A computing resource scheduling optimization method for large models

Abstract

The present invention discloses a computing resource scheduling optimization method for a large model, which relates to the technical field of resource scheduling optimization. Tasks are divided into two categories: abnormal tasks and ordinary tasks, and a weight value is matched according to the resource processing capacity value and the number of requests of each computing node. The computing nodes are divided into ordinary computing nodes and buffer computing nodes. The first priority is abnormal tasks. The idle computing nodes in the ordinary computing nodes are first scheduled as buffer temporary computing nodes, and the load capacity of the buffer computing nodes is temporarily increased to match the task requirements of abnormal tasks. The problem of facing sudden and abnormal task requirements can be effectively solved, and the problem of facing sudden and abnormal task requirements can be effectively solved. The dynamic computing resource scheduling optimization effect is achieved, and the accuracy of the calculation results is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of resource scheduling optimization, and in particular to a computing resource scheduling optimization method for a large model. Background Art

[0002] In the exploration of the technical field of large model computing resource scheduling optimization, a series of efficient methods have emerged, which are characterized by high resource utilization, fast training speed, strong scalability, significant cost-effectiveness, superior flexibility and excellent system stability. The core of these methods is to accurately match computing resources with model training requirements, aiming to maximize resource utilization efficiency and accelerate the model development process.

[0003] However, when actually deploying these optimization strategies, specific application scenarios and specific needs must be carefully considered. Different models and tasks have different resource requirements. Improper scheduling strategies may lead to performance bottlenecks or resource waste. Therefore, it is crucial to accurately match strategies and needs.

[0004] Most existing technologies currently use prediction models as the key support for resource scheduling decisions, and their performance and accuracy directly depend on the training and optimization of historical data. However, it is worth noting that historical data often cannot fully capture the dynamics of future resource demand, especially when faced with sudden and abnormal task demands, and it is impossible to predict. In addition, traditional static priority strategies are difficult to adapt to rapidly changing resource demands, and although dynamic priority strategies are more flexible, they may also introduce additional management complexity and uncertainty. Summary of the invention

[0005] In view of the above situation, the present invention divides tasks into two categories, abnormal tasks and ordinary tasks, and makes targeted optimization of computing resource scheduling, which can effectively solve the problem of sudden and abnormal task demands, achieve dynamic computing resource scheduling optimization effect, and improve the accuracy of calculation results.

[0006] The technical solution includes:

[0007] S1, receiving computing tasks, removing duplicate and invalid computing tasks, and then classifying all tasks into normal tasks and abnormal tasks. Tasks with sudden and abnormal demands are classified as abnormal tasks, and the rest are normal tasks;

[0008] S2, classify the common tasks to be calculated into important and urgent A, important but not urgent B, unimportant but urgent C, unimportant and not urgent D according to the execution time and resource requirements of the tasks, and then calculate the total required resource sizes of tasks A, B, C, and D in turn, A1, B1, C1, and D1;

[0009] S3, divide the computing nodes into ordinary computing nodes and buffer computing nodes, match a weight value according to the resource processing capacity value and the number of requests of each computing node, and calculate the total resource processing capacity value M of all ordinary computing nodes;

[0010] S4, the first priority is the abnormal task, and the idle computing nodes in the normal computing nodes are scheduled as buffer temporary computing nodes. After the abnormal task processing is completed, the buffer temporary computing nodes are re-marked as normal computing nodes;

[0011] The buffer computing node with a high weight value is marked as the target matching node first. For the same weight values, the number of connections of the abnormal task to be assigned to the buffer computing node with the same weight value is compared. The buffer computing node with the least number of connections is the target node until the abnormal task matching is completed;

[0012] S5, the second priority is ordinary tasks, which are assigned in order of A, B, C, and D tasks.

[0013] Prioritize important and urgent Class A tasks and calculate the difference A1-M. If the difference A1-M is positive, it means that the load range of ordinary computing nodes is exceeded. The idle buffer computing nodes are scheduled as temporary computing nodes. After the ordinary tasks are processed, the temporary computing nodes are re-marked as buffer computing nodes.

