A task scheduling method and system based on an ITSM platform

By dividing the ITSM platform node cluster into two sets and using real performance baselines to calibrate historical data, the problems of hotspot fixation and resource imbalance were solved, achieving dynamic balancing and efficient operation of task scheduling.

CN122363860APending Publication Date: 2026-07-10上海闻驿信息技术有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610699892.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-20
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

In the existing ITSM platform's task scheduling mechanism, historical success rate bias leads to problems such as hot spot solidification, self-fulfilling oracle loops, and resource utilization imbalance. This results in hot spot nodes becoming overloaded and crashing while other nodes are idle, leading to a decrease in overall cluster throughput and cascading delays.

Method used

The node cluster is divided into a first scheduling set and a second scheduling set. An unbiased performance baseline calibration and a periodic node rotation mechanism are adopted. By calibrating historical data with real performance baselines, the node ratio is dynamically adjusted to achieve a balanced distribution of resources.

Benefits of technology

It breaks the positive feedback loop of historical success rates leading to selection, prevents hotspots from becoming entrenched, corrects biases in self-fulfilling oracle models, promptly blocks overloaded hotspot nodes, revitalizes idle resources, and restores the cluster to uniform and efficient operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122363860A_ABST
    Figure CN122363860A_ABST
Patent Text Reader

Abstract

This invention discloses a task scheduling method and system based on an ITSM platform, belonging to the field of IT service management technology. The method includes the following steps: acquiring an available execution node cluster; collecting task execution records of nodes within a second scheduling set to generate a true performance baseline; performing periodic node rotation operations between the first and second scheduling sets based on historical execution data and the true performance baseline; calibrating the historical execution data of the first scheduling set; monitoring the real-time load distribution of the first scheduling set to trigger emergency node rotation operations; performing emergency node rotation operations and resetting historical data for nodes that have undergone rotation; and updating the initial node quantity ratio. This invention solves the problems of hotspot solidification, self-fulfilling oracle, and cluster resource utilization imbalance caused by historical affinity bias in existing technologies by dividing execution nodes into a first scheduling set and a second scheduling set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of IT service management technology, and in particular to a task scheduling method and system based on an ITSM platform. Background Technology

[0002] As enterprises become increasingly information-driven, ITSM platforms are responsible for scheduling and allocating a large number of tasks, such as work order processing and change execution. Existing ITSM platforms typically use scheduling methods based on historical execution data. By recording the historical success rate and average processing time of each execution node, they calculate the affinity score between the task and the node, and then assign the task to the node with the best historical performance.

[0003] However, this scheduling mechanism based on historical success rates is prone to the following problems in long-term operation: a few nodes with excellent historical performance are repeatedly selected, forming a hot spot phenomenon; hot nodes continuously undertake tasks far exceeding reasonable capacity, leading to an increase in their execution time and failure rate, while the prediction model based on historical data regards these high-load states as the norm, further reinforcing the recommendation of hot nodes, falling into a self-fulfilling oracle loop; eventually, hot nodes overload and crash, while other nodes have idle resources, causing a decrease in the overall throughput of the cluster and cascading delays. Summary of the Invention

[0004] This application provides a task scheduling method and system based on the ITSM platform, which solves the problems of hotspot solidification, self-proving oracle, and unbalanced cluster resource utilization caused by historical affinity bias in the prior art, and realizes adaptive scheduling and balanced resource distribution based on unbiased performance baseline.

[0005] This application provides a task scheduling method based on an ITSM platform, including: obtaining an available execution node cluster, and dividing the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node number ratio;

[0006] Tasks are assigned to nodes in the first scheduling set for execution based on historical execution data, and tasks are assigned to nodes in the second scheduling set for execution based on sequential polling allocation logic;

[0007] Collect task execution records from nodes within the second scheduling set to generate a true performance baseline;

[0008] Based on the historical execution data and the actual performance baseline, periodic node rotation operations are performed between the first scheduling set and the second scheduling set;

[0009] The historical execution data of the first scheduling set is calibrated using the real performance baseline;

[0010] Monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations;

[0011] Perform the emergency node rotation operation and reset the historical data of the nodes that have been rotated;

[0012] Based on the task execution failure index, the dynamic ratio of the first scheduling set and the second scheduling set is adaptively adjusted, and the initial node quantity ratio is updated.

[0013] Furthermore, the step of allocating tasks to nodes within the first scheduling set for execution based on historical execution data includes:

[0014] Configure a regular scheduling logic control unit for the first scheduling set;

[0015] For each data packet of a task to be executed, the task attribute information is parsed, and the execution data of the same type of task in the past preset sliding time window is extracted from each available execution node in the first scheduling set.

[0016] Based on similar task execution data, obtain the task success rate parameter and average execution time parameter for each available execution node, and obtain the historical execution affinity score for each available execution node;

[0017] The resource monitoring probe obtains the current resource availability metrics of each available execution node in real time.

[0018] The conventional scheduling logic control unit performs a weighted average calculation by combining historical execution affinity scores with current resource availability indicators to obtain a comprehensive matching score for each available execution node.

[0019] Iterate through the overall matching scores of all available execution nodes, select the available execution node with the highest overall matching score, mark the available execution node as the target distribution node, establish a network transmission connection between the data packet to be executed and the target distribution node, and distribute the data packet to be executed to the target distribution node for processing.

[0020] Furthermore, the step of allocating tasks to nodes within the second scheduling set for execution based on the sequential polling allocation logic includes:

[0021] Configure a random exploration scheduling logic control unit for the second scheduling set;

[0022] Construct and maintain in memory a circular index pointer to a list of available execution nodes within the second scheduling set;

[0023] Whenever a data packet of a task to be executed is received, the available execution node corresponding to the current position pointed to by the loop index pointer is directly extracted, the available execution node is set as the target allocation node, and the data packet of the task to be executed is distributed to the target allocation node;

[0024] After the distribution is complete, the circular index pointer is moved forward in the list of available execution nodes to the next adjacent node position;

[0025] When it is detected that the total number of available execution nodes in the second scheduling set has changed dynamically due to the system running status, the updated total number of available execution nodes is obtained and the historical index value corresponding to the original loop index pointer is extracted.

