the right side to obtain the difference of the average bandwidth ratios. The average bandwidth ratios of the left side and the right side are added to obtain the sum of the average bandwidth ratios. The difference of the average bandwidth ratios is divided by the sum of the average bandwidth ratios and multiplied by half of the difference between adjacent initial quantization representative values ​​to obtain the boundary correction amount.

[0117] The cost sorting sequence is re-divided according to the corrected quantization interval boundary, and the quantization representative value of each quantization interval is updated. The quantization representative value is obtained by multiplying the collaborative cost value by the reciprocal of the corresponding link available bandwidth ratio, summing the results, and dividing by the sum of the reciprocals of the link available bandwidth ratio.

[0118] The quantization interval boundary is recalculated and corrected based on the updated quantization representative value until the difference between two adjacent quantization representative values ​​meets the quantization convergence threshold, thus obtaining the target quantization representative value and the target quantization interval boundary. The quantization convergence threshold is obtained by arranging the initial quantization representative values ​​in ascending order, calculating the difference between adjacent initial quantization representative values ​​to obtain a representative value interval sequence, dividing the sum of the differences in the representative value interval sequence by the number of differences to obtain the average representative value interval, and dividing the average representative value interval by the number of quantization intervals to obtain the quantization convergence threshold. When the average representative value interval is equal to 0, the quantization convergence threshold is set to 0.001.

[0119] In Example 1, the updated quantization representative values ​​are arranged in ascending order. Half of the sum of adjacent updated quantization representative values ​​is calculated to obtain the candidate quantization interval boundary. Based on the candidate quantization interval boundary, the available bandwidth ratio of the links on both sides of the boundary is determined, and the average bandwidth ratio on the left and right sides is calculated respectively. The difference between the average bandwidth ratio on the left and the average bandwidth ratio on the right is divided by the sum of the two, and multiplied by half of the difference between adjacent updated quantization representative values ​​to obtain the update boundary correction amount. The candidate quantization interval boundary is added to the update boundary correction amount to obtain the corrected quantization interval boundary.

[0120] The quantization collaborative cost matrix and the quantization residual matrix are generated based on the target quantization representative value, the target quantization interval boundary, and the collaborative cost matrix. The quantization collaborative cost matrix is ​​obtained by replacing the collaborative cost value with the target quantization representative value within the corresponding target quantization interval and arranging them in the order of the tasks to be scheduled and the candidate computing resources. The quantization residual matrix is ​​obtained by subtracting the collaborative cost matrix from the quantization collaborative cost matrix element by element.

[0121] In this embodiment, the establishment of the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix specifically involves:

[0122] The task demand value corresponding to each task to be scheduled is calculated based on the task data to be scheduled, and the resource supply value corresponding to each candidate computing power resource is calculated based on the candidate computing power resource data. The task demand value is obtained by multiplying the normalized task computing power demand by the normalized task communication dependency. The resource supply value is obtained by multiplying the normalized available computing power by the normalized task queue capacity and then dividing by the normalized load rate.

[0123] Arrange the task demand values ​​and resource supply values ​​according to the order of the tasks to be scheduled and the candidate computing resources, respectively, to obtain the task demand marginal vector and resource supply marginal vector, and calculate the total demand and total supply respectively.

[0124] In Example 1, according to the order of the tasks to be scheduled in the cross-domain scheduling data, the task demand values ​​corresponding to each task to be scheduled are sequentially written into the task demand marginal vector; according to the order of the candidate computing resources in the cross-domain scheduling data, the resource supply values ​​corresponding to each candidate computing resources are sequentially written into the resource supply marginal vector; all task demand values ​​in the task demand marginal vector are added together to obtain the total demand; all resource supply values ​​in the resource supply marginal vector are added together to obtain the total supply.

[0125] Based on the total demand and total supply, total balance is performed, and the task demand marginal vector, resource supply marginal vector, quantitative collaboration cost matrix, and quantitative residual matrix are updated. When the total supply is greater than the total demand, virtual task nodes are added, and the difference between the total supply and the total demand is determined as the task demand value of the virtual task node. Corresponding rows with all values ​​of 0 are added to the quantitative collaboration cost matrix and the quantitative residual matrix. When the total demand is greater than the total supply, virtual resource nodes are added, and the difference between the total demand and the total supply is determined as the resource supply value of the virtual resource node. Corresponding columns with all values ​​of 0 are added to the quantitative collaboration cost matrix and the quantitative residual matrix. When the total demand equals the total supply, the task demand marginal vector, resource supply marginal vector, quantitative collaboration cost matrix, and quantitative residual matrix remain unchanged.