[0014] On the contrary, if the difference A1-M is negative, the computing node with a higher weight value is marked as the target matching node first. For those with the same weight value, the number of connections of the A-type tasks to be assigned to the computing nodes with the same weight value is compared, and the computing node with the least number of connections is the target node until the matching of the A-type tasks is completed. Finally, the B, C, and D-type tasks are assigned in turn according to the matching method of the A-type tasks.

[0015] Furthermore, in the steps S4 and S5, the abnormal task is set as the first priority. When there is no idle ordinary computing node in the system, or the existing buffer computing node is occupied by the ordinary task in step S5, the system stops scheduling the buffer computing node, suspends the execution of the task, and the suspended task is saved to the cache database;

[0016] The system suspends the ordinary computing nodes processing the class D tasks, and the suspended class D tasks are saved to the cache database. The suspended ordinary computing nodes are temporarily converted into buffer computing nodes, and the temporarily converted buffer computing nodes are marked as buffer temporary computing nodes, and the execution continues according to step S4;

[0017] Finally, when the abnormal task is completed, the buffer computing node continues to execute the corresponding suspended task in the cache database, the suspended ordinary computing node ends the call, and the ordinary computing node continues to execute the corresponding suspended task in its cache database.

[0018] Furthermore, in the steps S4 and S5, when there are no abnormal tasks in the system and there are a large number of idle ordinary computing nodes, in the process of executing step S5, firstly, the sum K of the resource processing capacity values ​​of the buffer computing nodes and the idle ordinary computing nodes is calculated, and the difference of K-A1 is calculated. If the difference is positive, the A-type tasks are copied. Then, the difference of K-A1-B1 is calculated. If the difference is positive, the B-type tasks are copied. Then, the difference of K-A1-B1-C1 is calculated. If the difference is positive, the C-type tasks are copied. At the same time, the difference of K-A1-B1-D1 is calculated. If the difference is positive, the D-type tasks are copied. Finally, all the copied tasks are copied according to step S4. Step S5 performs task resource matching processing. If the copied task computing resource processing result X and the copied task computing resource processing result Y are relatively the same, they are outputted normally. Otherwise, if the two results X and Y are relatively different, they are matched with other different computing nodes for further task computing resource processing. The processing result is recorded as Z. Z is compared with X and Y respectively. The calculation result of Z is used as the standard. If X and Z are different, the computing node of the computing resource processing result X is located and marked as an abnormal computing node. If Y and Z are different, the computing node of the computing resource processing result X is located and marked as an abnormal computing node, and the relevant personnel are notified to perform maintenance or algorithm detection on it.

[0019] On the contrary, if any difference is negative, it is considered that the copy task cannot be continued, and step S5 is executed.

[0020] Furthermore, in the weight value setting in step S3, the weight value of the common computing node = (the resource processing capacity value of the common computing node / the sum of the resource processing capacity values ​​of all common computing nodes) × (1 / the number of requests) × adjustment coefficient;

[0021] Among them, the adjustment coefficient is experimentally set according to the big data model, and the value is 1 when the number of requests is 0;

[0022] The weight value configuration of the buffer computing node is the same as the weight value formula of the ordinary computing node.

[0023] Furthermore, in step S4, when an idle computing node among ordinary computing nodes is regarded as a buffer temporary computing node, the sum Q of the buffer temporary computing nodes is calculated;

[0024] When executing step S5, priority is given to matching important and urgent tasks of category A, the difference is calculated as A1-(MQ), and then step S5 is continued normally;

[0025] When the matching of important and urgent Class A tasks is completed, Class B tasks are assigned and the difference is calculated as B1-(MQ-A1). When Class C tasks are assigned, the difference is calculated as C1-(MQ-A1-B1). When Class D tasks are assigned, the difference is calculated as D1-(MQ-A1-B1-C1).