[0026] Perform a modulo operation on the historical index value and the updated total number of available execution nodes.

[0027] Furthermore, the steps of collecting task execution records from nodes within the second scheduling set to generate a true performance baseline include:

[0028] By deploying the monitoring agent module on the available execution nodes in the second scheduling set, a task execution log data with an independent format is generated and recorded for each data packet of the task to be executed that is sent to the available execution node;

[0029] When the task is finished, the task execution log data is asynchronously written to a completely independent data storage area;

[0030] Read all task execution log data of all available execution nodes in the second scheduling set in batches from the independent data storage area at preset time intervals, and calculate the actual average execution time data of each available execution node for each type of task;

[0031] The tasks running on each available execution node are divided into multiple ascending-order utilization ranges based on the average CPU utilization value during their execution.

[0032] Calculate the ratio of the actual execution time of each task within each occupancy range to the baseline execution time of that task on an idle resource node. Extract the non-linear increasing characteristic sequence of the ratio parameter as the occupancy range gradually increases. Directly map the non-linear increasing characteristic sequence to construct the performance decay curve parameter of the available execution node. Combine the real average execution time data with the performance decay curve parameter to encapsulate the real performance baseline.

[0033] Furthermore, the steps for performing periodic node rotation operations between the first and second scheduling sets based on historical execution data and actual performance baselines include:

[0034] Deploy a global control process component independent of the task allocation engine, and set a fixed time span parameter as a complete scheduling cycle through the scheduling cycle timer module;

[0035] At the end of each complete scheduling cycle, the global control process component automatically triggers the node set rotation process.

[0036] The global control process component first iterates through and reads the historical execution affinity scores of all available execution nodes in the first scheduling set. It then performs serialization sorting based on the descending order of the historical execution affinity scores from high to low. It then extracts a group of available execution nodes that are at the top of the sorted sequence and whose number is equal to the first preset rounding ratio of the current total number of nodes in the first scheduling set. This group of available execution nodes is then logically removed from the first scheduling set.

[0037] The synchronous global control process component traverses and reads the default rate index values ​​of all available execution nodes in the second scheduling set within the current complete scheduling cycle.

[0038] Filter out a batch of available execution nodes whose cumulative total number of received task records is not less than the preset minimum sample threshold, and select a group of peer replacement nodes whose default rate index is at the lowest level and whose number is equal to the number of nodes removed from the first scheduling set from the batch of available execution nodes. Logically remove the group of peer replacement nodes from the second scheduling set.

[0039] A set of available execution nodes removed from the first scheduling set will be registered and incorporated into the address space of the second scheduling set, and a set of peer replacement nodes removed from the second scheduling set will be registered and incorporated into the address space of the first scheduling set.

[0040] After completing the bidirectional cross-set data migration, the latest configuration rule information is broadcast synchronously to the first scheduling set and the second scheduling set.

[0041] Furthermore, the step of calibrating the historical execution data of the first scheduling set using the real performance baseline includes:

[0042] The global control process component extracts a real performance baseline data structure from components within an independent data storage area that uniquely matches each available execution node within the first scheduling set;

[0043] For any available execution node within the first scheduling set, read the apparent performance data packet. The apparent performance data packet contains the apparent average duration and apparent average success rate of the available execution node in processing tasks within a predetermined historical time window.

[0044] Perform a difference comparison operation between the apparent average execution time value and the actual average execution time data of the same type of task in the actual performance baseline data structure;

[0045] When the apparent average execution time is less than the actual average execution time and the absolute difference between the two exceeds a preset relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is overestimated by historical statistical inertia, and a first correction factor coefficient with a numerical attribute strictly less than the natural number one is generated accordingly.

[0046] When the apparent average execution time is significantly greater than the actual average execution time and the absolute difference between the two exceeds the upper limit of the relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is underestimated by the historical low load state, and a second correction factor coefficient with a numerical attribute strictly greater than the natural number one is generated accordingly.

[0047] The generated first or second correction factor coefficient is sent as a configuration update instruction to the regular scheduling logic control unit in the first scheduling set.

[0048] When the conventional scheduling logic control unit starts calculating the historical execution affinity score of each available execution node in the next complete scheduling cycle, it is required to first perform a multiplication scaling process on the original affinity score read and the corresponding first correction factor coefficient or second correction factor coefficient, and then input it into the weighted merging operation link for processing.

[0049] Furthermore, the steps of monitoring the real-time load distribution of the first scheduling set to trigger emergency node rotation operations include:

[0050] Continuously and uninterruptedly scan the real-time concurrent workload status of each available execution node in the first scheduling set;

[0051] For each monitored available execution node, extract the number of active tasks that its internal processor is currently processing in parallel, and at the same time read the nominal maximum theoretical concurrent task capacity value specified in the node's hardware configuration table.

[0052] Calculate the standardized real-time load distribution coefficient for each available execution node;

[0053] The standardized real-time load distribution coefficients of all available execution nodes in the first scheduling set are pushed into a real-time dynamic comparison queue, and a maximum load extreme value parameter and a minimum load extreme value parameter are captured simultaneously from the real-time dynamic comparison queue.

[0054] The instantaneous imbalance ratio parameter, which reflects the severity of resource tilt within the entire first scheduling set, is calculated.

[0055] The instantaneous imbalance ratio parameter generated in real time is continuously compared with the emergency imbalance alarm threshold parameter to determine its magnitude.

[0056] Once it is determined that the instantaneous imbalance ratio parameter is not less than the emergency imbalance alarm threshold parameter, a hardware-level system interrupt signal is immediately generated.

[0057] Furthermore, the steps for performing emergency node rotation operations and resetting historical data for the rotated nodes include:

[0058] Mark the heavily overloaded node group as a forced removal state;

[0059] Select a group of idle nodes with the lowest load ranking and an absolute equal number to the heavily overloaded node group;

[0060] Nodes in the forced removal state are moved into the second scheduling set, and the idle node group is moved into the first scheduling set;

[0061] Distribute the latest node affiliation mapping table to the entire network;

[0062] Perform deep data cleaning to forcibly clear the historical execution affinity score records that have been accumulated in the regular scheduling logic control unit of the nodes transferred during this emergency rotation.