[0126] In Example 1, when the total supply exceeds the total demand, the difference between the total supply and the total demand is appended to the end of the task demand marginal vector, and a 0 value with the same length as the resource supply marginal vector is written into the last row of the quantized collaboration cost matrix and the quantized residual matrix; when the total demand exceeds the total supply, the difference between the total demand and the total supply is appended to the end of the resource supply marginal vector, and a 0 value with the same length as the task demand marginal vector is written into the last column of the quantized collaboration cost matrix and the quantized residual matrix; when the total demand equals the total supply, the task demand marginal vector, the resource supply marginal vector, the quantized collaboration cost matrix, and the quantized residual matrix remain unchanged.

[0127] The entropy regularization parameter is calculated based on the updated quantization residual matrix and cross-domain link data. The entropy regularization parameter is obtained by adding the mean of the absolute values ​​of the updated quantization residual matrix to the reciprocal of the mean of the normalized link available bandwidth ratio.

[0128] The initial transfer kernel matrix is ​​calculated based on the updated quantization co-cost matrix and the entropy regularization parameter. The initial transfer kernel matrix is ​​obtained by performing a negative exponential mapping after multiplying the updated quantization co-cost matrix by the reciprocal of the entropy regularization parameter.

[0129] The residual correction weight matrix is ​​calculated based on the updated quantization residual matrix. The residual correction weight matrix is ​​obtained by normalizing the absolute value of the updated quantization residual matrix and then performing negative exponential mapping.

[0130] In this embodiment, the improved Sinkhorn algorithm is specifically as follows:

[0131] The residual corrected transmission kernel matrix is ​​calculated based on the initial transmission kernel matrix and the residual corrected weight matrix. The residual corrected transmission kernel matrix is ​​obtained by multiplying the initial transmission kernel matrix and the residual corrected weight matrix element by element.

[0132] The task residual weight vector and the resource residual weight vector are calculated based on the residual correction weight matrix. The task residual weight vector is obtained by calculating the average of the differences between 1 and each element in the residual correction weight matrix by row and then adding 1. The resource residual weight vector is obtained by calculating the average of the differences between 1 and each element in the residual correction weight matrix by column and then adding 1.

[0133] Calculate the task correction bias and resource correction bias based on the residual correction transmission kernel matrix, task requirement marginal vector, resource supply marginal vector, task residual weight vector, and resource residual weight vector;

[0134] In Example 1, the task correction deviation is obtained by calculating the absolute value of the difference between the row sum vector of the residual correction transmission kernel matrix and the task demand margin vector, multiplying the absolute value of the difference element-wise with the task residual weight vector, and selecting the maximum value. The resource correction deviation is obtained by calculating the absolute value of the difference between the column sum vector of the residual correction transmission kernel matrix and the resource supply margin vector, multiplying the absolute value of the difference element-wise with the resource residual weight vector, and selecting the maximum value.

[0135] Compare the task correction deviation and the resource correction deviation. When the task correction deviation is greater than the resource correction deviation, generate a row scaling factor. When the resource correction deviation is not less than the task correction deviation, generate a column scaling factor.

[0136] In Example 1, when the task correction deviation is greater than the resource correction deviation, the sum of the elements in each row of the residual correction transmission kernel matrix is ​​calculated, and the task demand value in the task demand margin vector is divided by the sum of the elements in the corresponding row to obtain the scaling factor for each row; when the resource correction deviation is not less than the task correction deviation, the sum of the elements in each column of the residual correction transmission kernel matrix is ​​calculated, and the resource supply value in the resource supply margin vector is divided by the sum of the elements in the corresponding column to obtain the scaling factor for each column.

[0137] When generating row scaling factors, the residual correction transmission kernel matrix is ​​updated based on the row scaling factors. When generating column scaling factors, the residual correction transmission kernel matrix is ​​updated based on the column scaling factors to obtain the updated transmission kernel matrix. The updated transmission kernel matrix is ​​obtained by multiplying the row scaling factors by performing an exponentiation operation on each row scaling factor according to the corresponding task residual weight when generating row scaling factors, and by multiplying the column scaling factors by performing an exponentiation operation on each column scaling factor according to the corresponding resource residual weight.

