Heterogeneous computing power pool task adaptive migration and dynamic scheduling method and system

Abstract

The invention discloses a heterogeneous computing power pool task adaptive migration and dynamic scheduling method and system, and relates to the technical field of computer scheduling and artificial intelligence computing power. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method comprises the following steps: S1, constructing a scheduling mapping basic data set; s2, analyzing an adaptive structure difference between a task resource request and a node resource state based on the scheduling mapping basic data set, and analyzing resource load fluctuation characteristics in a node operation process; s3, performing multi-dimensional matching optimal path judgment in combination with the scheduling stability level and the node scheduling resource state; and S4, executing a task container transfer and context recovery process according to the path channel information and the node resource binding structure. The problem that the actual reasoning efficiency is obviously lower than the theoretical calculation capability due to the fact that queuing time consumption and context switching overhead of the GPU instruction are not tracked during multi-round concurrent task calling is solved.

Description

Technical Field

[0001] This invention relates to the fields of computer scheduling and artificial intelligence computing power technology, specifically to a method and system for adaptive migration and dynamic scheduling of tasks in heterogeneous computing power pools. Background Technology

[0002] With the increasing demands for training and inference of large-scale artificial intelligence models, heterogeneous computing power pools, as scheduling infrastructures that integrate multiple types of computing resources, are widely used to handle the dynamic distribution and resource coordination of multi-source heterogeneous tasks. These systems are gradually becoming key architectures supporting high-density computing in computing service platforms, intelligent data centers, and research institutions, and are continuously developing for typical scenarios such as complex model operation, high-concurrency calls, and rapid resource scheduling.

[0003] For example, the invention patent with announcement number CN119718692B discloses a heterogeneous computing power scheduling method, device, and storage medium. The method includes: responding to a heterogeneous computing power scheduling task sent by a user, classifying the heterogeneous computing power resources of the scheduling system according to the applicable scenarios and types of computing power resources to obtain resource adaptation scenarios and resource adaptation vendors; dividing the heterogeneous computing power resources and the heterogeneous computing power scheduling task into resource pools and task pools respectively; constructing a computing power scheduling objective function; solving the computing power scheduling objective function to obtain resource allocation results; monitoring the task pool, obtaining the current task from the task polling queue, and allocating the computing power resources in the resource pool to the current task according to the resource allocation results, the resource adaptation scenarios, and the resource adaptation vendors.

[0004] For example, the invention patent with announcement number CN120469784B discloses a method and system for dynamic scheduling of heterogeneous multi-task computing power, including: responding to multiple computing task demands from users, analyzing and sorting the priorities of computing tasks to obtain a task priority order; classifying heterogeneous computing resources according to their resource status, placing different categories of computing resources into corresponding resource pools to obtain multiple resource pools; combining the task priority order and the multiple resource pools, allocating corresponding categories of computing resources to each task through a preset resource matching mechanism; during task computing, when the computing resources of the first task are insufficient, preempting the computing resources of the second task through a preset resource preemption mechanism, increasing the priority of the second task after resource preemption, and reallocating computing resources; this application can allocate corresponding computing resources based on task priorities and perform dynamic resource scheduling when a computing resource gap occurs.

[0005] There are common problems such as the lack of identification of task queuing time, the lack of tracking of non-computation time consumed by GPU instruction execution and context switching, and the lack of dynamic perception of changes in node running status. Scheduling behavior is mostly based on the total task time, which is difficult to accurately reflect the actual execution efficiency and is prone to scheduling path imbalance, resource utilization deviation and task response delay.

[0006] To address the above issues, there is an urgent need for a method and system for adaptive migration and dynamic scheduling of tasks in heterogeneous computing power pools. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for adaptive migration and dynamic scheduling of tasks in heterogeneous computing power pools. This solves the problem that the actual inference efficiency is significantly lower than the theoretical computing power because the GPU instruction queuing time and context switching overhead are not tracked when tasks are called in multiple rounds of concurrent calls.

[0009] Technical solution

[0010] To achieve the above objectives, this invention provides the following technical solution: a heterogeneous computing power pool task adaptive migration and dynamic scheduling method, comprising the following steps: S1, collecting task resource request information, computing power node running status information, and scheduling environment status information, constructing a scheduling mapping basic dataset, and preprocessing the scheduling mapping basic dataset; S2, based on the scheduling mapping basic dataset, analyzing the adaptation structure differences between task resource requests and node resource status, analyzing the resource load fluctuation characteristics during node operation, classifying scheduling stability levels according to the judgment results, and adjusting the task scheduling priority strategy; S3, combining the scheduling stability level and node scheduling resource status to perform multi-dimensional matching and optimal path judgment, filtering target nodes according to the path optimization strategy, and generating node resource binding structure and path channel information; S4, performing task container transfer and context recovery process according to the path channel information and node resource binding structure, and feeding back the migration execution status and resource release information to the scheduling mapping basic dataset.

[0011] Furthermore, the specific steps for collecting task resource request information, computing node running status information, and scheduling environment status information to construct the scheduling mapping basic dataset are as follows: Collect multi-source scheduling information generated during the operation of the heterogeneous computing pool by the task submission interface, node resource monitoring system, task running log unit, queue scheduling management system, and scheduler global control component. This includes: resource dimension request structure, real-time available resource snapshots of nodes, resource category structure definitions, task queue length, historical resource request records, historical resource snapshots of nodes, historical load rate sequences of nodes, sliding window period count, and continuous period sampling interval time. All information is accompanied by a timestamp. Simultaneously, call the historical evaluation parameters of the queue status in the scheduling system configuration file. Align the multi-source scheduling information with the task execution cycle and node status cycle according to time, and combine the node resource scheduling path relationship to organize resource requests, node status, scheduling behavior, and timing control information to construct the scheduling mapping basic dataset.

[0012] Furthermore, the specific steps for preprocessing the scheduling mapping base dataset are as follows: Addressing the heterogeneous structure and periodic offset characteristics of the scheduling mapping base dataset across time, node, and resource dimensions, a unified cross-dimensional alignment mechanism is constructed. This involves identifying and removing abnormal switching segments and format drift segments using the median absolute deviation method of the sliding interval combined with a clustering algorithm based on periodic segment partitioning, unifying periodic boundaries and node state structures, and reconstructing the temporal continuity representation of the scheduling path. Utilizing the mapping relationship between node periodic behavior density and resource call trajectories, the structural representation of scheduling state data blocks is reconstructed based on a dynamic time series alignment method. Nonlinear segments are regularized and redundant path information is removed using trajectory compression rules. Finally, the entire scheduling mapping base dataset undergoes mean standardization and maximum / minimum value normalization.

