A method and system for dynamic resource scheduling of AI computing power modules
By acquiring the state and topology data of the AI computing module, analyzing the task into an execution dependency graph and resource demand vector, predicting candidate resource allocation schemes, and implementing local rescheduling when there is a state deviation, the problem of insufficient resources and blocking in AI task scheduling is solved, and efficient and stable resource utilization is achieved.
Patent Information
- Application Number
- CN202611287919.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-06
AI Technical Summary
In existing technologies, a single processor or a fixed-configuration computing board cannot simultaneously address task latency, energy consumption, cost, and scalability. AI task scheduling is prone to resource shortages or ineffective migrations, and differences exist between different AI computing modules, leading to task execution blockages and resource jitter.
By acquiring the operational status data and topology status data of each AI computing module, a resource status vector is formed. The AI task is analyzed into an execution dependency graph and a resource demand vector. The latency, energy consumption, and heat margin of the candidate resource allocation scheme are predicted. Local rescheduling is implemented when status deviation occurs. Historical execution profiles and online corrections are used to improve prediction accuracy and optimize resource allocation strategies.
It improves the accuracy and stability of resource scheduling, reduces scheduling jitter, ensures the continuity of task execution and efficient use of resources, and adapts to different model and task requirements.
Smart Images
Figure CN123048141A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence computing resource management technology, and more specifically, to a dynamic resource scheduling method and system for AI computing power modules. Background Technology
[0002] As the scale of artificial intelligence models, the amount of input data, and the demand for real-time inference continue to increase, a single processor or a fixed-configuration computing board can hardly simultaneously address task latency, energy consumption, cost, and scalability. To improve computing power density, existing computing devices typically package graphics processors, neural network processors, field-programmable gate arrays, dedicated inference chips, or other AI accelerators into multiple computing power modules, and form a shared computing power platform through high-speed interconnect interfaces.
[0003] Existing resource scheduling methods typically allocate AI tasks to a specific computing module based on current utilization, idle GPU memory, or preset priorities. Some solutions can perform simple load balancing or task migration. However, AI tasks are characterized by significant differences in operator types, large model parameter sizes, frequent intermediate feature data transmission, diverse parallel processing methods, and execution latency sensitivity to temperature and power consumption. Allocating tasks solely based on instantaneous utilization can easily lead to situations where sufficient computing resources are available but local storage or interconnect bandwidth is insufficient, causing task blocking during execution.
[0004] Furthermore, different AI computing modules may differ in instruction sets, supported precision, core size, storage tiers, interconnect bandwidth, and heat dissipation capabilities. Task migration also introduces overhead for model loading, cache rebuilding, checkpoint recovery, and cross-module data transfer. If the scheduler directly performs a global reordering when the load changes, it can easily cause a large number of invalid migrations and resource jitter; if reordering is not performed, service level requirements may not be met due to hotspot temperature rise, interconnect congestion, or the arrival of high-priority tasks.
[0005] Therefore, there is a need for a dynamic resource scheduling technology that can comprehensively sense computing, storage, interconnection, power consumption, temperature and health status, combine AI task dependencies and historical execution characteristics to predict the effectiveness of candidate solutions, and implement local and stable rescheduling when state deviations occur. Summary of the Invention
[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a dynamic resource scheduling method for AI computing power modules to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: Acquire the operational status data of each AI computing power module and the topology status data between modules to form a corresponding resource status vector; The AI task to be scheduled is obtained. By extracting the operator attributes in the task model computation graph, operators whose memory requirements are greater than the threshold of the local storage capacity available in a single module are tensor-segmented, or the layers with large amounts of cross-module data transmission are used as boundaries. Thus, the AI task is parsed into an execution dependency graph and a resource requirement vector. The nodes of the execution dependency graph correspond to the segmented operators, model segments, data batches or training stages, and the edges of the execution dependency graph correspond to execution dependency relationships or data transmission relationships. Based on the resource demand vector and preset hard constraints, a candidate resource set corresponding to each node is generated from the allocable resources of the multiple AI computing power modules. Based on the resource state vector, the topology state data, and the historical execution profile, the task completion delay, energy consumption, heat margin, inter-module communication overhead, and reconfiguration or migration overhead of each candidate resource allocation scheme are predicted, and the scheduling priority of each candidate resource allocation scheme is determined according to the prediction results. Based on the execution dependency graph and the scheduling priority, AI accelerator core groups, local storage areas and interconnect bandwidth resources are allocated to the nodes, and a resource scheduling scheme is generated and distributed. During the execution of the AI task, the actual running status is continuously acquired, the actual running status is compared with the predicted running status corresponding to the resource scheduling scheme, and when the rescheduling trigger condition is met, the affected nodes and their associated nodes are identified. Incremental rescheduling is implemented on the task subgraph composed of the affected nodes and their associated nodes, while keeping the resource allocation relationship outside the task subgraph unchanged.