[0063] After the cleaning is complete, these newly added nodes to the first scheduling set are reassigned a globally initialized baseline affinity score.

[0064] This application provides a task scheduling system based on an ITSM platform, used to implement a task scheduling method based on an ITSM platform, including:

[0065] The module includes: node cluster acquisition module, node execution module, baseline acquisition module, rotation operation module, data calibration module, aggregate monitoring module, data reset module, and adaptive adjustment module.

[0066] The node cluster module is used to obtain an available execution node cluster and divide the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node number ratio.

[0067] The node execution module is used to allocate tasks to nodes in the first scheduling set based on historical execution data, and to allocate tasks to nodes in the second scheduling set based on sequential polling allocation logic;

[0068] The baseline acquisition module is used to collect task execution records of nodes within the second scheduling set to generate a true performance baseline;

[0069] The rotation operation module is used to perform periodic node rotation operations between the first scheduling set and the second scheduling set based on the historical execution data and the actual performance baseline.

[0070] The data calibration module is used to calibrate the historical execution data of the first scheduling set using the real performance baseline;

[0071] The set monitoring module is used to monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations;

[0072] The data reset module is used to perform the emergency node rotation operation and reset the historical data of the node that has been rotated.

[0073] The adaptive adjustment module is used to dynamically and adaptively adjust the ratio of the first scheduling set and the second scheduling set based on the task execution failure index, and update the initial node quantity ratio.

[0074] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0075] By dividing nodes into first and second scheduling sets and employing round-robin allocation without historical bias, the true performance baseline of nodes is obtained, breaking the positive feedback loop of "the higher the historical success rate, the more likely they are to be selected," and curbing the solidification of hotspots. Furthermore, by periodically rotating hotspot nodes out of the first scheduling set and using the true baseline to calibrate historical data, affinity is reduced for overestimated nodes and increased for underestimated nodes, correcting model bias caused by self-fulfilling oracles and avoiding the vicious cycle of the busiest nodes being recommended more often. Going further, by monitoring load in real time and triggering emergency rotation, overloaded nodes are replaced with idle nodes and their historical data is reset, promptly preventing hotspot nodes from approaching their physical limits, preventing response time deterioration and downstream cascading delays, while simultaneously revitalizing idle resources, restoring the cluster from an unbalanced state to uniform and efficient operation. Attached Figure Description

[0076] Figure 1 A flowchart of a task scheduling method based on an ITSM platform provided in this application embodiment;

[0077] Figure 2 This is a schematic diagram of the structure of a task scheduling system based on the ITSM platform provided in an embodiment of this application. Detailed Implementation

[0078] This application provides a task scheduling method and system based on the ITSM platform, which solves the problems of hotspot solidification and self-proving oracle caused by historical affinity bias in the prior art. By dividing nodes into a first scheduling set and a second scheduling set and using the unbiased execution records of the second scheduling set to calibrate the historical affinity score of the first scheduling set, the forced rotation of hotspot nodes and dynamic balance of resource distribution are achieved.

[0079] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0080] like Figure 1 The diagram shown is a flowchart of a task scheduling method based on an ITSM platform provided in an embodiment of this application. The method is applied to a task scheduling system based on an ITSM platform and includes the following steps: obtaining an available execution node cluster and dividing the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node number ratio;

[0081] Tasks are assigned to nodes in the first scheduling set for execution based on historical execution data, and tasks are assigned to nodes in the second scheduling set for execution based on sequential polling allocation logic;

[0082] Collect task execution records from nodes within the second scheduling set to generate a true performance baseline;

[0083] Based on the historical execution data and the actual performance baseline, periodic node rotation operations are performed between the first scheduling set and the second scheduling set;

[0084] The historical execution data of the first scheduling set is calibrated using the real performance baseline;

[0085] Monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations;

[0086] Perform the emergency node rotation operation and reset the historical data of the nodes that have been rotated;

[0087] Based on the task execution failure index, the dynamic ratio of the first scheduling set and the second scheduling set is adaptively adjusted, and the initial node quantity ratio is updated.

[0088] In this embodiment, during the initialization phase, the global available execution node cluster is obtained through the platform's underlying resource manager interface. The initial node quantity ratio parameter, pre-written in the persistent configuration center server, is read. Based on the initial node quantity ratio parameter, the available execution node cluster is logically divided into a first scheduling set and a second scheduling set in the address space. After the division operation is completed, the scheduling engine initiates the dual-mode allocation logic in parallel: receiving tasks are distributed to nodes in the first scheduling set for execution based on historical execution data; in parallel, tasks are distributed to nodes in the second scheduling set for execution based on sequential round-robin allocation logic.

[0089] During the operation of the second scheduling set, the background control process collects task execution records of the nodes and generates a true performance baseline based on these records. When the timer reaches a set periodic threshold, a periodic node rotation operation is performed between the first and second scheduling sets based on historical execution data and the true performance baseline. Within the rotation period, the background control process uses the true performance baseline to perform difference comparison and coefficient calibration on the historical execution data of the first scheduling set.

[0090] Outside of regular scheduling cycles, the high-frequency real-time load monitor sub-thread monitors the real-time load distribution status of the first scheduling set. When the load distribution status triggers an imbalance alarm rule, it directly wakes up and executes an emergency node rotation operation, while simultaneously resetting and cleaning the historical data records of the nodes that underwent rotation in the database. At the end of each regular scheduling cycle, it summarizes task execution failure indicators and performs a dynamic ratio adaptive adjustment operation on the first and second scheduling sets based on these indicators to update the initial node quantity ratio parameters and overwrite them into the persistent configuration center server.

[0091] Furthermore, the step of allocating tasks to nodes within the first scheduling set for execution based on historical execution data includes:

[0092] Configure a regular scheduling logic control unit for the first scheduling set. The regular scheduling logic control unit receives data packets of tasks to be executed from the task queue.