[0026] Due to the adoption of the above technical solution, the present invention has the following advantages compared with the prior art:

[0027] 1. Tasks are divided into two categories: abnormal tasks and ordinary tasks. A weight value is matched according to the resource processing capacity value and the number of requests of each computing node. The computing nodes are divided into ordinary computing nodes and buffer computing nodes. The first priority is abnormal tasks. The idle computing nodes in the ordinary computing nodes are scheduled as temporary buffer computing nodes. The load capacity of the buffer computing nodes is temporarily increased to match the task requirements of abnormal tasks. This can effectively solve the problem of sudden and abnormal task requirements.

[0028] 2. The second priority is ordinary tasks. Ordinary tasks are assigned in order according to the four categories of A, B, C, and D. Important and urgent A-class tasks are matched first. The difference A1-M is negative. Computing nodes with high weight values ​​are marked as target matching nodes first. For those with the same weight values, the number of connections of the A-class tasks to be assigned to the computing nodes with the same weight values ​​is compared. The computing node with the least number of connections is the target node until the matching of A-class tasks is completed. According to the matching method of A-class tasks, B, C, and D-class tasks are assigned in turn to balance the load of computing nodes.

[0029] 3. The abnormal task is set as the first priority. When there is no idle ordinary computing node in the system, or the existing buffer computing node is occupied by the ordinary task in step S5, the system will first terminate the scheduling of the buffer computing node, suspend the tasks it is executing, and safely save these tasks in the cache database; when there is no abnormal task in the system and there are a large number of idle ordinary computing nodes, during the execution of step S5, the A-type task is copied, and the computing resource processing result X of the copied task is relatively the same as the computing resource processing result Y of the copied task, and the normal output is output. Otherwise, the two results X and Y are relatively different, and other different computing nodes are matched to further perform task computing resource processing. The processing result is recorded as Z, and Z is compared with X and Y respectively. The calculation result of Z is used as the standard. If X is different from Z, the computing node with the computing resource processing result X is located and marked as an abnormal computing node. If Y is different from Z, the computing node with the computing resource processing result X is located and marked as an abnormal computing node, and the relevant personnel are notified to perform maintenance or algorithm detection on it, thereby achieving the dynamic computing resource scheduling optimization effect and improving the accuracy of the calculation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The present invention is a flowchart of a method for optimizing computing resource scheduling for large models. DETAILED DESCRIPTION

[0031] The above and other technical contents, features and effects of the present invention are described in detail below with reference to the attached Figure 1The detailed description of the embodiments will clearly show that the structural contents mentioned in the following embodiments are all based on the drawings in the specification.

[0032] Since tasks with sudden and abnormal demands have three major task characteristics: variable and unknown, difficult to predict the size of resource requirements, and high timeliness, most existing technologies currently use prediction models as the key support for resource scheduling decisions. When faced with sudden and abnormal task requirements, they cannot be predicted. Therefore, this solution includes:

[0033] S1, receiving computing tasks, removing duplicate and invalid computing tasks, and then classifying all tasks into normal tasks and abnormal tasks. Tasks with sudden and abnormal demands are classified as abnormal tasks, and the rest are normal tasks;

[0034] S2, classify the common tasks to be calculated into important and urgent A, important but not urgent B, unimportant but urgent C, unimportant and not urgent D according to the execution time and resource requirements of the tasks, and then calculate the total required resource sizes of tasks A, B, C, and D in turn, A1, B1, C1, and D1;

[0035] S3, divide the computing nodes into ordinary computing nodes and buffer computing nodes, match a weight value according to the resource processing capacity value and the number of requests of each computing node, and calculate the total resource processing capacity value M of all ordinary computing nodes;

[0036] Tasks are divided into two categories: abnormal tasks and ordinary tasks. A weight value is matched according to the resource processing capacity value and the number of requests of each computing node. The computing nodes are divided into ordinary computing nodes and buffer computing nodes. The first priority is abnormal tasks, so as to match the task requirements of abnormal tasks. It can effectively solve the problem of sudden and abnormal task requirements and make targeted processing. Due to the three major task characteristics of abnormal tasks, namely, the variability and uncertainty, the difficulty in predicting the size of resource requirements, and the high timeliness, buffer computing nodes are specially reserved to handle abnormal tasks in a targeted manner.