[0008] Preferably, the resource status vector includes at least two of the following: accelerator computing unit utilization, available local storage capacity, storage access bandwidth, interconnect bandwidth utilization, queue length to be executed, operating temperature, power consumption, fault health, current loaded model or configuration status, and communication distance with other AI computing power modules; the operating status data is collected using different sampling periods and then formed into the resource status vector after time alignment, outlier filtering, and sliding window aggregation.
[0009] Preferably, the resource requirement vector includes at least two of the following: target computation type, computational load, weight data volume, intermediate feature data volume, video memory or on-chip storage requirement, input / output bandwidth requirement, precision mode, maximum allowable latency, task priority, data isolation level, and parallelizable mode; the execution dependency graph also records the earliest executable time, critical path attributes, and allowable splitting granularity of each node.
[0010] Preferably, the hard constraints include at least one of the following: instruction set or operator compatibility constraints, model accuracy compatibility constraints, local storage capacity constraints, data isolation constraints, maximum power consumption constraints, maximum operating temperature constraints, and service level constraints; when generating the candidate resource set, resources that do not meet the hard constraints are first removed, and then the remaining resources are topology pruned according to the inter-module communication distance, the scale of available continuous resources, and the current model residency status.
[0011] Preferably, the historical execution profile is indexed according to model identifier, operator type, input scale, precision mode, resource slice size, and operating temperature range; when predicting each candidate resource allocation scheme, the historical execution profile matching the current task characteristics is called to obtain the initial prediction value; after the task within the current execution window is completed and measured data is collected, the corresponding record in the historical execution profile is asynchronously updated using an exponentially weighted moving average algorithm; for example, the historical profile update formula for latency is: This is used as the initial prediction value for subsequent scheduling of similar tasks. The smoothing weighting coefficient is between 0 and 1 (e.g., preferably 0.2). The online correction of measured power consumption and temperature rise uses the same weighted smoothing logic to dynamically smooth transient noise.
[0012] Preferably, when determining the scheduling priority, the task completion delay, energy consumption, thermal margin, inter-module communication overhead, reconfiguration or migration overhead, resource fragmentation, and task fairness are normalized, and corresponding weight combinations are selected according to the task category; for real-time tasks, the maximum allowable delay and thermal safety constraints are satisfied first; for throughput tasks, under the premise of satisfying the hard constraints, the throughput per unit energy consumption and continuous resource utilization are improved first.
[0013] Preferably, when a single AI computing power module cannot meet the resource demand vector, the AI task is divided into multiple task segments that can be executed in parallel or in pipeline according to the execution dependency graph, and at least one parallel method is selected from data parallelism, tensor parallelism, pipeline parallelism and operator parallelism to allocate the task segments to multiple AI computing power modules, while reserving interconnect bandwidth and data buffer areas for adjacent task segments.
[0014] Preferably, the rescheduling triggering conditions include at least one of the following: prediction completion delay deviation exceeds the delay deviation threshold, operating temperature exceeds the temperature warning threshold, power consumption exceeds the power consumption budget, fault health is lower than the health threshold, interconnection congestion lasts for more than a preset window, and high-priority task arrival. When implementing incremental rescheduling, resource scaling or core replacement within the same AI computing power module is preferred. Cross-module migration is only performed when the scheduling target cannot be met. Migration loss and scheduling jitter are reduced through checkpoints, model parameter reuse, cache preheating, and dual-threshold hysteresis mechanisms.