[0013] Further, the specific steps for parsing the adaptation structure differences between task resource requests and node resource states based on the scheduling mapping basic dataset are as follows: According to the current periodic resource dimension request structure, read the call fields of all resource types and extract the request values ​​of the i-th and j-th resource types; synchronously read the contents of all resource fields of the current node from the node's real-time available resource snapshot and extract the available values ​​of the i-th and j-th resource types; parse the resource category structure definition, extract the number of valid resource dimension identifiers, and obtain the total number of resource categories; divide the request value of the i-th resource type by the available value of the i-th resource type to obtain the adaptation ratio value of the i-th resource type; divide the request value of the j-th resource type by the available value of the j-th resource type, iterate through all resource categories, sum them, and divide by... The total number of resource categories is used to obtain the average resource adaptation ratio. The average resource adaptation ratio is subtracted from the adaptation ratio of the i-th resource category, and the absolute value of the difference is taken to obtain the adaptation deviation of the i-th resource category. All resource adaptation deviations are summed and divided by the total number of resource categories to obtain the heterogeneous adaptation dispersion value. The heterogeneous adaptation dispersion value is compared with the structural deviation control threshold in real time. When the heterogeneous adaptation dispersion value is greater than or equal to the structural deviation control threshold, the current task is marked as a structurally unbalanced task, the target node candidate list is reconstructed, and node paths with structural mismatch are removed. When the heterogeneous adaptation dispersion value is less than the structural deviation control threshold, the current task is recorded as a structurally well-adapted task and written to the structural adaptation cache, and continues to participate in the task scheduling process of the current node.

[0014] Further, the specific steps for analyzing the resource load fluctuation characteristics during node operation are as follows: Select a data segment marked with a task cycle from the historical resource request records, read the resource type and request intensity fields, and obtain the historical resource request value of the i-th type; combine the time index in the node's historical resource snapshot to extract the resource occupancy record corresponding to the task cycle and obtain the available value of the i-th type of resource; locate the record interval of the s-th period in the node's historical load rate sequence, extract the resource utilization rate at the corresponding time point, and obtain the node load rate of the s-th period; obtain the total number of resource categories, the number of sliding cycles, and the continuous cycle sampling interval time; analyze the node operation... Resource load fluctuation characteristics during the process: Subtract the available value of the historical resource of the i-th type from the historical resource request value of the i-th type, and divide by the historical resource request value of the i-th type to obtain the resource adaptation deviation ratio; Multiply the resource adaptation deviation ratio by the derivative of the node load rate of the s-th period with respect to the continuous period sampling interval to obtain the deviation product term of the i-th type of resource in the s-th period; Sum the deviation product terms corresponding to the total number of resource categories to obtain the cumulative resource deviation value of the s-th period; Sum the cumulative resource deviation values ​​of all periods covered by continuously backtracking k periods from the current period, and divide by the number of sliding window periods to obtain the weighted deviation value of resource fluctuation.

[0015] Further, the specific steps for classifying scheduling stability levels and adjusting task scheduling priority strategies based on the judgment results are as follows: The resource fluctuation weighted deviation value is compared in real time with the fluctuation deviation threshold, which includes a primary fluctuation deviation threshold and a secondary fluctuation deviation threshold. When the resource fluctuation weighted deviation value is greater than or equal to the primary fluctuation deviation threshold, the current task is marked as a high-risk task for resource conflict, the scheduling priority and resource allocation process is suspended, and the value is written to the node resource optimization record area. When the resource fluctuation weighted deviation value is greater than the secondary fluctuation deviation threshold but less than the primary fluctuation deviation threshold, the current task is marked as a scheduling unstable task, the scheduling priority is reduced, and the value is written to the scheduling task review queue. When the resource fluctuation weighted deviation value is less than or equal to the secondary fluctuation deviation threshold, the current task is marked as a resource scheduling stable task, written to the stable state record area, and included in the priority scheduling pool.

[0016] Further, the specific steps for performing multi-dimensional matching and optimal path determination by combining scheduling stability level and node scheduling resource status are as follows: Read the number of tasks to be executed by the bound node within the current period in the task queue length to obtain the current queued task number of the node; according to the time window defined in the historical evaluation parameters of the queue status, select the corresponding interval of the queue record from the task queue length sequence, calculate the arithmetic mean, and obtain the historical average queued task number of the node; obtain the resource fluctuation weighted deviation value and the heterogeneous adaptation dispersion value; select stable resource scheduling tasks from the stable state record area to perform multi-dimensional matching and optimal path determination; multiply the resource fluctuation weighted deviation value and the heterogeneous adaptation dispersion value to obtain the joint adaptation fluctuation product value; divide the current queued task number of the node by the historical average queued task number of the node to obtain the queue pressure ratio; multiply the joint adaptation fluctuation product value by the queue pressure ratio to obtain the multi-factor comprehensive product term; add one to the multi-factor comprehensive product term and take the natural logarithm to obtain the task migration determination value.

[0017] Further, the specific steps for selecting target nodes and generating node resource binding structures and path channel information based on the path optimization strategy are as follows: The task migration judgment value is compared with migration thresholds T1, T2, and T3, and migration control levels are classified according to the comparison results; when the task migration judgment value is greater than or equal to T3, it is classified as a path reconstruction level; when the task migration judgment value is greater than or equal to T2 and less than T3, it is classified as a delayed switching level; when the task migration judgment value is greater than or equal to T1 and less than T2, it is classified as a mild congestion level; when the task migration judgment value is less than T1, it is classified as a strategy maintenance level; when the task migration judgment value is at the path reconstruction level, the migration process begins. The migration path planning process involves filtering the target node candidate pool, invoking the resource adaptation scoring mechanism to complete node reselection, writing the task to the target node task initialization queue, and releasing the current node resources. When the task migration judgment value is at the delayed switching level, it is marked as a delayed migration task, a node switching suggestion list is generated, and it is written to the scheduling candidate cache. When the task migration judgment value is at the mild congestion level, the task adaptation status and the current node identifier are recorded and written to the node load analysis unit for subsequent resource scheduling status evaluation. When the task migration judgment value is at the policy maintenance level, the corresponding task record in the scheduling candidate cache is cleared, and the binding information in the target node candidate path list is deleted.

[0018] Further, the specific steps for executing the task container transfer and context recovery process based on path channel information and node resource binding structure, and feeding back the migration execution status and resource release information to the scheduling mapping base dataset are as follows: Extract the task migration judgment value at the path reconstruction level, and perform a hot migration feasibility judgment operation, including: checking whether the running status meets the context encapsulation conditions, confirming that no uninterruptible resources are bound and that the system is not in a scheduling locked state; comparing the target node container environment compatibility, and verifying whether the migration path channel supports uninterrupted transmission; reading the task meta-information and migration mode definition in the configuration strategy, and generating a hot migration feasibility marker; if the hot migration conditions are met, a hot migration data channel is established based on the path channel information, and the running context is dynamically transferred and the target node instantiation is restored; if the conditions are not met, the cold migration process is entered and the running status is encapsulated; regardless of whether it is a hot migration or a cold migration, container packaging, context transmission, and target node status restoration are completed, source node resources are released, migration time, startup status, and resource adaptation structure are recorded, and the migration results, target node running performance, and resource usage change information are fed back to the scheduling mapping base dataset to perform scheduling parameter updates and resource adaptation strategy adjustments.