[0037] S4, the first priority is the abnormal task. First, the idle computing nodes in the ordinary computing nodes are scheduled as temporary buffer computing nodes, and the load capacity of the buffer computing nodes is temporarily increased to match the task requirements of the abnormal task. After the abnormal task is processed, the temporary buffer computing nodes are re-marked as ordinary computing nodes;

[0038] On the basis of improving the load capacity of buffer computing nodes, buffer computing nodes with high weight values ​​are marked as target matching nodes in priority. For those with the same weight values, the number of connections of abnormal tasks to be assigned to buffer computing nodes with the same weight values ​​is compared, and the buffer computing node with the least number of connections is the target node until the abnormal task matching is completed, thereby balancing the load of buffer computing nodes.

[0039] S5, the second priority is ordinary tasks, which are assigned in order of A, B, C, and D tasks.

[0040] Prioritize matching important and urgent Class A tasks, calculate the difference A1-M, and if the difference A1-M is positive, it means that the load range of ordinary computing nodes is exceeded. Schedule idle buffer computing nodes as temporary computing nodes to temporarily increase the load capacity of computing nodes to match the task requirements of Class A tasks. After the ordinary tasks are processed, the temporary computing nodes are re-marked as buffer computing nodes.

[0041] On the contrary, if the difference A1-M is negative, the computing nodes with higher weight values ​​are marked as target matching nodes first. For those with the same weight values, the number of connections of the A-type tasks to be assigned to the computing nodes with the same weight values ​​is compared, and the computing node with the least number of connections is the target node until the matching of the A-type tasks is completed. This can balance the load of the computing nodes and assign B, C, and D-type tasks.

[0042] Furthermore, in the steps S4 and S5, the abnormal task is set as the first priority. When there is no idle ordinary computing node in the system, or the existing buffer computing node is occupied by the ordinary task in step S5, the system stops scheduling the buffer computing node, suspends the execution of the task, and the suspended task is saved to the cache database;

[0043] The system suspends the ordinary computing nodes processing the class D tasks, and the suspended class D tasks are saved to the cache database. The suspended ordinary computing nodes are temporarily converted into buffer computing nodes, and the temporarily converted buffer computing nodes are marked as buffer temporary computing nodes, and the execution continues according to step S4;

[0044] Finally, when the abnormal task is completed, the buffer computing node continues to execute the corresponding suspended task in the cache database, the suspended ordinary computing node ends the call, and the ordinary computing node continues to execute the corresponding suspended task in its cache database.

[0045] Furthermore, in the steps S4 and S5, when there are no abnormal tasks in the system and there are a large number of idle ordinary computing nodes, in the process of executing step S5, firstly, the sum K of the resource processing capacity values ​​of the buffer computing nodes and the idle ordinary computing nodes is calculated, and the difference of K-A1 is calculated. If the difference is positive, the A-type tasks are copied. Then, the difference of K-A1-B1 is calculated. If the difference is positive, the B-type tasks are copied. Then, the difference of K-A1-B1-C1 is calculated. If the difference is positive, the C-type tasks are copied. At the same time, the difference of K-A1-B1-D1 is calculated. If the difference is positive, the D-type tasks are copied. Finally, all the copied tasks are copied according to step S4. Step S5 performs task resource matching processing. If the copied task computing resource processing result X and the copied task computing resource processing result Y are relatively the same, they are outputted normally. Otherwise, if the two results X and Y are relatively different, they are matched with other different computing nodes for further task computing resource processing. The processing result is recorded as Z. Z is compared with X and Y respectively. The calculation result of Z is used as the standard. If X and Z are different, the computing node of the computing resource processing result X is located and marked as an abnormal computing node. If Y and Z are different, the computing node of the computing resource processing result X is located and marked as an abnormal computing node, and the relevant personnel are notified to perform maintenance or algorithm detection on it.