[0138] The updated transmission kernel matrix is ​​used as the residual corrected transmission kernel matrix to recalculate the task correction bias and resource correction bias. When both the task correction bias and resource correction bias meet the convergence threshold, the updated transmission kernel matrix is ​​determined as the task-resource soft matching matrix.

[0139] In Example 1, after the initial calculation of the task correction deviation and the resource correction deviation, the task correction deviation and the resource correction deviation are added together to obtain the total deviation value; the number of elements in the task demand marginal vector is added together with the number of elements in the resource supply marginal vector to obtain the total number of marginal elements; the total deviation value is divided by the total number of marginal elements to obtain the convergence threshold; when the total deviation value is equal to 0, the convergence threshold is set to 0.001.

[0140] In this embodiment, determining the task resource matching relationship based on the task resource soft matching matrix specifically involves:

[0141] A task processing sequence is generated based on the task resource soft matching matrix, the task requirement marginal vector and the task data to be scheduled. The task processing sequence is obtained by deleting the row and column elements corresponding to the virtual task nodes and virtual resource nodes, arranging the tasks to be scheduled in descending order according to the task requirement value, and arranging them in descending order according to the task data volume when the task requirement values ​​are the same.

[0142] A candidate sequence of task resources is generated based on the task processing sequence and the task resource soft matching matrix. The candidate sequence of task resources is obtained by arranging the candidate computing resources corresponding to each task to be scheduled in descending order according to the soft matching coefficient.

[0143] The link occupancy requirement value is calculated based on the task data to be scheduled. The link occupancy requirement value is obtained by multiplying the normalized task data volume by the normalized task communication dependency.

[0144] Candidate computing resources are verified based on the candidate sequence of task resources, candidate computing resource data, cross-domain link data, and link occupancy requirements, and the matching relationship of task resources is determined.

[0145] In Example 1, candidate computing resources are selected one by one according to the order of the candidate task resource sequence. The available computing power and task queue capacity corresponding to the current candidate computing resource are obtained from the candidate computing resource data, and the normalized available bandwidth of the current candidate computing resource is obtained from the cross-domain link data. The available computing power is compared with the computing power requirement of the task to be scheduled, and the normalized available bandwidth of the link is compared with the link occupancy requirement. When the available computing power is not less than the computing power requirement, the task queue capacity is greater than 0, and the normalized available bandwidth of the link is not less than the link occupancy requirement, a correspondence is established between the task to be scheduled and the current candidate computing resource to determine the task resource matching relationship. If the aforementioned conditions are not met simultaneously, the next candidate computing resource in the candidate task resource sequence is selected for further verification.

[0146] When a candidate computing resource fails the verification, the task resource candidate sequence is updated;

[0147] In Example 1, the updated task resource candidate sequence is obtained by setting the corresponding soft matching coefficient to 0, calculating the link transfer value and transfer weight corresponding to the remaining candidate computing power resources, and allocating the soft matching coefficients that are set to zero to the remaining soft matching coefficients according to the transfer weight. The link transfer value is obtained by dividing the proportion of available bandwidth of the link by the network latency, and the transfer weight is obtained by dividing each link transfer value by the sum of the link transfer values.

[0148] When the number of remaining candidate computing resources is 0, the unmatched status of the corresponding task to be scheduled is recorded. When the task resource matching relationship is determined, the candidate computing resource data and cross-domain link data are updated, and the task resource matching relationship and unmatched status are collected according to the task processing sequence to obtain the task resource hard matching table.

[0149] In Example 1, the update of candidate computing power resource data is performed by deducting the computing power requirement of the task and the capacity of one task queue. The update of cross-domain link data is performed by multiplying the link occupancy requirement value by the available link bandwidth to obtain the link bandwidth deduction value, and then deducting the link bandwidth deduction value from the available link bandwidth.

[0150] In this embodiment, the step of establishing a cross-domain communication connection based on the task resource hard matching table, allocating computing power quotas and sending data of tasks to be scheduled, and triggering candidate computing power resources to execute tasks to be scheduled, specifically involves:

[0151] Generate a task execution sequence based on the task resource hard match table;

[0152] In Example 1, the task resource matching relationship and unmatched status corresponding to each scheduled task are extracted from the task resource hard matching table; according to the arrangement position of the scheduled tasks in the task processing sequence, the corresponding task resource matching relationship is searched sequentially; the found task resource matching relationship is written into the task execution sequence according to the task processing sequence; the unmatched status is kept at the position of the corresponding scheduled task and is not written into the task execution sequence; after all the scheduled tasks are searched and arranged, the task execution sequence is obtained.