[0019] The second aspect of this invention provides a heterogeneous computing power pool task adaptive migration and dynamic scheduling system, comprising: a computing power mapping acquisition module, a task adaptation judgment module, a resource matching scheduling module, and a migration execution feedback module. The computing power mapping acquisition module is used to collect task resource request information, computing power node running status information, and scheduling environment status information, construct a scheduling mapping basic dataset, and preprocess the scheduling mapping basic dataset. The task adaptation judgment module is used to analyze the adaptation structure differences between task resource requests and node resource states based on the scheduling mapping basic dataset, analyze the resource load fluctuation characteristics during node operation, classify scheduling stability levels according to the judgment results, and adjust the task scheduling priority strategy. The resource matching scheduling module is used to perform multi-dimensional matching and optimal path judgment by combining the scheduling stability level and node scheduling resource state, filter target nodes according to the path optimization strategy, and generate node resource binding structure and path channel information. The migration execution feedback module is used to perform task container transfer and context recovery processes according to the path channel information and node resource binding structure, and feed back the migration execution status and resource release information to the scheduling mapping basic dataset.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) By constructing a scheduling mapping basic dataset and performing preprocessing operations, this invention can realize a unified structural expression of multi-source scheduling information in the time and resource dimensions, so that the system has stable resource status parsing capabilities and highly consistent task scheduling input when facing heterogeneous node resource differences and task cycle offsets.

[0023] (2) Based on the joint analysis mechanism of heterogeneous adaptation dispersion value and resource fluctuation weighted deviation value, this invention can dynamically identify the imbalance of resource adaptation structure and the characteristics of node load fluctuation, thereby improving the system's response sensitivity and judgment accuracy to changes in multidimensional adaptation relationship and operating trend during resource allocation.

[0024] (3) By constructing a task migration judgment value and introducing a multi-level migration control hierarchy, this invention enables the scheduling strategy to make a comprehensive judgment on the task scheduling status, node resource pressure and queuing congestion, so that the system can maintain a dynamic balance between task execution and resource utilization during the scheduling path selection process.

[0025] (4) This invention uses the path channel information and node resource binding structure to trigger the migration execution feedback mechanism, which can realize the state maintenance and execution recovery of the task context during the migration process, and use the migration results to update the scheduling parameters, thereby enhancing the continuity and iterative optimization capability of the scheduling control process.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a flowchart of the heterogeneous computing power pool task adaptive migration and dynamic scheduling method of the present invention;

[0028] Figure 2 This is a structural diagram of the heterogeneous computing power pool task adaptive migration and dynamic scheduling system of the present invention;

[0029] Figure 3 This is a trend chart of the task migration determination value of the present invention;

[0030] Figure 4 This is a diagram of the computing node network topology of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-4 This invention provides a technical solution: a method for adaptive migration and dynamic scheduling of tasks in a heterogeneous computing power pool, comprising the following steps: S1, collecting task resource request information, computing node running status information, and scheduling environment status information, constructing a scheduling mapping basic dataset, and preprocessing the scheduling mapping basic dataset; S2, based on the scheduling mapping basic dataset, analyzing the adaptation structure differences between task resource requests and node resource status, analyzing the resource load fluctuation characteristics during node operation, classifying scheduling stability levels according to the judgment results, and adjusting the task scheduling priority strategy; S3, combining the scheduling stability level and node scheduling resource status to perform multi-dimensional matching and optimal path judgment, filtering target nodes according to the path optimization strategy, and generating node resource binding structure and path channel information; S4, performing task container transfer and context recovery process according to the path channel information and node resource binding structure, and feeding back the migration execution status and resource release information to the scheduling mapping basic dataset.

[0033] Specifically, the steps for collecting task resource request information, computing node running status information, and scheduling environment status information to construct the scheduling mapping basic dataset are as follows: Collect multi-source scheduling information generated during the operation of the heterogeneous computing pool by the task submission interface, node resource monitoring system, task running log unit, queue scheduling management system, and scheduler global control component. This multi-source scheduling information includes: a resource dimension request structure generated based on the task submission structure, used to describe the static request parameters of various tasks in multiple resource dimensions such as model type, concurrency, memory requirements, bandwidth requirements, computation cycle, and IO read / write rate; a real-time available resource snapshot of nodes generated by the node monitoring system, used to reflect the resource remaining capacity of each node at the sampling time, including the number of CPU cores, GPU cores and remaining memory, free memory space, instantaneous power consumption level of the node, network interface bandwidth, and connection status; a resource category structure definition extracted through the preset resource classification definition in the system configuration file, used to clarify the resource types, dimension order, and allocation restrictions currently available for scheduling in the system; a task queue length generated based on the task entry order and waiting status recorded in the scheduling queue, used to quantify the current scheduling pressure; and a node history database composed of historical resource request records formed by combining historical task submission records and node status snapshots. Historical resource snapshots are used to support time-series analysis of resource request behavior and node carrying capacity; historical node load rate sequences calculated by calling the node load monitoring module are used to identify the fluctuation trend of node operating pressure over continuous periods; the number of sliding window periods and the continuous period sampling interval determined by the sliding window parameters set in the system operation log are used to set the time span and frequency in subsequent dynamic feature extraction; all collected information is accompanied by a timestamp in a uniform format to ensure the synchronization and comparability of cross-module data; at the same time, historical evaluation parameters of queuing status in the scheduling system configuration file are called, including the average queuing waiting time and resource... The core fields of allocation failure rate, scheduling retry count, and wait-to-execute ratio are used to enhance the interpretability of scheduling behavior in the time dimension. All collected multi-source scheduling information is uniformly aligned in time according to the task execution cycle and node status cycle. A smoothing method based on exponential weighted moving average is used to fill in the time periods with sampling gaps. A mapping index is established based on node ID and task ID. Resource requests, node status, scheduling behavior, and timing control information are organized in combination with the node resource scheduling path relationship recorded in the scheduler. Finally, a basic dataset for scheduling mapping is constructed, which serves as the sole input source for subsequent adaptation judgment, resource matching, and path planning modules.

[0034] In this implementation plan, the construction of a scheduling mapping basic dataset can achieve the unified integration of task resource request structure, node running status characteristics and scheduling environment time sequence information, open up the logical relationship path between multi-source data, and provide a scheduling foundation support with clear structure, complete time sequence and node task mapping relationship for subsequent modules, significantly improving the adaptability of scheduling strategy and the accuracy of resource allocation strategy in heterogeneous computing power pool.