[0046] On the contrary, if any difference is negative, it is considered that the copy task cannot be continued, and step S5 is executed.

[0047] In order to achieve the dynamic computing resource scheduling optimization effect, when there are no idle ordinary computing nodes in the system, or the existing buffer computing nodes are occupied by ordinary tasks in step S5, the system will first terminate the scheduling of the buffer computing nodes, suspend the tasks being executed, and safely save these tasks in the cache database; or when there are no abnormal tasks in the system and there are a large number of idle ordinary computing nodes, during the execution of step S5, the A-type task is copied, and the copied task computing resource processing result X and the copied task computing resource processing result Y are relatively the same, and are output normally. Otherwise, the two results X and Y are relatively different, and other different computing nodes are matched to further perform task computing resource processing. The processing result is recorded as Z, and Z is compared with X and Y respectively. The calculation result of Z is used as the standard. If X is different from Z, the computing node of the computing resource processing result X is located and marked as an abnormal computing node. If Y is different from Z, the computing node of the computing resource processing result X is located and marked as an abnormal computing node, and the relevant personnel are notified to perform maintenance or algorithm detection on it, thereby achieving and improving the accuracy of the computing results.

[0048] Further, the weight value setting in step S3 specifically refers to the following method: the weight value of the common computing node = (the resource processing capacity value of the common computing node / the sum of the resource processing capacity values ​​of all common computing nodes) × (1 / the number of requests) × adjustment coefficient;

[0049] The adjustment coefficient is experimentally set according to the big data model to ensure the rationality and balance of the weight value. When the number of requests is 0, the value is 1.

[0050] The weight value configuration of the buffer computing node is the same as the weight value formula of the ordinary computing node;

[0051] And the adjustment coefficient changes dynamically. As the computing resource scheduling optimization application data increases and the application scenario becomes more complex, the value of the adjustment coefficient is dynamically adjusted according to the specific computing feedback results.

[0052] The tasks with sudden and abnormal demands have three major characteristics: variable and unknown, difficult to predict the size of resource requirements, and high timeliness;

[0053] Uncertainty: Unlike ordinary tasks, which can be predicted and planned, they appear suddenly and have changing requirements, which cannot be estimated in advance.

[0054] The resource requirements are difficult to estimate. Since there is little reference data for tasks with sudden or abnormal requirements, it is impossible to estimate the amount of resources required for the tasks.

[0055] Tasks with high timeliness, sudden and abnormal demands need to be handled in a short time, otherwise serious consequences will occur.

[0056] Furthermore, in step S4, when an idle computing node among ordinary computing nodes is regarded as a buffer temporary computing node, the total number Q of buffer temporary computing nodes needs to be calculated at the same time;

[0057] When executing step S5, priority is given to matching important and urgent tasks of category A, the difference is calculated as A1-(MQ), and then step S5 is continued normally;

[0058] When the matching of important and urgent Class A tasks is completed, the difference when assigning Class B tasks is calculated as B1-(MQ-A1), the difference when assigning Class C tasks is calculated as C1-(MQ-A1-B1), and the difference when assigning Class D tasks is calculated as D1-(MQ-A1-B1-C1).

[0059] When the present invention is used specifically, on the basis of the existing technology, computing tasks are received, repeated and invalid computing tasks are removed, and then all tasks are divided into ordinary tasks and abnormal tasks. Tasks with sudden and abnormal requirements are classified as abnormal tasks, and the rest are ordinary tasks.