[0093] For each data packet of a task to be executed, the task attribute information is parsed, and the execution data of the same type of task in the past preset sliding time window is extracted from each available execution node in the first scheduling set.

[0094] Based on the execution data of similar tasks, the task success rate parameter and the average execution time parameter of each available execution node are obtained. The task success rate parameter and the average execution time parameter are fused together to calculate the historical execution affinity score of each available execution node.

[0095] The resource monitoring probe obtains the current resource availability indicators of each available execution node in real time. The current resource availability indicators include the CPU idle rate and the amount of memory free.

[0096] The conventional scheduling logic control unit performs a weighted average calculation by combining historical execution affinity scores with current resource availability indicators to obtain a comprehensive matching score for each available execution node.

[0097] Iterate through the overall matching scores of all available execution nodes, select the available execution node with the highest overall matching score, mark the available execution node as the target distribution node, establish a network transmission connection between the data packet to be executed and the target distribution node, and distribute the data packet to be executed to the target distribution node for processing.

[0098] In this embodiment, for the step of allocating tasks to nodes within the first scheduling set based on historical execution data, a conventional scheduling logic control unit is configured for the first scheduling set.

[0099] When the conventional scheduling logic control unit receives the data packet of the task to be executed from the task queue, it parses and extracts the task attribute information, and retrieves the execution data of the same type of task from the underlying historical relational database for each available execution node in the first scheduling set within the past preset sliding time window.

[0100] The historical execution affinity score for each available execution node is calculated using the following formula. :

[0101] ;

[0102] in, This represents the task success rate of the available execution node within the sliding time window. It is obtained by dividing the number of successfully executed task records by the total number of task records allocated to that node. This parameter represents the average execution time of similar tasks performed by this node. It is obtained by calculating the arithmetic mean of the actual execution times of similar tasks within a sliding time window. This represents the standard estimated time for similar tasks in a cluster baseline environment, calculated in advance based on the entire historical sample. and These are normalized weight coefficients, and satisfy... .

[0103] Simultaneously, the resource monitoring probe obtains the current resource availability metrics of each available execution node in real time through the kernel calls of the operation. The calculation formula is:

[0104] ;

[0105] in, This represents the CPU idle rate. This is the memory free rate value (obtained by dividing the number of available memory bytes by the total number of physical memory bytes). These are parameters for adjusting resource preference weights. The conventional scheduling logic control unit ultimately calculates the overall matching score. : in, and satisfy Traverse all available execution nodes. The node with the largest value is selected, marked as the target distribution node, a network transmission connection is established, and the data packet of the task to be executed is distributed.

[0106] Furthermore, the step of allocating tasks to nodes within the second scheduling set for execution based on the sequential polling allocation logic includes:

[0107] A random exploration scheduling logic control unit is configured for the second scheduling set. When the random exploration scheduling logic control unit receives the data packet of the task to be executed, it actively masks the historical execution data information of the data packet of the task to be executed and the current resource availability historical statistics of all available execution nodes in the second scheduling set, and uses memoryless sequential round-robin allocation logic to process the data packet of the task to be executed.

[0108] Construct and maintain in memory a circular index pointer to a list of available execution nodes within the second scheduling set;

[0109] Whenever a data packet of a task to be executed is received, the available execution node corresponding to the current position pointed to by the loop index pointer is directly extracted, the available execution node is set as the target allocation node, and the data packet of the task to be executed is distributed to the target allocation node;

[0110] After the distribution is completed, the circular index pointer is moved forward in the list of available execution nodes to the next adjacent node position, forming a cyclical allocation mechanism;

[0111] When it is detected that the total number of available execution nodes in the second scheduling set has changed dynamically due to the system running status, the updated total number of available execution nodes is obtained and the historical index value corresponding to the original loop index pointer is extracted.

[0112] Perform a modulo operation on the historical index value and the updated total number of available execution nodes. Based on the result of the modulo operation, recalculate and determine the new position of the circular index pointer. This ensures that the circular index pointer maintains a compliant pointing state even after the number of nodes changes dynamically, preventing out-of-bounds data allocation errors.

[0113] In this embodiment, for the step of allocating tasks to nodes within the second scheduling set based on sequential polling allocation logic, a random exploration scheduling logic control unit is configured for the second scheduling set. When this unit receives a data packet of a task to be executed, it actively blocks database query requests for historical execution data and requests to call resource monitoring probes.

[0114] The random exploration scheduling logic control unit constructs a circular index pointer in the memory stack area to a list of available execution nodes within the second scheduling set. In the initial state, Upon receiving a data packet for a task to be executed, extract... The available execution node corresponding to the current location is used as the target allocation node and the task is distributed.

[0115] Then, the pointer stepping operation is performed: When a node addition / deletion heartbeat signal is received, and a change in the total number of available execution nodes in the second scheduling set is detected, the updated total number of available execution nodes is obtained. Extract the historical index value corresponding to the original circular index pointer. And perform a modulo operation to reposition the pointer: Will Assigning a value to the loop index pointer ensures that the addressing range of the loop index pointer is strictly defined within... to Within a closed interval, to prevent memory access exceptions caused by array out-of-bounds.

[0116] Furthermore, the steps of collecting task execution records from nodes within the second scheduling set to generate a true performance baseline include:

[0117] By deploying the monitoring agent module on the available execution nodes in the second scheduling set, a task execution log data with an independent format is generated and recorded for each data packet of the task to be executed that is sent to the available execution node;

[0118] The task execution log data specifically includes the task's unique identifier, the physical identifier of the available execution node, the actual start timestamp of the task at the operating system level, the actual end timestamp after the task is completed, and the average CPU utilization and average physical memory utilization of the available execution node throughout the entire lifecycle of the task.

[0119] When the task is finished, the task execution log data is asynchronously written to a completely independent data storage area, which is physically isolated from the affinity calculation module of the regular scheduling logic control unit.

[0120] The background control process reads all task execution log data of all available execution nodes in the second scheduling set from an independent data storage area at preset time intervals, and calculates the actual average execution time data of each available execution node for each type of task.