[0035] Specifically, the preprocessing steps for the scheduling mapping dataset are as follows: Addressing the heterogeneous structure and periodic offset characteristics of the scheduling mapping dataset across time, node, and resource dimensions, a unified cross-dimensional alignment mechanism is constructed. Abnormal offset values ​​in scheduling behavior are identified using the median absolute deviation method of the sliding interval. Combined with the K-Medoids clustering algorithm based on periodic segment partitioning, cluster analysis is performed on the scheduling path differences generated by different nodes in asynchronous states. Abnormal switching segments and format drift segments are identified and removed, unifying periodic boundaries and node state structures, reconstructing the temporal continuity representation of scheduling paths, and improving the data's consistency across periods and multiple nodes. Comparability: Utilizing the mapping relationship between node periodic behavior density and resource call trajectory, an improved dynamic time warping algorithm is used to perform time-series matching of task execution trajectories under different nodes. Trajectory compression rules are used to warp nonlinear segments of resource calls, eliminating redundant segments in path records caused by task fluctuations or node switching, ensuring that scheduling status data blocks possess complete closed-loop path representation capabilities. The scheduling mapping base dataset undergoes mean standardization and maximum / minimum normalization to ensure that data across resource dimensions have uniform scale and distribution characteristics, avoiding bias caused by numerical magnitude differences in subsequent model analysis, and ensuring the comparability and stability of cross-module algorithm processing results.

[0036] In this implementation scheme, the preprocessing operation improves the structural consistency and temporal integrity of the scheduling mapping base dataset across cycles, nodes, and multiple resource dimensions, effectively avoids interference from abnormal segments and trajectory drift, ensures that the resource scheduling path is expressed continuously and clearly, constructs a standardized input foundation that adapts to dynamic scheduling criteria, and enhances the robustness and data universality of subsequent computational structures.

[0037] Specifically, the steps for analyzing the adaptation structure differences between task resource requests and node resource states based on the scheduling mapping basic dataset are as follows: Based on the current period's resource dimension request structure, read the call fields of all resource types, extract the request values ​​of the i-th and j-th resource types, and clarify the static demand of the current task in different resource dimensions; synchronously read all resource field contents of the current node from the node's real-time available resource snapshot, extract the available values ​​of the i-th and j-th resource types as a reference for the node's current carrying capacity; parse the resource category structure definition, extract the number of valid resource dimension identifiers, obtain the total number of resource categories, and ensure that all resource dimensions have a unified weight basis in subsequent calculations; divide the i-th resource request value by the i-th resource available value to obtain the i-th resource adaptation ratio, representing the instantaneous adaptation degree of a single dimension; divide the j-th resource request value by the j-th resource available value, sum the results after traversing all resource categories, divide by the total number of resource categories, and obtain the average resource adaptation ratio, constituting the system's overall... A reference benchmark for resource adaptation level is established. The resource adaptation ratio of the i-th type is subtracted from the mean resource adaptation ratio, and the absolute value of the difference is taken to obtain the resource adaptation deviation term of the i-th type, used to measure the disturbance magnitude of the single-dimensional adaptation value on the overall adaptation balance. All resource adaptation deviation terms are summed and divided by the total number of resource categories to obtain the heterogeneous adaptation dispersion value, which serves as a quantitative indicator of multi-dimensional resource structure balance. The heterogeneous adaptation dispersion value is compared with the structural deviation control threshold in real time to determine whether there are structural adaptation differences between tasks and nodes. When the heterogeneous adaptation dispersion value is greater than or equal to the structural deviation control threshold, the current task is marked as a structurally unbalanced task, the target node candidate list is reconstructed, and structurally mismatched node paths are removed to avoid local resource bottlenecks after scheduling. When the heterogeneous adaptation dispersion value is less than the structural deviation control threshold, the current task is recorded as a structurally well-adapted task and written to the structural adaptation cache, continuing to participate in the task scheduling process of the current node, improving the matching degree of resource allocation and the stability of the scheduling path.

[0038] The specific calculation method for the heterogeneous adaptation dispersion value is as follows:

[0039]

[0040] In the formula, This represents the heterogeneous adaptation dispersion value. This represents the request value for the i-th type of resource. This represents the available value of the i-th type of resource. This represents the request value for the j-th type of resource. This represents the available value of the j-th type of resource. This indicates the total number of resource categories.

[0041] In this implementation scheme, the heterogeneous adaptation dispersion value can quantify the structural deviation between multidimensional resources, clearly reveal the adaptation balance between task and node resource distribution, and realize the early identification of potential bottleneck paths and automatic elimination of mismatched nodes through real-time comparison, effectively improving the structural matching quality of resource scheduling and the accuracy of scheduling path selection.

[0042] Specifically, the steps for analyzing the resource load fluctuation characteristics during node operation are as follows: Select data segments marked with task cycles from historical resource request records, read the resource type and request intensity fields, and obtain the historical resource request value of type i, which serves as the theoretical resource demand for the task in the past cycle; Combine the time index in the node's historical resource snapshots to extract resource occupancy records corresponding to the task cycle, and obtain the available value of type i resources, reflecting the actual amount of such resources available to the node during the same period; Locate the record interval of period s in the node's historical load rate sequence, extract the resource utilization rate at the corresponding time point, and obtain the node load rate of period s, used to capture the node's resource operation pressure status; Obtain the total number of resource categories, the number of sliding cycles, and the continuous cycle sampling interval to construct a complete time window and resource dimension range; Analyze the resource load fluctuation characteristics during node operation: [The text then abruptly shifts to a different topic:] ...the historical resource... The resource adaptation deviation ratio is obtained by subtracting the available value of the i-th type of historical resource from the requested value and dividing it by the requested value of the i-th type of historical resource. This ratio represents the relative strength of the difference between demand and available resources. The resource adaptation deviation ratio is multiplied by the derivative of the node load rate in the s-th period with respect to the continuous period sampling interval to form a deviation with time-varying weights. This deviation product term for the i-th type of resource in the s-th period is obtained, which characterizes the synchronous change of resource tension status with node load trends in a short period. The deviation product term for each type of resource is summed to obtain the cumulative resource deviation value in the s-th period, which serves as the joint expression of resource adaptation and load fluctuation in this period. The cumulative resource deviation values ​​for all periods covered by continuously backtracking k periods from the current period are summed and divided by the number of sliding window periods to obtain the resource fluctuation weighted deviation value, which is used to quantify the resource scheduling stability and trend anomaly strength of nodes in the recent time window.

[0043] The specific calculation method for the weighted deviation value of resource fluctuations is as follows:

[0044]

[0045] In the formula, This represents the weighted deviation value of resource fluctuations. This represents the request value for the i-th type of historical resource. This represents the available value of the i-th type of historical resource. This represents the node load rate in period s. Indicates the number of sliding window periods. Indicates the continuous periodic sampling interval time. This indicates the total number of resource categories.