[0153] Establish cross-domain communication connections between the task to be scheduled and the candidate computing resources according to the task execution sequence, and calculate the connection waiting time. The connection waiting time is obtained by adding the network latency to the product of the network latency and the link bandwidth utilization.

[0154] Resource status processing is performed based on the establishment status of the cross-domain communication connection. When the cross-domain communication connection is established within the connection waiting time, the computing power quota is allocated according to the computing power requirement of the task. When the cross-domain communication connection is not established within the connection waiting time, the available computing power and task queue capacity of the corresponding candidate computing power resources are restored, the available bandwidth of the corresponding cross-domain link is restored, and the corresponding task to be scheduled is recorded as unmatched.

[0155] The basic transmission volume is calculated based on the data of the tasks to be scheduled and the cross-domain link data, and the data of the tasks to be scheduled is sent in segments according to the basic transmission volume.

[0156] In Example 1, the basic transmission volume is obtained by multiplying the task data volume by the normalized link available bandwidth ratio and dividing by the product of the normalized network latency and the normalized link bandwidth occupancy rate.

[0157] After each data transmission is completed, update the cross-domain link data and recalculate the base transmission volume;

[0158] In Example 1, the update of cross-domain link data is performed by collecting the percentage of available bandwidth of the link, the bandwidth utilization rate of the link, and the network latency and performing normalization processing. When the remaining task data volume is less than the recalculated base transmission volume, the remaining task data volume is determined as the base transmission volume.

[0159] The tasks to be scheduled are executed based on the segmented data of the tasks to be scheduled and the computing power quota, and the task execution results are obtained.

[0160] In Example 1, the task execution result is obtained by receiving each segment of task data to be scheduled through candidate computing power resources, executing the corresponding task data to be scheduled according to the computing power quota, and returning the segment execution result. The segment execution result is obtained when the amount of data sent reaches the amount of task data and all segment execution results are returned.

[0161] By collecting task execution results and unmatched states according to the task execution sequence, the dynamic scheduling results of cross-domain computing resources are obtained.

[0162] Example 2: A cross-domain computing resource dynamic scheduling system, comprising:

[0163] The scheduling data processing module collects data on tasks to be scheduled, candidate computing resources, and cross-domain link data, and processes them to obtain cross-domain scheduling data.

[0164] The collaborative cost construction module calculates data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, generates corresponding cost weights, and establishes a collaborative cost matrix.

[0165] The quantization processing module performs quantization processing on the collaborative cost matrix to obtain the quantized collaborative cost matrix and the quantized residual matrix.

[0166] The transmission parameter construction module establishes the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix based on the task data to be scheduled, candidate computing power resource data, quantized collaborative cost matrix, and quantized residual matrix, and performs total balance.

[0167] The soft matching solution module executes the improved Sinkhorn algorithm based on the task requirement margin vector, resource supply margin vector, initial transmission kernel matrix, and residual correction weight matrix to obtain the task resource soft matching matrix.

[0168] The hard matching generation module verifies candidate computing resources and cross-domain links based on the task resource soft matching matrix, determines the task resource matching relationship, and generates a task resource hard matching table.

[0169] The scheduling and execution module establishes a cross-domain communication connection based on the task resource hard matching table, allocates computing power quotas, sends the data of the tasks to be scheduled in segments, and triggers the candidate computing power resources to execute the tasks to be scheduled, thereby obtaining the dynamic scheduling results of cross-domain computing power resources.

[0170] Example 3: This example provides a computer storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of any of the methods described above.

[0171] The computer storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples of computer storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), registers, hard disks, optical fibers, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof, or any other form of computer storage medium in the art.

[0172] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside within an Application Specific Integrated Circuit (ASIC). In embodiments of the invention, the computer storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0173] Example 4:

[0174] To verify the feasibility of this invention in practice, it was applied to a multi-network domain computing node collaborative processing scenario. The application target was continuously arriving data analysis tasks. The scenario included five tasks to be scheduled, five candidate computing resources, and corresponding cross-domain links, with each candidate computing resource distributed across different network domains. Existing scheduling methods primarily rely on weighted scoring based on available computing power, load rate, task queue capacity, network latency, and available link bandwidth. When the task data volume, task computing power requirements, and task communication dependencies change simultaneously, the score for a single indicator is insufficient to reflect the collaborative costs between data migration, computing power execution, and link occupancy. Furthermore, fluctuations in link status can lead to inconsistencies between task resource matching relationships and actual execution conditions.