[0121] The tasks running on each available execution node are divided into multiple ascending-order utilization ranges based on the average CPU utilization value during their execution.

[0122] Calculate the ratio of the actual execution time of each task within each occupancy range to the baseline execution time of that task on an idle resource node. Extract the non-linear increasing characteristic sequence of the ratio parameter as the occupancy range gradually increases. Directly map the non-linear increasing characteristic sequence to construct the performance decay curve parameter of the available execution node. Combine the real average execution time data with the performance decay curve parameter to encapsulate the real performance baseline.

[0123] In this embodiment, for the step of collecting task execution records from nodes within the second scheduling set to generate a true performance baseline, the monitoring agent module deployed on the available execution nodes captures the actual start timestamps of the tasks in kernel mode. Compared with the actual end timestamp The actual time consumed by the calculation task .

[0124] exist to Within a given time interval, CPU utilization is acquired at a fixed sampling frequency, and its arithmetic mean is calculated as the average CPU utilization value. This information is encapsulated into task execution log data in an independent format and written to a completely physically isolated independent data storage area via an asynchronous message queue. The background control process reads all task execution log data from a specified past time window period.

[0125] When constructing the performance degradation curve parameters, tasks are divided into multiple ascending utilization intervals based on the average CPU utilization (each interval is defined as 10% utilization). The actual execution time of tasks within each utilization interval is then calculated. Compared to the baseline time consumed by this task on an idle resource node ratio parameter :

[0126] ;

[0127] Extracting multiple occupancy rate ranges The sequence set is fitted with a polynomial using the least squares method, outputting the coefficients of the nonlinearly increasing characteristic sequence equation. These coefficients are then combined and encapsulated with the actual average execution time data of the node within the time window to generate a true performance baseline.

[0128] Furthermore, the steps for performing periodic node rotation operations between the first and second scheduling sets based on historical execution data and actual performance baselines include:

[0129] Deploy a global control process component independent of the task allocation engine, and set a fixed time span parameter as a complete scheduling cycle through the scheduling cycle timer module;

[0130] At the end of each complete scheduling cycle, the global control process component automatically triggers the node set rotation process.

[0131] The global control process component first iterates through and reads the historical execution affinity scores of all available execution nodes in the first scheduling set. It then performs serialization sorting based on the descending order of the historical execution affinity scores from high to low. It then extracts a group of available execution nodes that are at the top of the sorted sequence and whose number is equal to the first preset rounding ratio of the current total number of nodes in the first scheduling set. This group of available execution nodes is then logically removed from the first scheduling set.

[0132] The synchronous global control process component traverses and reads the default rate index values ​​of all available execution nodes in the second scheduling set within the current complete scheduling cycle. The default rate index value is obtained by dividing the sum of the number of node execution failure records and the number of timeout default records by the total number of task records received by the node.

[0133] Filter out a batch of available execution nodes whose cumulative total number of received task records is not less than the preset minimum sample threshold, and select a group of peer replacement nodes whose default rate index is at the lowest level and whose number is equal to the number of nodes removed from the first scheduling set from the batch of available execution nodes. Logically remove the group of peer replacement nodes from the second scheduling set.

[0134] A set of available execution nodes removed from the first scheduling set will be registered and incorporated into the address space of the second scheduling set, and a set of peer replacement nodes removed from the second scheduling set will be registered and incorporated into the address space of the first scheduling set.

[0135] After completing the bidirectional cross-set data migration, the latest configuration rule information is broadcast synchronously to the first scheduling set and the second scheduling set.

[0136] In this embodiment, for the step of performing periodic node rotation, a scheduling period timer module is initialized internally within the global control process component, independent of the task allocation engine, with the time span parameter set to a fixed number of seconds. Resetting the timer triggers the rotation process. The global control process component iterates through all nodes in the first scheduling set, sorting them in descending order based on their historical execution affinity scores. The rotation is then determined according to the current total number of nodes in the first scheduling set. Calculate the number of segments corresponding to the first preset rounding ratio. : Before truncating the sorted sequence Each node is logically removed from the first scheduling set.

[0137] Simultaneously, the second scheduling set nodes are traversed to calculate the default rate index value within the current complete scheduling cycle. :

[0138] ;

[0139] in, The number of execution failure records. This represents the number of overdue default records. To receive the total number of task records, filter out... Nodes not lower than the preset minimum sample threshold, according to Perform ascending sorting and extract the first part of the sorted sequence. Each node is logically removed from the second scheduling set as a peer replacement node.

[0140] By updating the underlying routing mapping table, nodes removed from the first scheduling set are added to the routing mapping table of the second scheduling set, and nodes removed from the second scheduling set are added to the routing mapping table of the first scheduling set.

[0141] Furthermore, the step of calibrating the historical execution data of the first scheduling set using the real performance baseline includes:

[0142] The global control process component extracts a real performance baseline data structure from components within an independent data storage area that uniquely matches each available execution node within the first scheduling set;

[0143] For any available execution node within the first scheduling set, the global control process component reads the apparent performance data packet implicitly represented by the historical execution affinity score from the regular scheduling logic control unit. The apparent performance data packet contains the apparent average duration and apparent average success rate of the available execution node in processing tasks within a given historical time window.

[0144] Perform a difference comparison operation between the apparent average execution time value and the actual average execution time data of the same type of task in the actual performance baseline data structure;

[0145] When the apparent average execution time is less than the actual average execution time and the absolute difference between the two exceeds a preset relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is overestimated by historical statistical inertia, and a first correction factor coefficient with a numerical attribute strictly less than the natural number one is generated accordingly.

[0146] When the apparent average execution time is significantly greater than the actual average execution time and the absolute difference between the two exceeds the upper limit of the relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is underestimated by the historical low load state, and a second correction factor coefficient with a numerical attribute strictly greater than the natural number one is generated accordingly.

[0147] The generated first or second correction factor coefficient is sent as a configuration update instruction to the regular scheduling logic control unit in the first scheduling set.

[0148] When the conventional scheduling logic control unit starts calculating the historical execution affinity score of each available execution node in the next complete scheduling cycle, it is required to first perform a multiplication scaling process on the original affinity score read and the corresponding first correction factor coefficient or second correction factor coefficient, and then input it into the weighted merging operation link for processing.