[0046] In this implementation scheme, the resource fluctuation weighted deviation value characterizes the dynamic linkage between the node resource adaptation status and the load change trend. It can accurately reflect the fluctuation intensity of the scheduling environment within a continuous period, capture short-term overload or resource deterioration trends, provide time-weighted feature support for scheduling decisions, and enhance the scheduling strategy's responsiveness to the evolution of node operating status.

[0047] Specifically, the steps for classifying scheduling stability levels and adjusting task scheduling priority strategies based on the judgment results are as follows: The weighted deviation value of resource fluctuations is compared in real time with the fluctuation deviation threshold. The fluctuation deviation threshold includes a primary fluctuation deviation threshold and a secondary fluctuation deviation threshold, used to define high-risk fluctuation ranges and slightly unstable ranges, respectively. When the weighted deviation value of resource fluctuations is greater than or equal to the primary fluctuation deviation threshold, the current task is marked as a high-risk task for resource conflict, the scheduling priority and resource allocation process is suspended, and the data is written to the node resource optimization record area to drive subsequent resource mirroring structure adjustments and node load transfer operations. When the weighted deviation value of resource fluctuations is greater than the secondary fluctuation deviation threshold but less than the primary fluctuation deviation threshold, the current task is marked as a high-risk task for resource conflict. Tasks that are unstable in scheduling are downgraded in priority and added to the task review queue. A periodic re-evaluation mechanism tracks the continuity and direction of fluctuation trends. When the weighted deviation of resource fluctuations is less than or equal to the second-level fluctuation deviation threshold, the current task is marked as a stable resource scheduling task, written to the stable state record area, and included in the priority scheduling pool as a high-availability task for the scheduler to select first. A lag and jitter suppression mechanism is introduced, setting a judgment lag time and state maintenance period for the critical range of the weighted deviation of resource fluctuations. A state level change is triggered only when the judgment results of three consecutive periods are consistent, avoiding frequent upgrades and downgrades caused by instantaneous fluctuations, stabilizing the response rhythm of the scheduling strategy, and improving the overall scheduling robustness of the system.

[0048] In this implementation plan, the scheduling stability level classification and priority strategy adjustment mechanism can accurately identify resource conflicts and scheduling instability. Combined with the lag and jitter suppression mechanism, it can effectively control the frequency of task state switching, maintain the continuity and stability of task level determination, avoid repeated fluctuations in scheduling paths, and improve the rhythm coordination of resource allocation and the stable carrying capacity of the scheduling system.

[0049] Specifically, the steps for multi-dimensional matching and optimal path determination, combining scheduling stability level and node scheduling resource status, are as follows: First, read the number of pending tasks bound to the node within the current period from the task queue length to obtain the current queued task count, used to measure the current congestion level of scheduling resources. Second, based on the time window defined in the historical evaluation parameters of the queuing status, select the corresponding interval of queuing records from the task queue length sequence, calculate the arithmetic mean, and obtain the historical average queued task count of the node, serving as a time reference benchmark for congestion status. Third, obtain the resource fluctuation weighted deviation value and heterogeneous adaptation dispersion value to express the degree of matching between node load status and resources. Fourth, select stable resource scheduling tasks from the stable state record area to perform multi-dimensional matching and optimal path determination, ensuring that candidate tasks have continuous and stable scheduling performance. Fifth, combine the resource fluctuation weighted deviation value with the heterogeneous adaptation dispersion value. Multiplying the adaptation dispersion values ​​yields the joint adaptation fluctuation product, which constitutes the adaptation-stability coupling quantity of scheduling pressure. Dividing the current number of queued tasks by the node's historical average number of queued tasks yields the queuing pressure ratio, reflecting the node's relative load level. Multiplying the joint adaptation fluctuation product by the queuing pressure ratio yields a multi-factor comprehensive product term, which introduces a joint expression of node structure matching, operational stability, and scheduling congestion when constructing the task migration decision value. Adding one to the multi-factor comprehensive product term and taking its natural logarithm yields the task migration decision value. Introducing the logarithmic function can suppress the threshold judgment sensitivity caused by the sharp rise of the product term in extreme congestion scenarios, avoiding frequent triggering of migration judgments in the boundary critical region, making the overall migration strategy response more buffered and tolerant, and improving the stability of migration decisions and the convergence smoothness of the decision curve.

[0050] The specific calculation method for the task migration decision value is as follows:

[0051]

[0052] In the formula, This represents the task migration decision value. This represents the weighted deviation value of resource fluctuations. This represents the heterogeneous adaptation dispersion value. This indicates the number of tasks currently queued at the node. This represents the historical average number of queued tasks for a node.

[0053] Table 1 shows the task migration determination value data table provided in the embodiments of this application. The resource fluctuation weighted deviation value for determination 1 is set to 0.80, the heterogeneous adaptation dispersion value is set to 1.20, the current queued task count is set to 6.0, and the historical average queued task count is set to 5.0; the resource fluctuation weighted deviation value for determination 2 is set to 0.65, the heterogeneous adaptation dispersion value is set to 0.95, the current queued task count is set to 5.0, and the historical average queued task count is set to 6.0; the resource fluctuation weighted deviation value for determination 3 is set to 0.95, and the heterogeneous adaptation dispersion value is set to 1. .00, the current number of queued tasks is set to 6.0, and the historical average number of queued tasks is set to 6.0; the resource fluctuation weighted deviation value for judgment 4 is set to 0.50, the heterogeneous adaptation dispersion value is set to 0.80, the current number of queued tasks is set to 4.0, and the historical average number of queued tasks is set to 6.0; the resource fluctuation weighted deviation value for judgment 5 is set to 1.10, the heterogeneous adaptation dispersion value is set to 1.30, the current number of queued tasks is set to 7.0, and the historical average number of queued tasks is set to 5.0.

[0054] Table 1 Task Migration Judgment Value Data Table

[0055]

[0056] like Figure 3 The figure shows a trend chart of task migration judgment values ​​provided in this application embodiment. According to the data in the image and table, the threshold for Level 3 migration judgment is 1.00, the threshold for Level 2 migration judgment is 0.40, and the threshold for Level 1 migration judgment is 0.60. The task migration judgment values ​​corresponding to the five sets of data fluctuate between 0.24 and 1.10, exhibiting a multi-peak oscillation characteristic. Judgment 5 has a task migration judgment value of 1.10, exceeding the Level 3 migration judgment threshold, indicating that the current task resource adaptation structure is significantly abnormal and the queuing pressure is significant, placing it in a high-priority migration state requiring immediate path reconstruction. Judgments 1 and 3 are 0.77 and 0.67 respectively, falling between Level 1 and Level 3, indicating a certain degree of instability in the task's running state, suggesting it enter the node candidate pool for switching. Judgment 2 has a value of 0.42, falling between Level 1 and Level 2, belonging to the slightly abnormal scheduling range. Judgment 4 has a value of 0.24, below the Level 2 migration judgment threshold, indicating the task is in a stable scheduling state. This figure can be used to intuitively determine whether the task scheduling state meets the migration trigger conditions, providing a basis for subsequent path selection and node switching operations.