[0015] This invention also provides a dynamic resource scheduling system for AI computing power modules, comprising: The module includes a state perception and prediction module, a task parsing and requirement modeling module, a candidate generation and scheduling evaluation module, a resource orchestration and feedback correction module, and local execution agents deployed on multiple AI computing power modules. The state perception and prediction module is used to acquire the operating status data of each AI computing power module and the topology status data between modules, form a resource status vector, and predict the running results of candidate resource allocation schemes based on historical execution profiles. The task parsing and requirement modeling module is used to parse the AI task to be scheduled into an execution dependency graph and a resource requirement vector. The candidate generation and scheduling evaluation module is used to generate a set of candidate resources based on hard constraints, predict and evaluate the task completion delay, energy consumption, heat margin, inter-module communication overhead, and reconfiguration or migration overhead of each candidate resource allocation scheme, and determine the scheduling priority. The resource orchestration and feedback correction module is used to generate a resource scheduling scheme based on the execution dependency graph and the scheduling priority, and to perform incremental rescheduling on the affected task subgraph when the rescheduling trigger condition is met. The local execution agent is used to configure the AI accelerator core group, local storage area and interconnect bandwidth resources of the corresponding AI computing power module according to the resource scheduling scheme, execute AI tasks and report the actual running status.
[0016] Preferably, the resource orchestration and feedback correction module has a hierarchical structure, including a global scheduling controller and local scheduling controllers corresponding to each AI computing power module; the global scheduling controller maintains resource scheduling schemes with version numbers and validity periods, and the local scheduling controller performs resource occupation, release and migration operations according to resource leases, and when version conflicts, communication interruptions or AI computing power module failures are detected, the most recent executable resource scheduling scheme is retained or the unfinished tasks are switched to backup AI computing power modules according to a preset rollback strategy.
[0017] The technical effects and advantages of this invention are as follows: This invention incorporates computing power, local storage, interconnect bandwidth, queue status, power consumption, temperature, health, and model residency status into a resource state vector, avoiding resource misjudgments caused by a single utilization metric and improving the actual executability of candidate resources. Based on this, by establishing an AI task execution dependency graph and resource demand vector, scheduling is performed at the granularity of operators, model segmentation, data batching, or training phases. It also identifies critical paths and cross-module data transmission relationships, making the scheduling results more closely reflect the real execution characteristics of AI tasks. To improve prediction accuracy, a combination of historical execution profiles and online correction is used to adapt to performance differences under different models, input scales, accuracy modes, and temperature ranges, thereby more accurately predicting latency, energy consumption, and temperature rise. Regarding optimization objectives, task completion latency, energy consumption, heat margin, communication overhead, migration overhead, resource fragmentation, and fairness are evaluated simultaneously. Different optimization focuses can be selected based on the needs of real-time tasks, throughput tasks, or energy-saving tasks, enhancing the adaptability of the scheduling strategy. When system changes occur, the system identifies the affected task subgraphs and performs incremental rescheduling, maintaining the original resource allocation of unaffected nodes and reducing the overhead of duplicate model loading, cache invalidation, and cross-module migration caused by global reordering. Finally, by leveraging mechanisms such as prioritizing scaling within the same module, postponing cross-module migration, dual-threshold hysteresis, resource leases, and version control, scheduling jitter is effectively suppressed, and execution continuity and consistency of the scheduling scheme are guaranteed in the event of a failure. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the method flow of the present invention.
[0019] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0020] 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.
[0021] Example 1 like Figure 1As shown, this implementation operates on a computing device comprising multiple AI computing modules. The scheduling controller communicates with the local execution agent in each AI computing module. The local execution agent reads data from the accelerator driver layer, storage management unit, interconnect controller, power management unit, and temperature sensor via an out-of-band management interface (such as the Intelligent Platform Management Interface (IPMI)) or an in-band system API (e.g., periodically polling the registers of the Baseboard Management Controller (BMC) within each module via the I2C / SMBus bus to obtain real-time power consumption and temperature values), and is capable of performing core group partitioning, storage area allocation, model loading, bandwidth limiting, task initiation, task pausing, checkpoint generation, and resource release operations.