[0060] The common tasks to be calculated are divided into important and urgent A, important but not urgent B, unimportant but urgent C, unimportant and not urgent D according to the execution time and resource requirements of the tasks, and then the total required resource sizes of tasks A1, B1, C1, and D are calculated in turn.

[0061] The computing nodes are divided into ordinary computing nodes and buffer computing nodes. A weight value is matched according to the resource processing capability value and the number of requests of each computing node. The higher the resource processing capability value and the lower the number of requests, the larger the matching weight value. The total resource processing capability value M of all ordinary computing nodes is calculated.

[0062] The first priority is abnormal tasks. The idle computing nodes in the normal computing nodes are scheduled as temporary buffer computing nodes. After the abnormal tasks are processed, the temporary buffer computing nodes are re-marked as normal computing nodes.

[0063] The buffer computing nodes with higher weight values ​​are marked as target matching nodes first. For those with the same weight values, the number of connections of the abnormal tasks to be assigned to the buffer computing nodes with the same weight values ​​is compared. The buffer computing node with the least number of connections is the target node until the abnormal task matching is completed.

[0064] The second priority is ordinary tasks, which are assigned in order of A, B, C, and D.

[0065] Prioritize important and urgent Class A tasks and calculate the difference A1-M. If the difference A1-M is positive, it means that the load range of ordinary computing nodes is exceeded. The idle buffer computing nodes are scheduled as temporary computing nodes. After the ordinary tasks are processed, the temporary computing nodes are re-marked as buffer computing nodes.

[0066] On the contrary, if the difference A1-M is negative, the computing node with a high weight value is marked as the target matching node first. For the same weight values, the number of connections of the A-class tasks to be assigned to the computing nodes with the same weight value is compared, and the computing node with the least number of connections is the target node until the matching of the A-class tasks is completed. Finally, the B, C, and D-class tasks are assigned in turn according to the matching method of the A-class tasks. By dividing the tasks into two categories, abnormal tasks and ordinary tasks, and making targeted optimization of computing resource scheduling, it can effectively solve the problem of sudden and abnormal task requirements, achieve the optimization effect of dynamic computing resource scheduling, and improve the accuracy of the calculation results.

[0067] The above is a further detailed description of the present invention in combination with a specific implementation method, and it cannot be determined that the specific implementation of the present invention is limited to this; for technical personnel in the technical field to which the present invention belongs and related technical fields, based on the technical solution of the present invention, the expansion and replacement of operating methods and data should all fall within the protection scope of the present invention.

Claims

1. A computing resource scheduling optimization method for a large model, characterized in that: include: S1, receiving computing tasks, removing duplicate and invalid computing tasks, and then classifying all tasks into normal tasks and abnormal tasks. Tasks with sudden and abnormal demands are classified as abnormal tasks, and the rest are normal tasks; S2, classify the common tasks to be calculated into important and urgent A, important but not urgent B, unimportant but urgent C, unimportant and not urgent D according to the execution time and resource requirements of the tasks, and then calculate the total required resource sizes of tasks A, B, C, and D in turn, A1, B1, C1, and D1; S3, divide the computing nodes into ordinary computing nodes and buffer computing nodes, match a weight value according to the resource processing capacity value and the number of requests of each computing node, and calculate the total resource processing capacity value M of all ordinary computing nodes; S4, the first priority is the abnormal task, and the idle computing nodes in the normal computing nodes are scheduled as buffer temporary computing nodes. After the abnormal task processing is completed, the buffer temporary computing nodes are re-marked as normal computing nodes; The buffer computing node with a high weight value is marked as the target matching node first. For the same weight values, the number of connections of the abnormal task to be assigned to the buffer computing node with the same weight value is compared. The buffer computing node with the least number of connections is the target node until the abnormal task matching is completed; S5, the second priority is ordinary tasks, which are assigned in order of A, B, C, and D tasks. Prioritize important and urgent Class A tasks and calculate the difference A1-M. If the difference A1-M is positive, it means that the load range of ordinary computing nodes is exceeded. The idle buffer computing nodes are scheduled as temporary computing nodes. After the ordinary tasks are processed, the temporary computing nodes are re-marked as buffer computing nodes. On the contrary, if the difference A1-M is negative, the computing node with a higher weight value is marked as the target matching node first. If the weight values ​​are the same, the number of connections of the A-class tasks to be assigned to the computing nodes with the same weight value is compared, and the computing node with the least number of connections is the target node until the matching of the A-class tasks is completed. Finally, the B, C, and D-class tasks are assigned in turn according to the matching method of the A-class tasks. In the steps S4 and S5, the abnormal task is set to the first priority. When there is no idle ordinary computing node in the system, or the existing buffer computing node is occupied by the ordinary task in step S5, the system stops scheduling the buffer computing node, suspends the execution of the task, and saves the suspended task to the cache database; The system suspends the ordinary computing nodes processing the class D tasks, and the suspended class D tasks are saved to the cache database. The suspended ordinary computing nodes are temporarily converted into buffer computing nodes, and the temporarily converted buffer computing nodes are marked as buffer temporary computing nodes, and the execution continues according to step S4; Finally, when the abnormal task is completed, the buffer computing node continues to execute the corresponding suspended task in the cache database, the suspended ordinary computing node ends the call, and the ordinary computing node continues to execute the corresponding suspended task in its cache database.