[0175] During implementation, data on tasks to be scheduled, candidate computing resources, and cross-domain links are collected at 1-second intervals. Task data includes task data volume, computing power requirements, and communication dependencies. Candidate computing resource data includes available computing power, load rate, and task queue capacity. Cross-domain link data includes network latency, available link bandwidth, available link bandwidth percentage, and link bandwidth occupancy rate. The system corrects the collection time of candidate computing resource data and cross-domain link data based on network latency, and determines the association timeframe by summing the product of network latency and link bandwidth occupancy rate. Within the association timeframe, data with the smallest correction time difference is selected for association. If the available link bandwidth percentage is missing, the link correction data with the second smallest correction time difference is selected for replacement. If other data items are missing, the data change rate is calculated based on the two valid values ​​before the missing time, and the missing data is processed by combining the next valid value, the corresponding time difference, and the available link bandwidth percentage. Complete association data is normalized according to the maximum and minimum values ​​of each data item, and the normalization result is incremented by 1 to form cross-domain scheduling data.

[0176] Based on cross-domain scheduling data, the data migration cost, computing power execution cost, and link occupancy cost are calculated separately. Migration modulation cost and link modulation cost are then derived using task communication dependency and link bandwidth utilization. Migration weight, computing power weight, and link weight are calculated separately according to their corresponding costs and network latency, load rate, link bandwidth utilization, and available link bandwidth percentage. Each of the three weight values ​​is then divided by the sum of the weight values ​​to obtain the corresponding cost weight. The system performs a weighted summation of the migration modulation cost, computing power execution cost, and link modulation cost according to their respective cost weights to establish a collaborative cost matrix.

[0177] In this embodiment, the number of quantization intervals is set to 4. The collaborative cost values ​​are sorted in ascending order and divided into continuous subsequences. The initial quantization representative value is determined based on the mean of each continuous subsequence, and the initial quantization interval boundary is determined based on the mean of adjacent initial quantization representative values. The system corrects the quantization interval boundary using the mean of the available bandwidth ratio of the links on both sides of the boundary, and updates the quantization representative value according to the reciprocal of the available bandwidth ratio of the links. Quantization iteration stops when the difference between two adjacent quantization representative values ​​is less than 0.001. This process obtains the quantization collaborative cost matrix after 7 iterations. The quantization residual matrix is ​​obtained by subtracting the collaborative cost matrix from the quantization collaborative cost matrix element by element.

[0178] The task requirement value is obtained by multiplying the normalized task computing power requirement by the task communication dependency. The resource supply value is obtained by multiplying the normalized available resource computing power by the task queue capacity and then dividing by the load rate. In this verification, the total supply exceeds the total demand. The system adds virtual task nodes, and the difference between the total supply and total demand is determined as the task requirement value of the virtual task node. A corresponding row with a value of 0 is added to the quantized collaborative cost matrix and the quantized residual matrix. The system determines the entropy regularization parameter based on the mean absolute value of the quantized residual matrix and the mean proportion of available link bandwidth, and generates the initial transmission kernel matrix and the residual correction weight matrix.

[0179] When executing the improved Sinkhorn algorithm, the initial transfer kernel matrix is ​​multiplied element-wise with the residual correction weight matrix to obtain the residual correction transfer kernel matrix. The task residual weight vector and resource residual weight vector are then calculated based on the residual correction weight matrix. In each iteration, the task correction deviation and resource correction deviation are calculated. Row scaling is performed when the task correction deviation is large, and column scaling is performed when the resource correction deviation is not less than the task correction deviation. The scaling factor is exponentially operated on according to the corresponding residual weight, and the transfer kernel matrix is ​​updated. The convergence threshold is set to 0.001. After 18 iterations, both the task correction deviation and the resource correction deviation satisfy the convergence threshold, resulting in the task-resource soft-match matrix.

[0180] The system deletes the row and column elements corresponding to virtual task nodes and virtual resource nodes, generates a task processing sequence in descending order of task requirement values, and generates a candidate sequence of task resources in descending order of soft matching coefficients. The link occupancy requirement value is obtained by multiplying the normalized task data volume by the task communication dependency. A task resource matching relationship is determined when the candidate computing power resources simultaneously satisfy the following conditions: available computing power is not less than the task computing power requirement, task queue capacity is greater than 0, and normalized available link bandwidth is not less than the link occupancy requirement value. The soft matching coefficient corresponding to the candidate computing power resources that fail the verification is set to 0, and they are redistributed to the remaining soft matching coefficients based on the link transfer value formed by the proportion of available link bandwidth and network latency, ultimately obtaining the task resource hard matching table.