[0022] During the state awareness phase, the scheduling controller acquires resource information according to different sampling periods. Computational utilization, queue length, and interconnect bandwidth occupancy can use shorter sampling periods, while temperature and power consumption can use medium sampling periods. Fault health and model residency status can be updated when an event occurs. Data from different sources, after timestamp alignment, missing value imputation, outlier filtering, and sliding window statistics, forms the resource state vector of the i-th AI computing module at time t: in, Indicates available computing power. Indicates available local storage capacity. Indicates available interconnect bandwidth. Indicates queue load. Indicates power consumption. Indicates the operating temperature. Indicates the health status of the fault. Indicates the model or configuration residency status. This represents the topological distance or communication cost between modules. To eliminate dimensional differences, each dimension can be normalized according to preset upper and lower bounds or sliding window statistics.
[0023] During the task parsing phase, the scheduler controller reads the model computation graph, task description, input scale, precision mode, deadline, priority, isolation level, and allowed parallelism, converting the task into a directed acyclic execution dependency graph G=(V,E). Each node in the node set V can correspond to an operator, a cluster of consecutive operators, a model segment, a data batch, or a training phase; the edge set E represents the execution order and data transmission relationships. For the j-th node, a resource requirement vector is constructed: in, For computational load or required accelerator core size, For local storage needs, The input / output bandwidth requirements are given, as are the weights and the amount of intermediate feature data. For the maximum allowable delay, Priority For isolation level, To allow for parallel processing, the scheduler controller can also calculate the earliest start time, latest finish time, critical path weight, and descalable granularity of nodes based on the execution dependency graph.
[0024] During the candidate resource generation phase, the scheduling controller first performs hard constraint screening. If an AI computing module does not support the operators, precision modes, or instruction sets required by the node, or if its available local storage capacity is less than the node's requirements, or if the isolation level, power consumption limit, temperature limit, and deadline constraints cannot be met, then the corresponding resource of that module is removed from the candidate set. For the remaining resources, pruning is performed according to the model's residency status, the size of consecutive idle core groups, topology distance, and data source location. Resources whose models are already resided and are located in the same module or adjacent modules as their predecessor nodes receive higher candidate rankings.
[0025] During the prediction phase, the scheduling controller retrieves records from the historical execution profile database that match the current model identifier, node type, input scale, accuracy mode, resource slice size, and temperature range. The prediction completion delay for node j in candidate scheme r can be expressed as: in, To delay queuing, To calculate the delay, For storage access latency, To reduce cross-module transmission latency, This includes latency for model loading, core reconfiguration, or cache warm-up. Computation latency and storage access latency can be obtained through regression analysis of historical data, or estimated by dividing computational load by effective computing power and data volume by effective storage bandwidth.
[0026] Predicted heat margins corresponding to candidate solutions It can be represented as: in, The maximum allowable operating temperature for the candidate resource. The current temperature. The predicted temperature rise generated for execution node j. When the value is less than the safety margin, candidate solutions are eliminated or subject to a significant penalty. Predicted energy consumption can be obtained by multiplying the predicted average power consumption by the predicted execution time.
[0027] A normalized multi-objective evaluation is performed on the retained candidate solutions. Taking the case where smaller values are considered better, the overall cost can be expressed as: Where N(·) represents the normalization function, Indicates predicted energy consumption. Indicates communication overhead. This indicates the overhead of reconfiguration or migration. Indicates resource fragmentation. , , , , and Weights are set based on task category. For real-time tasks, increase... and For throughput tasks, increase the weights related to continuous resource utilization and throughput per unit energy consumption; for energy-saving tasks, increase the weights related to energy consumption. Scheduling priority can be determined by the reciprocal of the overall cost, ranking, or hierarchical constraint results.
[0028] The scheduling controller selects nodes according to the topological order of the execution dependency graph and the critical path priority principle, choosing resources with lower overall costs from the candidate set. The allocation includes at least accelerator cores, local storage areas, and interconnect bandwidth. For data requiring cross-module transmission, interconnect time windows and receive buffers are reserved before the preceding node starts to avoid insufficient communication resources despite allocated computing resources. For nodes that allow parallelism, data parallelism, tensor parallelism, pipelined parallelism, or operator parallelism can be employed, respectively.