2. A computing resource scheduling optimization method for a large model according to claim 1, characterized in that: In the steps S4 and S5, when there are no abnormal tasks in the system and there are a large number of idle ordinary computing nodes, in the process of executing step S5, first calculate the sum K of the resource processing capacity values ​​of the buffer computing nodes and the idle ordinary computing nodes, and calculate the difference of K-A1. If the difference is positive, copy the A-type tasks, calculate the difference of K-A1-B1. If the difference is positive, copy the B-type tasks, then calculate the difference of K-A1-B1-C1. If the difference is positive, copy the C-type tasks, and at the same time calculate the difference of K-A1-B1-D1. If the difference is positive, copy the D-type tasks, and finally copy all the copied tasks according to step S5. Perform task resource matching processing. If the copied task computing resource processing result X and the copied task computing resource processing result Y are similar, they are outputted normally. Otherwise, if the two results X and Y are different, they are matched with other different computing nodes for further task computing resource processing. The processing result is recorded as Z. Z is compared with X and Y respectively. The calculation result of Z is used as the standard. If X and Z are different, locate the computing node with computing resource processing result X and mark it as an abnormal computing node. If Y and Z are different, locate the computing node with computing resource processing result X and mark it as an abnormal computing node. Notify relevant personnel to perform maintenance or algorithm detection on it. On the contrary, if any difference is negative, it is considered that the copy task cannot be continued, and step S5 is executed.

3. A computing resource scheduling optimization method for a large model according to claim 2, characterized in that: The weight value setting in step S3 is as follows: the weight value of the common computing node = (the resource processing capacity value of the common computing node / the sum of the resource processing capacity values ​​of all common computing nodes) × (1 / the number of requests) × the adjustment coefficient; The adjustment coefficient is set according to the big data model experiment, and the value is 1 when the number of requests is 0; The weight value configuration of the buffer computing node is the same as the weight value formula of the ordinary computing node.

4. A computing resource scheduling optimization method for a large model according to claim 3, characterized in that: In the step S4, when the idle computing nodes among the common computing nodes are regarded as buffer temporary computing nodes, the sum Q of the buffer temporary computing nodes is calculated; When executing step S5, priority is given to matching important and urgent tasks of category A, the difference is calculated as A1-(MQ), and then step S5 is continued normally; When the matching of important and urgent Class A tasks is completed, Class B tasks are assigned and the difference is calculated as B1-(MQ-A1). When Class C tasks are assigned, the difference is calculated as C1-(MQ-A1-B1). When Class D tasks are assigned, the difference is calculated as D1-(MQ-A1-B1-C1).

Images