[0149] In this embodiment, for the step of calibrating the historical execution data of the first scheduling set using the true performance baseline, the global control process component records the apparent average duration value inside the conventional scheduling logic control unit as... The actual average execution time data extracted from the real performance baseline data structure is denoted as... .

[0150] Set the upper limit of the relative deviation tolerance threshold as Perform the difference comparison logic operation:

[0151] Condition 1: If The node performance was determined to be overestimated. The first correction factor coefficient was generated. ,in Through formula The calculation shows that at this time The numerical property of is strictly less than the natural number one.

[0152] Condition 2: If The performance of the determined nodes is severely underestimated. A second correction factor coefficient is generated. ,in Through formula The calculation shows that at this time Its numerical attribute is greater than the natural number one.

[0153] When the regular scheduling logic control unit starts the next complete scheduling cycle, the queried original affinity score is... Perform multiplication value scaling to generate calibrated historical execution affinity scores. :

[0154] ;

[0155] in This refers to either the received first correction factor coefficient or the second correction factor coefficient. After calculation, [the following will be done]. It is then incorporated into the weighted merging operation chain for processing.

[0156] Furthermore, the steps of monitoring the real-time load distribution of the first scheduling set to trigger emergency node rotation operations include:

[0157] The global control process component activates the high-frequency real-time load monitor sub-thread, which continuously and uninterruptedly scans the real-time concurrent workload status of each available execution node in the first scheduling set.

[0158] For each monitored available execution node, extract the number of active tasks that its internal processor is currently processing in parallel, and at the same time read the nominal maximum theoretical concurrent task capacity value specified in the node's hardware configuration table.

[0159] The number of currently active tasks being processed in parallel is used as the numerator parameter, and the nominal maximum theoretical concurrent task capacity is used as the denominator parameter. A division calculation is performed to obtain the standardized real-time load distribution coefficient for each available execution node.

[0160] The standardized real-time load distribution coefficients of all available execution nodes in the first scheduling set are pushed into a real-time dynamic comparison queue. An extreme value search algorithm is used to simultaneously capture a maximum load extreme value parameter and a minimum load extreme value parameter from the real-time dynamic comparison queue.

[0161] Extract the maximum load extreme value parameter and divide it by the minimum load extreme value parameter to calculate the instantaneous imbalance ratio parameter that reflects the severity of resource tilt within the entire first scheduling set;

[0162] An emergency imbalance alarm threshold parameter issued by the operation and maintenance management terminal is pre-loaded in memory, and the instantaneous imbalance ratio parameter calculated in real time is continuously compared with the emergency imbalance alarm threshold parameter to determine its magnitude.

[0163] Once it is determined that the instantaneous imbalance ratio parameter is not less than the emergency imbalance alarm threshold parameter, the global control process component immediately generates a hardware-level system interrupt signal, forcibly bypasses and suspends the countdown blocking logic of the ordinary scheduling cycle timer module, and immediately activates the emergency node rotation processing program with the highest system execution priority, thereby triggering physical state intervention and control at the underlying architecture level.

[0164] In this embodiment, for the step of monitoring the real-time load distribution of the first scheduling set to trigger emergency node rotation operations, the high-frequency real-time load monitor sub-thread performs a check on each controlled available execution node. Extract the number of active tasks currently being processed in parallel. Read the nominal maximum theoretical concurrent task capacity value. Calculate the standardized real-time load distribution coefficient :

[0165] ;

[0166] All nodes Push the value into the real-time dynamic comparison queue. Call the extreme value search algorithm function to output the maximum value in the queue. and minimum value .

[0167] Calculate the instantaneous imbalance ratio parameter :

[0168] ;

[0169] When the judgment condition When it was established ( (For the emergency imbalance alarm threshold parameter pre-loaded in memory), a hardware-level interrupt signal is sent to the interrupt controller of the operation. By modifying the priority bit of the process scheduling table, the countdown blocking logic of the ordinary scheduling cycle timer module is forcibly suspended, and the memory entry point of the emergency node rotation handler is immediately activated.

[0170] Furthermore, the steps for performing emergency node rotation operations and resetting historical data for the rotated nodes include:

[0171] When the emergency node rotation process is successfully activated by the system interrupt signal, the global control process component immediately rearranges all available execution nodes in the first scheduling set according to the descending order of the standardized real-time load distribution coefficient from high to low, and cuts out a group of heavily overloaded nodes with the highest load ranking and a number that accounts for the second preset rounding ratio of the current total number of nodes in the first scheduling set, and marks the group of heavily overloaded nodes as a forced removal state.

[0172] Simultaneously, within the second scheduling set, the nodes are rearranged according to the same ascending order of load coefficient from low to high, and a group of idle nodes with the lowest load ranking and an absolute equal number to the group of overloaded nodes is selected.

[0173] Perform atomic swap operations at the physical pointer level to move nodes in the forced-out state into the second scheduling set, and move the group of idle nodes into the first scheduling set;

[0174] After the transfer operation is completed, the status refresh broadcast mechanism is initiated to send the latest node affiliation mapping table to the entire network;

[0175] The global control process component performs a deep data cleaning action, forcibly clearing the historical execution affinity score records that the transferred nodes have accumulated in the regular scheduling logic control unit during this emergency rotation. Specifically, the cleaning action is to directly erase all mapping entries in the underlying historical affinity relationship database table of these transferred nodes.

[0176] After the cleaning is completed, a globally initialized baseline affinity score is reassigned to the nodes newly added to the first scheduling set. This baseline affinity score is calculated by extracting the arithmetic mean of the remaining affinity scores of all normal nodes currently remaining in the first scheduling set. Finally, the monitoring alarm interception status bit of the instantaneous imbalance ratio parameter is reset and the normal system monitoring cycle is restored.

[0177] In this embodiment, for the steps of performing emergency node rotation operations and resetting historical data for nodes that have undergone rotation, after the interrupt signal is activated, the global control process component calculates the interception quantity within the first scheduling set. .