[0029] During the task execution phase, the local execution agent completes model loading, resource slicing configuration, and task initiation according to the scheduling scheme, and periodically reports actual latency, utilization, power consumption, temperature, interconnect congestion, and error information. The scheduling controller compares the actual state with the predicted state. State deviation metrics can be defined: in, Indicates the degree to which resource utilization exceeds limits. Indicates the degree to which power consumption exceeds the budget. Indicates the degree to which the temperature exceeds the warning threshold. Indicates the risk of failure. , , ,and The coefficient is denoted as Δ(t). When Δ(t) exceeds the upper trigger threshold for multiple consecutive sampling windows, rescheduling is initiated; the rescheduling state is revoked only when Δ(t) drops below the lower recovery threshold and remains below the preset window, thus forming a double threshold hysteresis.
[0030] in, The specific calculation method is as follows: The specific calculation method is as follows: The specific calculation method is as follows: During rescheduling, the node causing the deviation is first located. A breadth-first search (BFS) or depth-first search (DFS) algorithm is used to expand forward and backward along the edges of the execution dependency graph. When the expansion reaches a node whose predicted completion time does not exceed the slack time set in the original scheduling scheme, or when a node is already in execution and cannot be interrupted, the expansion search stops. The set of associated nodes satisfying the above traversal conditions forms an affected task subgraph. Incremental rescheduling is then performed on the task subgraph composed of the affected nodes and their associated nodes. The scheduling controller maintains the nodes and resource allocation relationships outside the task subgraph unchanged and prioritizes adding core groups, adjusting local storage quotas, or replacing low-priority tasks within the original AI computing power module. Cross-module migration is only selected when the original module cannot meet the requirements.
[0031] For candidate migration schemes, the migration cost is... It can be represented as: in, This represents the amount of state data that needs to be migrated. For effective transmission bandwidth, Tcheckpoint is the time when a checkpoint is generated. For model or parameter loading time, For cache warm-up time, This represents the computational loss due to the lack of checkpoints. Cross-module migration is only performed if the expected benefits of the remaining tasks after migration exceed the sum of migration costs and safety margins.
[0032] 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 dynamic resource scheduling method for an AI computing power module, characterized in that, include: Acquire the operational status data of each AI computing power module and the topology status data between modules to form a corresponding resource status vector; The AI task to be scheduled is obtained. By extracting the operator attributes in the task model computation graph, operators whose memory requirements are greater than the threshold of the local storage capacity available in a single module are tensor-segmented, or the layers with large amounts of cross-module data transmission are used as boundaries. Thus, the AI task is parsed into an execution dependency graph and a resource requirement vector. The nodes of the execution dependency graph correspond to the segmented operators, model segments, data batches or training stages, and the edges of the execution dependency graph correspond to execution dependency relationships or data transmission relationships. Based on the resource demand vector and preset hard constraints, a candidate resource set corresponding to each node is generated from the allocable resources of the multiple AI computing power modules. Based on the resource state vector, the topology state data, and the historical execution profile, the task completion delay, energy consumption, heat margin, inter-module communication overhead, and reconfiguration or migration overhead of each candidate resource allocation scheme are predicted, and the scheduling priority of each candidate resource allocation scheme is determined according to the prediction results. Based on the execution dependency graph and the scheduling priority, AI accelerator core groups, local storage areas and interconnect bandwidth resources are allocated to the nodes, and a resource scheduling scheme is generated and distributed. During the execution of the AI task, the actual running status is continuously acquired, the actual running status is compared with the predicted running status corresponding to the resource scheduling scheme, and when the rescheduling trigger condition is met, the affected nodes and their associated nodes are identified. Incremental rescheduling is implemented on the task subgraph composed of the affected nodes and their associated nodes, while keeping the resource allocation relationship outside the task subgraph unchanged.
2. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: The resource status vector includes at least two of the following: accelerator computing unit utilization, available local storage capacity, storage access bandwidth, interconnect bandwidth utilization, queue length to be executed, operating temperature, power consumption, fault health, current loaded model or configuration status, and communication distance with other AI computing power modules. The operating status data is collected using different sampling periods and is formed into the resource status vector after time alignment, outlier filtering, and sliding window aggregation.
3. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: The resource requirement vector includes at least two of the following: target computation type, computational load, weight data volume, intermediate feature data volume, video memory or on-chip storage requirements, input / output bandwidth requirements, precision mode, maximum allowable latency, task priority, data isolation level, and parallelization mode; the execution dependency graph also records the earliest executable time, critical path attributes, and allowable splitting granularity of each node.
4. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: The hard constraints include at least one of the following: instruction set or operator compatibility constraints, model accuracy compatibility constraints, local storage capacity constraints, data isolation constraints, maximum power consumption constraints, maximum operating temperature constraints, and service level constraints. When generating the candidate resource set, resources that do not meet the hard constraints are first removed, and then the remaining resources are topology pruned according to the inter-module communication distance, the scale of available continuous resources, and the current model residency status.
5. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: The historical execution profile is indexed according to model identifier, operator type, input scale, precision mode, resource slice size, and operating temperature range. When predicting each candidate resource allocation scheme, the historical execution profile that matches the current task characteristics is called to obtain the initial prediction value, and the initial prediction value is corrected online using the measured data in the current execution window and the exponential weighted moving average algorithm.
6. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: When determining the scheduling priority, the task completion delay, energy consumption, thermal margin, inter-module communication overhead, reconfiguration or migration overhead, resource fragmentation, and task fairness are normalized, and the corresponding weight combination is selected according to the task category. For real-time tasks, the maximum allowable delay and thermal safety constraints are given priority. For throughput tasks, under the premise of meeting the hard constraints, the throughput per unit energy consumption and continuous resource utilization are given priority.
7. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: When a single AI computing power module cannot meet the resource demand vector, the AI task is divided into multiple task segments that can be executed in parallel or in pipeline according to the execution dependency graph. At least one parallel method is selected from data parallelism, tensor parallelism, pipeline parallelism and operator parallelism. The task segments are allocated to multiple AI computing power modules, while reserving interconnect bandwidth and data buffer areas for adjacent task segments.
8. The dynamic resource scheduling method for an AI computing power module according to claim 1, characterized in that: The rescheduling triggering conditions include at least one of the following: prediction completion delay deviation exceeds the delay deviation threshold, operating temperature exceeds the temperature warning threshold, power consumption exceeds the power consumption budget, fault health is lower than the health threshold, interconnection congestion continues for more than the preset window, and high-priority tasks arrive. When implementing incremental rescheduling, resource scaling or core replacement within the same AI computing power module is preferred. Cross-module migration is only performed when the scheduling target cannot be met. Migration loss and scheduling jitter are reduced through checkpoints, model parameter reuse, cache preheating, and dual-threshold hysteresis mechanisms.
9. A dynamic resource scheduling system for an AI computing power module, using the dynamic resource scheduling method for an AI computing power module as described in any one of claims 1-8, characterized in that, include: The module includes a state perception and prediction module, a task parsing and requirement modeling module, a candidate generation and scheduling evaluation module, a resource orchestration and feedback correction module, and local execution agents deployed on multiple AI computing power modules. The state perception and prediction module is used to acquire the operating status data of each AI computing power module and the topology status data between modules, form a resource status vector, and predict the running results of candidate resource allocation schemes based on historical execution profiles. The task parsing and requirement modeling module is used to parse the AI task to be scheduled into an execution dependency graph and a resource requirement vector. The candidate generation and scheduling evaluation module is used to generate a set of candidate resources based on hard constraints, predict and evaluate the task completion delay, energy consumption, heat margin, inter-module communication overhead, and reconfiguration or migration overhead of each candidate resource allocation scheme, and determine the scheduling priority. The resource orchestration and feedback correction module is used to generate a resource scheduling scheme based on the execution dependency graph and the scheduling priority, and to perform incremental rescheduling on the affected task subgraph when the rescheduling trigger condition is met. The local execution agent is used to configure the AI accelerator core group, local storage area and interconnect bandwidth resources of the corresponding AI computing power module according to the resource scheduling scheme, execute AI tasks and report the actual running status.
10. A dynamic resource scheduling system for an AI computing power module according to claim 1, characterized in that: The resource orchestration and feedback correction module has a hierarchical structure, including a global scheduling controller and local scheduling controllers corresponding to each AI computing power module; The global scheduling controller maintains resource scheduling schemes with version numbers and validity periods. The local scheduling controller performs resource occupation, release and migration operations based on resource leases. When version conflicts, communication interruptions or AI computing power module failures are detected, the most recent executable resource scheduling scheme is retained or the unfinished tasks are switched to the backup AI computing power module according to the preset rollback strategy.