[0181] Table 1 Verification Results of Cross-Domain Computing Resource Dynamic Scheduling

[0182] Task Number Task data volume (GB) Unit of task computing power requirement Task communication dependency Network latency in milliseconds Link available bandwidth percentage Soft matching coefficient Matching resource number Number of iterations Scheduling completion time s T1 12.4 38 0.82 26 0.76 0.84 R3 18 8.7 T2 8.6 24 0.55 18 0.81 0.79 R1 18 5.9 T3 15.2 42 0.91 31 0.69 0.81 R5 18 10.8 T4 6.8 19 0.47 14 0.86 0.77 R2 18 4.6 T5 10.7 31 0.73 23 0.78 0.82 R4 18 7.4

[0183] As shown in Table 1, tasks T1 and T3, which have high communication dependencies, were matched to candidate computing resources with a high proportion of available link bandwidth and sufficient available computing power to meet the task's computing power requirements, with soft matching coefficients of 0.84 and 0.81, respectively. Task T4 has lower data volume and computing power requirements; the matching process prioritizes lower network latency and a higher proportion of available link bandwidth, with a scheduling completion time of 4.6 seconds. All five tasks to be scheduled formed stable soft matching matrices for their resources after 18 improved Sinkhorn iterations. Hard matching tables for their resources were then formed by verifying available computing power, task queue capacity, and available link bandwidth. No task resource matching relationships exceeded the resource supply capacity.

[0184] The system generates a task execution sequence based on a hard-match table of task resources. It establishes cross-domain communication connections between the tasks to be scheduled and candidate computing resources, and calculates the connection waiting time based on network latency and link bandwidth utilization. After the cross-domain communication connection is established within the connection waiting time, computing power quotas are allocated according to the task's computing power requirements. The system also calculates the basic transmission volume based on the task data volume, the normalized percentage of available link bandwidth, link bandwidth utilization, and network latency. After each data transmission, cross-domain link data is re-collected and the basic transmission volume is updated. Candidate computing resources receive each segment of task data to be scheduled, execute the corresponding data according to their computing power quotas, and return the segmented execution results. Once the transmitted data volume reaches the task data volume and all segmented execution results are returned, the system aggregates the segmented execution results and unmatched states to obtain the dynamic scheduling result of the cross-domain computing resources. This embodiment demonstrates that the present invention can incorporate task requirements, resource supply, and cross-domain link status into a unified scheduling process, and completes task resource matching and execution coordination through quantized residual correction, improved Sinkhorn iteration, hard-match verification, and dynamic segmented transmission.

[0185] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic scheduling of cross-domain computing resources, characterized in that, include: Collect data on tasks to be scheduled, candidate computing resources, and cross-domain link data, perform normalization processing and corresponding correlation to obtain cross-domain scheduling data; Calculate the data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, and establish a collaborative cost matrix; The collaborative cost matrix is ​​quantized to obtain the quantized collaborative cost matrix and the quantized residual matrix; Based on the task data to be scheduled, candidate computing resource data, quantized collaborative cost matrix and quantized residual matrix, establish the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix and residual correction weight matrix, and complete the total balance; The improved Sinkhorn algorithm is executed based on the task requirement marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix to obtain the task resource soft matching matrix. The task resource matching relationship is determined based on the task resource soft matching matrix, and the task resource hard matching table is obtained. Establish a cross-domain communication connection based on the task resource hard matching table, allocate computing power quotas and send data of tasks to be scheduled, trigger candidate computing power resources to execute the tasks to be scheduled, and obtain the dynamic scheduling results of cross-domain computing power resources.

2. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The collection of data on tasks to be scheduled, candidate computing resources, and cross-domain link data is normalized and correlated accordingly, specifically as follows: Data on tasks to be scheduled, candidate computing resources, and cross-domain links are collected separately according to the collection time. The data on tasks to be scheduled includes the amount of task data, the amount of computing power required by the task, and the communication dependency of the task. The data on candidate computing resources includes available computing power, load rate, and task queue capacity. The data on cross-domain links includes network latency, available bandwidth of the link, percentage of available bandwidth of the link, and bandwidth utilization rate of the link. Based on the data of the tasks to be scheduled, the candidate computing resources data, and the cross-domain link data, the collection time correction is performed, and the associated time width is calculated to obtain the resource correction data, link correction data, and associated time width; Based on the data of tasks to be scheduled, resource correction data, link correction data, and associated time span, candidate associated data are obtained by filtering the data. Validate the data items in the candidate related data and handle missing data to obtain complete related data; Normalize the data items in the complete associated data to obtain normalized associated data; The normalized correlation data is arranged in the order of the tasks to be scheduled, candidate computing resources, and cross-domain links to obtain cross-domain scheduling data.

3. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The process involves calculating data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, and establishing a collaborative cost matrix, specifically as follows: Based on the correspondence between the tasks to be scheduled, candidate computing resources, and cross-domain links, normalized task data volume, task computing power requirement, task communication dependency, available computing power, load rate, network latency, available link bandwidth, percentage of available link bandwidth, and link bandwidth occupancy rate are extracted from the cross-domain scheduling data to obtain candidate cost data. Calculate the data migration cost, computing power execution cost, and link occupancy cost based on the candidate cost data; The migration modulation cost and link modulation cost are calculated based on the data migration cost, link occupancy cost, normalized task communication dependency, and normalized link bandwidth occupancy. Calculate the migration weight value, computing power weight value, and link weight value based on the migration modulation cost, computing power execution cost, link modulation cost, and candidate cost data. Calculate the corresponding cost weights based on the migration weight value, computing power weight value, and link weight value; The collaborative cost value is calculated based on the migration modulation cost, computing power execution cost, link modulation cost and corresponding cost weights, and the collaborative cost values ​​are arranged in the order of the tasks to be scheduled and the candidate computing power resources to obtain the collaborative cost matrix.

4. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The quantization process performed on the collaborative cost matrix is ​​as follows: Extract the collaborative cost value and the corresponding link available bandwidth percentage from the collaborative cost matrix, and sort them to obtain the cost sorting sequence and bandwidth percentage sequence. Input the cost sorting sequence and bandwidth percentage sequence into the Lloyd-Max algorithm to divide the continuous subsequence and calculate the initial quantization representative value and the initial quantization interval boundary; The boundary correction amount is calculated based on the bandwidth ratio sequence, the initial quantization representative value, and the initial quantization interval boundary. The initial quantization interval boundary is then corrected to obtain the corrected quantization interval boundary. The cost sorting sequence is redefined based on the corrected quantization interval boundaries, and the quantization representative value of each quantization interval is updated. The quantization interval boundary is recalculated and corrected based on the updated quantization representative value until the difference between two adjacent quantization representative values ​​meets the quantization convergence threshold, thus obtaining the target quantization representative value and the target quantization interval boundary. The quantization collaborative cost matrix and quantization residual matrix are generated based on the target quantization representative value, the target quantization interval boundary, and the collaborative cost matrix.

5. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The establishment of the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix is ​​specifically as follows: Calculate the task requirement value corresponding to each task to be scheduled based on the task data to be scheduled, and calculate the resource supply value corresponding to each candidate computing resource based on the candidate computing resource data. Arrange the task demand values ​​and resource supply values ​​according to the order of the tasks to be scheduled and the candidate computing resources, respectively, to obtain the task demand marginal vector and resource supply marginal vector, and calculate the total demand and total supply respectively. Perform total balance based on total demand and total supply, and update the task demand marginal vector, resource supply marginal vector, quantified collaborative cost matrix, and quantified residual matrix; Calculate the entropy regularization parameter based on the updated quantized residual matrix and cross-domain link data; Calculate the initial transmission kernel matrix based on the updated quantization co-cost matrix and entropy regularization parameter; Calculate the residual correction weight matrix based on the updated quantization residual matrix.

6. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The improved Sinkhorn algorithm is as follows: Calculate the residual corrected transfer kernel matrix based on the initial transfer kernel matrix and the residual corrected weight matrix; Calculate the task residual weight vector and the resource residual weight vector based on the residual correction weight matrix; Calculate the task correction bias and resource correction bias based on the residual correction transmission kernel matrix, task requirement marginal vector, resource supply marginal vector, task residual weight vector, and resource residual weight vector; Compare the task correction deviation and the resource correction deviation. When the task correction deviation is greater than the resource correction deviation, generate a row scaling factor. When the resource correction deviation is not less than the task correction deviation, generate a column scaling factor. When generating row scaling factors, the residual correction transmission kernel matrix is ​​updated based on the row scaling factors. When generating column scaling factors, the residual correction transmission kernel matrix is ​​updated based on the column scaling factors to obtain the updated transmission kernel matrix. The updated transmission kernel matrix is ​​used as the residual corrected transmission kernel matrix to recalculate the task correction bias and resource correction bias. When both the task correction bias and resource correction bias meet the convergence threshold, the updated transmission kernel matrix is ​​determined as the task-resource soft matching matrix.

7. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The process of determining the task resource matching relationship based on the task resource soft matching matrix is ​​as follows: Generate a task processing sequence based on the task resource soft matching matrix, the task requirement marginal vector, and the task data to be scheduled. Generate a candidate sequence of task resources based on the task processing sequence and the task resource soft matching matrix; Calculate the link occupancy requirement value based on the data of the tasks to be scheduled; Candidate computing resources are verified based on the candidate sequence of task resources, candidate computing resource data, cross-domain link data, and link occupancy requirements, and the matching relationship of task resources is determined. When a candidate computing resource fails the verification, the task resource candidate sequence is updated; When the number of remaining candidate computing resources is 0, the unmatched status of the corresponding task to be scheduled is recorded. When the task resource matching relationship is determined, the candidate computing resource data and cross-domain link data are updated, and the task resource matching relationship and unmatched status are collected according to the task processing sequence to obtain the task resource hard matching table.

8. The method for dynamic scheduling of cross-domain computing resources according to claim 1, characterized in that, The process of establishing a cross-domain communication connection based on the task resource hard matching table, allocating computing power quotas and sending data of tasks to be scheduled, and triggering candidate computing power resources to execute tasks to be scheduled, specifically involves: Generate a task execution sequence based on the task resource hard match table; Establish cross-domain communication connections between the tasks to be scheduled and the candidate computing resources based on the task execution sequence, and calculate the connection waiting time. Resource status processing is performed based on the establishment status of the cross-domain communication connection. When the cross-domain communication connection is established within the connection waiting time, the computing power quota is allocated according to the computing power requirement of the task. When the cross-domain communication connection is not established within the connection waiting time, the available computing power and task queue capacity of the corresponding candidate computing power resources are restored, the available bandwidth of the corresponding cross-domain link is restored, and the corresponding task to be scheduled is recorded as unmatched. The basic transmission volume is calculated based on the data of the tasks to be scheduled and the cross-domain link data, and the data of the tasks to be scheduled is sent in segments according to the basic transmission volume. After each data transmission is completed, update the cross-domain link data and recalculate the base transmission volume; The tasks to be scheduled are executed based on the segmented data of the tasks to be scheduled and the computing power quota, and the task execution results are obtained. By collecting task execution results and unmatched states according to the task execution sequence, the dynamic scheduling results of cross-domain computing resources are obtained.

9. A cross-domain computing resource dynamic scheduling system, used to execute the cross-domain computing resource dynamic scheduling method according to any one of claims 1 to 8, characterized in that, include: The scheduling data processing module collects data on tasks to be scheduled, candidate computing resources, and cross-domain link data, and processes them to obtain cross-domain scheduling data. The collaborative cost construction module calculates data migration cost, computing power execution cost, and link occupancy cost based on cross-domain scheduling data, generates corresponding cost weights, and establishes a collaborative cost matrix. The quantization processing module performs quantization processing on the collaborative cost matrix to obtain the quantized collaborative cost matrix and the quantized residual matrix. The transmission parameter construction module establishes the task demand marginal vector, resource supply marginal vector, initial transmission kernel matrix, and residual correction weight matrix based on the task data to be scheduled, candidate computing power resource data, quantized collaborative cost matrix, and quantized residual matrix, and performs total balance. The soft matching solution module executes the improved Sinkhorn algorithm based on the task requirement margin vector, resource supply margin vector, initial transmission kernel matrix, and residual correction weight matrix to obtain the task resource soft matching matrix. The hard matching generation module verifies candidate computing resources and cross-domain links based on the task resource soft matching matrix, determines the task resource matching relationship, and generates a task resource hard matching table. The scheduling and execution module establishes a cross-domain communication connection based on the task resource hard matching table, allocates computing power quotas, sends the data of the tasks to be scheduled in segments, and triggers the candidate computing power resources to execute the tasks to be scheduled, thereby obtaining the dynamic scheduling results of cross-domain computing power resources.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the method of any one of claims 1-8.

Images