[0057] In this implementation plan, the task migration judgment value integrates resource matching status, operational stability, and scheduling congestion level to construct a multi-dimensional, cross-cycle joint judgment basis. Logarithmic compression is used to suppress the amplification effect of extreme values, enhance the judgment stability in high load ranges, avoid frequent boundary jumps, and improve the smoothness and fault tolerance robustness of the migration judgment process.

[0058] Specifically, the steps for selecting target nodes and generating node resource binding structures and path channel information based on the path optimization strategy are as follows: The task migration judgment value is compared with migration thresholds T1, T2, and T3, and migration control levels are classified according to the comparison results. When the task migration judgment value is greater than or equal to T3, it is classified as a path reconstruction level, indicating that the current running state of the task has seriously deviated from the resource adaptation balance range and scheduling stability range, and node switching and path reconstruction operations need to be performed immediately. When the task migration judgment value is greater than or equal to T2 and less than T3, it is classified as a delayed switching level, used to identify intermediate tasks in the transition phase, and postpone migration actions to observe the state trend. When the task migration judgment value is greater than or equal to T1 and less than T2, it is classified as a mild congestion level, suitable for tasks with local resource shortages but overall scheduling still tolerable. When the task migration judgment value is less than T1, it is classified as a strategy maintenance level, maintaining the original node resource binding state and stabilizing the current scheduling structure. When the task migration judgment value is at the path reconstruction level, the migration path planning process is entered, the target node candidate pool is filtered, and the resource adaptation scoring mechanism is invoked to complete node reselection. The resource adaptation scoring mechanism is based on a multi-objective scoring strategy, comprehensively considering resource matching degree, affinity and anti-affinity policy requirements, cross-availability zone scheduling cost, average communication latency in the node network topology, historical load jitter frequency and fault recovery time. A hierarchical weighted method is used to construct a comprehensive scoring structure to ensure that the selected nodes are adapted in all dimensions in terms of computing power, network location and policy compatibility. Based on the scoring results, a target node priority list is generated, the node with the highest score is selected as the target node, the task is written to the target node's task initialization queue and the current node's resources are released, completing the resource unbinding and path reconstruction operation. When the task migration judgment value is at the delayed switching level, it is marked as a delayed migration task, a node switching suggestion list is generated and written to the scheduling candidate cache for subsequent periodic review and invocation; when the task migration judgment value is at the mild congestion level, the task adaptation status and current node identifier are recorded and written to the node load analysis unit for the system to refer to during resource reallocation or batch optimization; when the task migration judgment value is at the policy maintenance level, the corresponding task record in the scheduling candidate cache is cleared and the binding information in the target node candidate path list is deleted, maintaining the original path and binding status unchanged to avoid invalid migration actions.

[0059] In this implementation scheme, the path optimization mechanism triggered by the task migration judgment value can realize the fine division of migration level and the dynamic screening of node candidate paths. Combined with the resource adaptation scoring mechanism, it evaluates multi-dimensional matching and scheduling cost, and ensures the global adaptability of target node selection while controlling the pace of migration actions, thereby improving the convergence efficiency of the task scheduling process and the utilization rate of node resources.

[0060] Specifically, the task container transfer and context recovery process is executed based on the path channel information and node resource binding structure, and the migration execution status and resource release information are fed back to the scheduling mapping base dataset. The specific steps are as follows: Extract the task migration judgment value at the path reconstruction level, and perform a hot migration feasibility judgment operation, including: checking whether the running status meets the context encapsulation conditions, judging whether there are service components or non-shared data mounting paths that need to rely on local persistent state. If the task does not involve local state usage and has a remote state synchronization mechanism, then the context encapsulation conditions are met; confirm that no uninterruptible resources are bound, including checking whether there are pass-through devices occupying, in the configuration locking stage, or in exclusive resource scheduling; compare the compatibility of the target node container environment, and judge the compatibility between the currently running image and the target node's supported image by image hash matching and dependency version consistency. If the image compatibility score is higher than the environment adaptation benchmark threshold dynamically generated by the system, and the dependency environment consistency evaluation meets the standard, then it is considered that the container environment is compatible, and the compatibility score threshold is passed. The baseline mirror group is compared and automatically fitted with historical migration success rate curves, with the adaptation range dynamically updated. The migration path channel is verified to support uninterrupted transmission. Path bandwidth and packet loss rate metrics are obtained in real-time through the link status monitoring interface. The path channel guarantee level in the scheduling network is dynamically matched according to the task migration level. Path bandwidth and packet loss rate must meet the transmission continuity requirements under the corresponding level to ensure the complete continuity of the container context during the switching process. Furthermore, based on the congestion level label and bandwidth usage curve of the transmission path, it is analyzed whether it has continuous and stable data carrying capacity. Combined with flow control policy configuration, it is determined whether the current path enables rate adjustment, congestion avoidance, and retransmission mechanisms to ensure uninterrupted transmission capability under high concurrency and sudden loads. Bandwidth and packet loss rate thresholds are derived through the scheduling system's adaptive channel performance evaluation model. These thresholds are dynamically adjusted based on path performance monitoring results, inter-node link usage statistics, historical migration interruption ratios, and retry success rates to ensure that the thresholds are set near the steady-state performance boundary, avoiding oversensitivity and judgment failure.Read the migration mode definition fields from the task metadata and configuration strategy to generate a hot migration feasibility marker. If the hot migration conditions are met, establish a hot migration data channel based on the path channel information, perform dynamic transfer of the runtime context and instantiation recovery of the target node, and complete the uninterrupted task takeover process. If the conditions are not met, enter the cold migration process, encapsulate the current runtime state, and rebuild the container environment and task context of the target node. Regardless of whether it is a hot or cold migration, container packaging, context transmission, and target node state recovery operations must be completed, while releasing the source node resources. Record key indicators of the migration process, including migration time, recovery time, state encapsulation time, task startup delay, resource reallocation time, service level protocol coverage change, resource utilization during the target node initialization phase, and task first available response time. All indicators are written to a data field consistent with the structure of the S1 phase. Feed back the migration results, target node runtime performance, and resource usage change information to the scheduling mapping base dataset to complete the scheduling parameter update and resource adaptation strategy adjustment, supporting the subsequent node load model construction and dynamic optimization of path selection strategy.

[0061] In this implementation plan, the task container transfer and context recovery process accurately identifies hot migration conditions through multiple feasibility criteria, dynamically selects hot migration and cold migration paths, and combines container environment, scheduling channel and resource status linkage control to ensure smooth task continuation and complete resource release. All indicators during the migration process are synchronously written into the scheduling mapping basic dataset, providing data closed-loop support for subsequent scheduling parameter correction and strategy optimization.