[0178] in accordance with Sort in descending order, extract the first... Several nodes form a heavily overloaded node group and are marked as being forced to be removed. Within the second scheduling set, based on... Sort in ascending order and extract the first part. These nodes form a group of idle nodes.

[0179] Under the protection of the memory lock mechanism, atomic swap operations at the physical pointer level are performed to update the ownership configuration of both sets.

[0180] After the transfer is complete, using database structured query language commands (such as the DELETE statement), all historical mapping entries for the newly transferred nodes to the second set are directly deleted from the underlying historical affinity relation database table. For nodes newly added to the first scheduling set, the number of previously normal nodes remaining in the first scheduling set (let's say the number is...) is calculated. The arithmetic mean of the remaining affinity scores is used as the baseline affinity score for global initialization. :

[0181]

[0182] Will Insert the data into the database table as the initial attribute values ​​for these newly added nodes. After the data insertion is complete, set the monitoring alarm interception status register for the instantaneous imbalance ratio parameter to zero.

[0183] like Figure 2 The diagram shown is a structural schematic of a task scheduling system based on an ITSM platform provided in this application embodiment. The task scheduling system based on an ITSM platform provided in this application embodiment includes: a node cluster acquisition module, a node execution module, a baseline acquisition module, a rotation operation module, a data calibration module, a collection monitoring module, a data reset module, and an adaptive adjustment module.

[0184] The node cluster module is used to obtain an available execution node cluster and divide the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node number ratio.

[0185] The node execution module is used to allocate tasks to nodes in the first scheduling set based on historical execution data, and to allocate tasks to nodes in the second scheduling set based on sequential polling allocation logic;

[0186] The baseline acquisition module is used to collect task execution records of nodes within the second scheduling set to generate a true performance baseline;

[0187] The rotation operation module is used to perform periodic node rotation operations between the first scheduling set and the second scheduling set based on the historical execution data and the actual performance baseline.

[0188] The data calibration module is used to calibrate the historical execution data of the first scheduling set using the real performance baseline;

[0189] The set monitoring module is used to monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations;

[0190] The data reset module is used to perform the emergency node rotation operation and reset the historical data of the node that has been rotated.

[0191] The adaptive adjustment module is used to dynamically and adaptively adjust the ratio of the first scheduling set and the second scheduling set based on the task execution failure index, and update the initial node quantity ratio.

[0192] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0193] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0194] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0195] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0196] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0197] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A task scheduling method based on an ITSM platform, characterized in that, Includes the following steps: Obtain an available execution node cluster, and divide the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node quantity ratio; Tasks are assigned to nodes in the first scheduling set for execution based on historical execution data, and tasks are assigned to nodes in the second scheduling set for execution based on sequential polling allocation logic; Collect task execution records from nodes within the second scheduling set to generate a true performance baseline; Based on the historical execution data and the actual performance baseline, periodic node rotation operations are performed between the first scheduling set and the second scheduling set; The historical execution data of the first scheduling set is calibrated using the real performance baseline; Monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations; Perform the emergency node rotation operation and reset the historical data of the nodes that have been rotated; Based on the task execution failure index, the dynamic ratio of the first scheduling set and the second scheduling set is adaptively adjusted, and the initial node quantity ratio is updated.

2. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps of assigning tasks to nodes within the first scheduling set based on historical execution data include: Configure a regular scheduling logic control unit for the first scheduling set; For each data packet of a task to be executed, the task attribute information is parsed, and the execution data of the same type of task in the past preset sliding time window is extracted from each available execution node in the first scheduling set. Based on similar task execution data, obtain the task success rate parameter and average execution time parameter for each available execution node, and obtain the historical execution affinity score for each available execution node; The resource monitoring probe obtains the current resource availability metrics of each available execution node in real time. The conventional scheduling logic control unit performs a weighted average calculation by combining historical execution affinity scores with current resource availability indicators to obtain a comprehensive matching score for each available execution node. Iterate through the overall matching scores of all available execution nodes, select the available execution node with the highest overall matching score, mark the available execution node as the target distribution node, establish a network transmission connection between the data packet to be executed and the target distribution node, and distribute the data packet to be executed to the target distribution node for processing.

3. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for assigning tasks to nodes within the second scheduling set for execution based on sequential polling allocation logic include: Configure a random exploration scheduling logic control unit for the second scheduling set; Construct and maintain in memory a circular index pointer to a list of available execution nodes within the second scheduling set; Whenever a data packet of a task to be executed is received, the available execution node corresponding to the current position pointed to by the loop index pointer is directly extracted, the available execution node is set as the target allocation node, and the data packet of the task to be executed is distributed to the target allocation node; After the distribution is complete, the circular index pointer is moved forward in the list of available execution nodes to the next adjacent node position; When it is detected that the total number of available execution nodes in the second scheduling set has changed dynamically due to the system running status, the updated total number of available execution nodes is obtained and the historical index value corresponding to the original loop index pointer is extracted. Perform a modulo operation on the historical index value and the updated total number of available execution nodes.

4. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for collecting task execution records from nodes within the second scheduling set to generate a true performance baseline include: By deploying the monitoring agent module on the available execution nodes in the second scheduling set, a task execution log data with an independent format is generated and recorded for each data packet of the task to be executed that is sent to the available execution node; When the task is finished, the task execution log data is asynchronously written to a completely independent data storage area; Read all task execution log data of all available execution nodes in the second scheduling set in batches from the independent data storage area at preset time intervals, and calculate the actual average execution time data of each available execution node for each type of task; The tasks running on each available execution node are divided into multiple ascending-order utilization ranges based on the average CPU utilization value during their execution. Calculate the ratio of the actual execution time of each task within each occupancy range to the baseline execution time of that task on an idle resource node. Extract the non-linear increasing characteristic sequence of the ratio parameter as the occupancy range gradually increases. Directly map the non-linear increasing characteristic sequence to construct the performance decay curve parameter of the available execution node. Combine the real average execution time data with the performance decay curve parameter to encapsulate the real performance baseline.

5. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for performing periodic node rotation operations between the first and second scheduling sets based on historical execution data and actual performance baselines include: Deploy a global control process component independent of the task allocation engine, and set a fixed time span parameter as a complete scheduling cycle through the scheduling cycle timer module; At the end of each complete scheduling cycle, the global control process component automatically triggers the node set rotation process. The global control process component first iterates through and reads the historical execution affinity scores of all available execution nodes in the first scheduling set. It then performs serialization sorting based on the descending order of the historical execution affinity scores from high to low. It then extracts a group of available execution nodes that are at the top of the sorted sequence and whose number is equal to the first preset rounding ratio of the current total number of nodes in the first scheduling set. This group of available execution nodes is then logically removed from the first scheduling set. The synchronous global control process component traverses and reads the default rate index values ​​of all available execution nodes in the second scheduling set within the current complete scheduling cycle. Filter out a batch of available execution nodes whose cumulative total number of received task records is not less than the preset minimum sample threshold, and select a group of peer replacement nodes whose default rate index is at the lowest level and whose number is equal to the number of nodes removed from the first scheduling set from the batch of available execution nodes. Logically remove the group of peer replacement nodes from the second scheduling set. A set of available execution nodes removed from the first scheduling set will be registered and incorporated into the address space of the second scheduling set, and a set of peer replacement nodes removed from the second scheduling set will be registered and incorporated into the address space of the first scheduling set. After completing the bidirectional cross-set data migration, the latest configuration rule information is broadcast synchronously to the first scheduling set and the second scheduling set.

6. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for calibrating the historical execution data of the first scheduling set using a true performance baseline include: The global control process component extracts a real performance baseline data structure from components within an independent data storage area that uniquely matches each available execution node within the first scheduling set; For any available execution node within the first scheduling set, read the apparent performance data packet. The apparent performance data packet contains the apparent average duration and apparent average success rate of the available execution node in processing tasks within a predetermined historical time window. Perform a difference comparison operation between the apparent average execution time value and the actual average execution time data of the same type of task in the actual performance baseline data structure; When the apparent average execution time is less than the actual average execution time and the absolute difference between the two exceeds a preset relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is overestimated by historical statistical inertia, and a first correction factor coefficient with a numerical attribute strictly less than the natural number one is generated accordingly. When the apparent average execution time is significantly greater than the actual average execution time and the absolute difference between the two exceeds the upper limit of the relative deviation tolerance threshold, it is determined that the performance capability of the available execution node is underestimated by the historical low load state, and a second correction factor coefficient with a numerical attribute strictly greater than the natural number one is generated accordingly. The generated first or second correction factor coefficient is sent as a configuration update instruction to the regular scheduling logic control unit in the first scheduling set. When the conventional scheduling logic control unit starts calculating the historical execution affinity score of each available execution node in the next complete scheduling cycle, it is required to first perform a multiplication scaling process on the original affinity score read and the corresponding first correction factor coefficient or second correction factor coefficient, and then input it into the weighted merging operation link for processing.

7. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for monitoring the real-time load distribution of the first scheduling set to trigger emergency node rotation operations include: Continuously and uninterruptedly scan the real-time concurrent workload status of each available execution node in the first scheduling set; For each monitored available execution node, extract the number of active tasks that its internal processor is currently processing in parallel, and at the same time read the nominal maximum theoretical concurrent task capacity value specified in the node's hardware configuration table. Calculate the standardized real-time load distribution coefficient for each available execution node; The standardized real-time load distribution coefficients of all available execution nodes in the first scheduling set are pushed into a real-time dynamic comparison queue, and a maximum load extreme value parameter and a minimum load extreme value parameter are captured simultaneously from the real-time dynamic comparison queue. The instantaneous imbalance ratio parameter, which reflects the severity of resource tilt within the entire first scheduling set, is calculated. The instantaneous imbalance ratio parameter generated in real time is continuously compared with the emergency imbalance alarm threshold parameter to determine its magnitude. Once it is determined that the instantaneous imbalance ratio parameter is not less than the emergency imbalance alarm threshold parameter, a hardware-level system interrupt signal is immediately generated.

8. The task scheduling method based on the ITSM platform as described in claim 1, characterized in that, The steps for performing emergency node rotation and resetting historical data for the rotated nodes include: Mark the heavily overloaded node group as a forced removal state; Select a group of idle nodes with the lowest load ranking and an absolute equal number to the heavily overloaded node group; Nodes in the forced removal state are moved into the second scheduling set, and the idle node group is moved into the first scheduling set; Distribute the latest node affiliation mapping table to the entire network; Perform deep data cleaning to forcibly clear the historical execution affinity score records that have been accumulated in the regular scheduling logic control unit of the nodes transferred during this emergency rotation. After the cleaning is complete, these newly added nodes to the first scheduling set are reassigned a globally initialized baseline affinity score.

9. A task scheduling system based on an ITSM platform, used to implement the task scheduling method based on an ITSM platform as described in any one of claims 1-8, characterized in that, include: The module includes: node cluster acquisition module, node execution module, baseline acquisition module, rotation operation module, data calibration module, aggregate monitoring module, data reset module, and adaptive adjustment module. The node cluster module is used to obtain an available execution node cluster and divide the available execution node cluster into a first scheduling set and a second scheduling set according to the initial node number ratio. The node execution module is used to allocate tasks to nodes in the first scheduling set based on historical execution data, and to allocate tasks to nodes in the second scheduling set based on sequential polling allocation logic; The baseline acquisition module is used to acquire task execution records of nodes within the second scheduling set to generate a true performance baseline; The rotation operation module is used to perform periodic node rotation operations between the first scheduling set and the second scheduling set based on the historical execution data and the actual performance baseline. The data calibration module is used to calibrate the historical execution data of the first scheduling set using the real performance baseline; The set monitoring module is used to monitor the real-time load distribution of the first scheduling set to trigger emergency node rotation operations; The data reset module is used to perform the emergency node rotation operation and reset the historical data of the node that has been rotated. The adaptive adjustment module is used to dynamically and adaptively adjust the ratio of the first scheduling set and the second scheduling set based on the task execution failure index, and update the initial node quantity ratio.