[0062] like Figure 2The diagram shown is a structural schematic of the heterogeneous computing power pool task adaptive migration and dynamic scheduling system provided in this application embodiment. The heterogeneous computing power pool task adaptive migration and dynamic scheduling system provided in this application embodiment applies a heterogeneous computing power pool task adaptive migration and dynamic scheduling method, including: a computing power mapping acquisition module, a task adaptation judgment module, a resource matching scheduling module, and a migration execution feedback module. The computing power mapping acquisition module is used to collect task resource request information, computing power node running status information, and scheduling environment status information, construct a scheduling mapping basic dataset, and preprocess the scheduling mapping basic dataset to achieve structural alignment and time synchronization of resource status, improving the parsability and cross-cycle applicability of scheduling data. The task adaptation judgment module is used to parse task resource requests and node resources based on the scheduling mapping basic dataset. The system adapts structural differences between states and analyzes resource load fluctuation characteristics during node operation. Based on the judgment results, it classifies scheduling stability levels and adjusts task scheduling priority strategies to achieve resource bottleneck identification and dynamic correction of scheduling strategies. The resource matching scheduling module combines scheduling stability levels and node scheduling resource states to perform multi-dimensional matching and optimal path judgment. Based on the path optimization strategy, it selects target nodes and generates node resource binding structures and path channel information to achieve adaptive matching of scheduling paths and robust control of target node selection. The migration execution feedback module performs task container transfer and context recovery processes based on path channel information and node resource binding structures, and feeds back migration execution status and resource release information to the scheduling mapping base dataset to achieve closed-loop feedback of migration results and dynamic optimization of system scheduling parameters.

[0063] In this implementation plan, the modular structure constructs a complete scheduling and control path, forming a closed-loop control chain from resource status acquisition, adaptation evaluation, path matching to container migration. This enables dynamic perception of task running status, precise adjustment of scheduling response, and controllable closed-loop migration execution, supporting the stable operation of tasks in heterogeneous computing power pools and the efficient implementation of cross-node resource scheduling.

[0064] like Figure 4 The diagram shows the network topology of the computing nodes provided in this embodiment. It illustrates the connections between five computing nodes, forming a directed interconnection structure with redundant paths. Node A has direct communication channels with both nodes B and C. Node B is further connected to nodes D and E. Node C is connected to node D, and node D is also interconnected with node E. This structure provides multi-path transmission capability, enabling the scheduling system to select the path with the least latency and optimal load among nodes for task migration, thus improving the overall resource scheduling flexibility and fault tolerance.

[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0066] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for adaptive migration and dynamic scheduling of tasks in heterogeneous computing power pools, characterized in that, Includes the following steps: S1: Collect task resource request information, computing node running status information and scheduling environment status information, construct the scheduling mapping basic dataset, and preprocess the scheduling mapping basic dataset; S2, based on the scheduling mapping basic dataset, analyzes the adaptation structure differences between task resource requests and node resource status, analyzes the resource load fluctuation characteristics during node operation, classifies scheduling stability levels according to the judgment results, and adjusts task scheduling priority strategies. S3 combines scheduling stability level and node scheduling resource status to perform multi-dimensional matching and optimal path judgment, and filters target nodes and generates node resource binding structure and path channel information according to path optimization strategy. S4 executes the task container transfer and context recovery process based on the path channel information and node resource binding structure, and feeds back the migration execution status and resource release information to the scheduling mapping base dataset.

2. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for collecting task resource request information, computing node running status information, and scheduling environment status information to construct the scheduling mapping basic dataset are as follows: During the operation of the heterogeneous computing pool, multi-source scheduling information generated by the task submission interface, node resource monitoring system, task execution log unit, queue scheduling management system, and scheduler global control component is collected. This information includes: resource dimension request structure, real-time available resource snapshots of nodes, resource category structure definitions, task queue length, historical resource request records, historical resource snapshots of nodes, historical load rate sequences of nodes, sliding window period count, and continuous period sampling interval time. All information is accompanied by a timestamp. At the same time, the historical evaluation parameters of queue status in the scheduling system configuration file are called. The multi-source scheduling information is time-aligned according to the task execution cycle and node status cycle. Resource requests, node status, scheduling behavior, and timing control information are organized in combination with the node resource scheduling path relationship to construct a scheduling mapping basic dataset.

3. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for preprocessing the scheduling mapping base dataset are as follows: To address the heterogeneous structure and periodic offset characteristics of the scheduling mapping dataset across time, node, and resource dimensions, a unified cross-dimensional alignment mechanism is constructed. This mechanism uses the median absolute deviation of the sliding interval combined with a clustering algorithm based on periodic segment partitioning to identify and remove abnormal switching segments and format drift segments, unifying periodic boundaries and node state structures, and reconstructing the temporal continuity of scheduling paths. Furthermore, leveraging the mapping relationship between node periodic behavior density and resource call trajectories, the structural representation of scheduling state data blocks is reconstructed based on dynamic time series alignment. Nonlinear segments are regularized and redundant path information is removed using trajectory compression rules. Finally, the entire scheduling mapping dataset undergoes mean standardization and maximum / minimum normalization.

4. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for parsing the adaptation structure differences between task resource requests and node resource states based on the scheduling mapping base dataset are as follows: Based on the current periodic resource dimension request structure, read the call fields of all resource types and extract the request values ​​of the i-th type of resource and the j-th type of resource; Synchronously read all resource field contents of the current node from the real-time available resource snapshot of the node, extract the available value of the i-th type of resource and the available value of the j-th type of resource; parse the resource category structure definition, extract the number of valid resource dimension identifiers, and obtain the total number of resource categories; Divide the request value of the i-th type of resource by the available value of the i-th type of resource to obtain the adaptation ratio value of the i-th type of resource; Divide the requested value of resource type j by the available value of resource type j, sum the values ​​after traversing all resource types, and divide by the total number of resource types to obtain the average resource adaptation ratio; subtract the average resource adaptation ratio from the resource adaptation ratio of type i, and take the absolute value of the difference to obtain the resource adaptation deviation of type i. Sum all resource adaptation deviations and divide by the total number of resource categories to obtain the heterogeneous adaptation dispersion value; The heterogeneous adaptation dispersion value is compared with the structural deviation control threshold in real time. When the heterogeneous adaptation dispersion value is greater than or equal to the structural deviation control threshold, the current task is marked as a structural adaptation imbalance task, the target node candidate list is reconstructed and the node paths with structural mismatch are removed. When the heterogeneous adaptation dispersion value is less than the structural deviation control threshold, the current task is recorded as a well-adapted task and written into the structural adaptation cache, and continues to participate in the task scheduling process of the current node.

5. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for analyzing the resource load fluctuation characteristics during the operation of the analysis node are as follows: Select the data segment marked with the task cycle from the historical resource request record, read the resource type and request intensity fields, and obtain the historical resource request value of the i-th type; combine the time index in the node's historical resource snapshot to extract the resource occupancy record corresponding to the task cycle and obtain the available value of the i-th type of resource; locate the record interval of the s-th period in the node's historical load rate sequence, extract the resource utilization rate at the corresponding time point, and obtain the node load rate of the s-th period; obtain the total number of resource categories, the number of sliding cycles, and the continuous cycle sampling interval time; Analyze the resource load fluctuation characteristics during node operation: Subtract the available value of the historical resource of the i-th type from the historical resource request value of the i-th type, and divide by the historical resource request value of the i-th type to obtain the resource adaptation deviation ratio; Multiply the resource adaptation deviation ratio by the derivative of the node load rate in the s-th period with respect to the continuous period sampling interval to obtain the deviation product term of the i-th type of resource in the s-th period; Sum the deviation product terms corresponding to the total number of resource categories to obtain the cumulative resource deviation value in the s-th period; Sum the cumulative resource deviation values ​​of all periods covered by continuously backtracking k periods from the current period, and divide by the number of sliding window periods to obtain the weighted deviation value of resource fluctuation.

6. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for classifying scheduling stability levels and adjusting task scheduling priority strategies based on the determination results are as follows: The weighted deviation value of resource fluctuation is compared with the fluctuation deviation threshold in real time. The fluctuation deviation threshold includes a first-level fluctuation deviation threshold and a second-level fluctuation deviation threshold. When the weighted deviation value of resource fluctuation is greater than or equal to the first-level fluctuation deviation threshold, the current task is marked as a high-risk task for resource conflict, the scheduling priority and resource allocation process is suspended, and it is written to the node resource optimization record area; when the weighted deviation value of resource fluctuation is greater than the second-level fluctuation deviation threshold but less than the first-level fluctuation deviation threshold, the current task is marked as an unstable scheduling task, the scheduling priority is reduced, and it is written to the scheduling task review queue; when the weighted deviation value of resource fluctuation is less than or equal to the second-level fluctuation deviation threshold, the current task is marked as a stable resource scheduling task, it is written to the stable state record area, and included in the priority scheduling pool.

7. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for performing multi-dimensional matching and optimal path determination by combining scheduling stability level and node scheduling resource status are as follows: Read the number of tasks to be executed by the bound node in the current period from the task queue length to obtain the current number of queued tasks of the node; select the queue records of the corresponding interval from the task queue length sequence according to the time window defined in the queue status history evaluation parameters, calculate the arithmetic mean, and obtain the historical average number of queued tasks of the node. Obtain the weighted deviation value of resource fluctuations and the heterogeneous adaptation dispersion value; The resource scheduling stable task execution multi-dimensional matching optimal path judgment is selected from the stable state record area; the resource fluctuation weighted deviation value is multiplied by the heterogeneous adaptation dispersion value to obtain the joint adaptation fluctuation product value; Divide the current number of queued tasks at a node by the node's historical average number of queued tasks to obtain the queuing pressure ratio; multiply the joint adaptation fluctuation product value by the queuing pressure ratio to obtain the multi-factor comprehensive product term; add one to the multi-factor comprehensive product term and take the natural logarithm to obtain the task migration judgment value.

8. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for selecting target nodes and generating node resource binding structures and path channel information based on the path optimization strategy are as follows: The task migration judgment value is compared with the migration thresholds T1, T2, and T3, and the migration control level is determined based on the comparison results. When the task migration judgment value is greater than or equal to T3, it is classified as path reconstruction level; when the task migration judgment value is greater than or equal to T2 and less than T3, it is classified as delayed switching level; when the task migration judgment value is greater than or equal to T1 and less than T2, it is classified as mild congestion level. When the task migration judgment value is less than T1, it is classified as a strategy maintenance level; When the task migration judgment value is at the path reconstruction level, the migration path planning process is entered, the target node candidate pool is filtered, the resource adaptation scoring mechanism is called to complete the node reselection, the task is written to the target node task initialization queue and the current node resources are released. When the task migration judgment value is at the delayed switching level, it is marked as a delayed migration task, a node switching suggestion list is generated and written to the scheduling candidate cache area; When the task migration judgment value is at the mild congestion level, record the task adaptation status and the current node identifier, and write them into the node load analysis unit for subsequent resource scheduling status evaluation. When the task migration determination value is at the policy maintenance level, clear the corresponding task record in the scheduling candidate cache and delete the binding information in the target node candidate path list.

9. The heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to claim 1, characterized in that: The specific steps for executing the task container transfer and context recovery process based on path channel information and node resource binding structure, and feeding back the migration execution status and resource release information to the scheduling mapping base dataset are as follows: Extract the migration judgment value of the task at the path reconstruction level, and perform a hot migration feasibility judgment operation, including: checking whether the running status meets the context encapsulation conditions, confirming that it is not bound to uninterruptible resources and is not in a scheduling lock state; comparing the compatibility of the target node container environment, verifying whether the migration path channel supports uninterrupted transmission; reading the task meta information and the migration mode definition in the configuration strategy, and generating a hot migration feasibility flag. If the hot migration conditions are met, a hot migration data channel is established based on the path channel information, and dynamic transfer of the runtime context and instantiation restoration of the target node are performed. If the conditions are not met, the cold migration process is entered and the runtime state is encapsulated. Regardless of whether it is hot or cold migration, container packaging, context transfer and target node state restoration are completed, source node resources are released, migration time, startup status and resource adaptation structure are recorded, and migration results, target node runtime performance and resource usage change information are fed back to the scheduling mapping base dataset to perform scheduling parameter updates and resource adaptation strategy adjustments.

10. A heterogeneous computing power pool task adaptive migration and dynamic scheduling system, applying the heterogeneous computing power pool task adaptive migration and dynamic scheduling method according to any one of claims 1-9, comprising: The computing power mapping acquisition module, task adaptation judgment module, resource matching and scheduling module, and migration execution feedback module are characterized by: The computing power mapping acquisition module is used to collect task resource request information, computing power node running status information and scheduling environment status information, construct a scheduling mapping basic dataset, and preprocess the scheduling mapping basic dataset; The task adaptation judgment module is used to parse the adaptation structure differences between task resource requests and node resource status based on the scheduling mapping basic dataset, analyze the resource load fluctuation characteristics during node operation, classify scheduling stability levels and adjust task scheduling priority strategies according to the judgment results. The resource matching and scheduling module is used to perform multi-dimensional matching and optimal path judgment by combining the scheduling stability level and the node scheduling resource status, and to filter target nodes and generate node resource binding structure and path channel information according to the path optimization strategy. The migration execution feedback module is used to perform task container transfer and context recovery processes based on path channel information and node resource binding structure, and to feed back the migration execution status and resource release information to the scheduling mapping base dataset.

Images