A computing task scheduling method, a computer device and a medium

Abstract

The present application relates to the technical field of computer data processing, and provides a kind of computing task scheduling method, computer equipment and medium, comprising: obtaining semantic feature data representing resource demand and model structure characteristics;Determine the strategy of activation scheduling or termination execution based on the feature difference degree of target state and current running state;In response to activation scheduling, obtain resource evolution trend data of candidate node future period;Determine target node and schedule execution based on the matching degree of semantic feature data and resource evolution trend data.The present application realizes the accurate matching of task demand and future resource of node and the dynamic intervention of invalid calculation, improves the utilization rate of heterogeneous cluster computing power and scheduling robustness.

Description

Technical Field

[0001] This application relates to the field of computer data processing technology, and in particular to a computing task scheduling method, computer equipment, and medium. Background Technology

[0002] Currently, in computing clusters composed of multiple heterogeneous computing nodes, task scheduling typically employs matching strategies based on static resource snapshots or fixed thresholds. However, this traditional approach only reflects the instantaneous resource status of nodes and cannot accurately characterize the deep semantic requirements of computing tasks for underlying hardware resources. It also struggles to predict the resource evolution trends of computing nodes over future periods, leading to a spatiotemporal mismatch between task demand and node supply, resulting in low scheduling accuracy. Furthermore, existing scheduling mechanisms lack effective monitoring and intervention methods for task execution states deviating from their target states. When a task becomes stuck in ineffective computation due to changes in data distribution or model drift, the system cannot promptly identify and terminate the task, causing valuable computing resources to be continuously occupied and severely impacting the overall throughput efficiency and real-time response of the cluster. Summary of the Invention

[0003] To address the problems of low task scheduling accuracy and lack of dynamic intervention mechanisms for invalid computation in existing computing power clusters, this application proposes a computing task scheduling method, computer equipment, and media to achieve precise matching between computing task requirements and the future resource status of nodes, as well as adaptive control of task execution status.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for scheduling computing tasks includes: acquiring semantic feature data of a computing task, the semantic feature data representing the resource requirements and model structure features of the computing task; determining an execution control strategy for the computing task based on the feature difference between the target state features and the current running state features of the computing task, the execution control strategy including activating scheduling or terminating execution; in response to the execution control strategy being activation scheduling, acquiring resource evolution trend data of candidate computing nodes, the resource evolution trend data representing the multidimensional resource state prediction values ​​of the candidate computing nodes in a future time period; determining a target computing node from the candidate computing nodes based on the matching degree between the semantic feature data and the resource evolution trend data, and scheduling the computing task to the target computing node for execution.

[0005] The above scheme overcomes the spatiotemporal limitations of traditional static scheduling by performing two-dimensional matching of semantic feature data representing the deep needs of tasks with resource evolution trend data representing the future supply capacity of nodes. At the same time, it introduces an execution control strategy based on feature difference degree to evaluate the necessity of task activation before scheduling and has the ability to terminate invalid tasks during runtime, thereby constructing a closed-loop mechanism of "precise scheduling-dynamic management".

[0006] As one implementation, determining the execution control strategy for the computing task based on the feature difference between the target state features and the current running state features includes: obtaining a first feature vector corresponding to the target state features and a second feature vector corresponding to the current running state features; calculating the similarity between the first feature vector and the second feature vector, and determining the feature difference based on the model architecture topology change of the computing task; determining the execution control strategy as activation scheduling in response to the feature difference being greater than a preset threshold; and determining the execution control strategy as termination execution in response to the feature difference not being greater than the preset threshold and the computing task being in a running state.

[0007] This implementation method quantifies the deviation between the target state and the current running state in the feature space and integrates the physical changes in the model architecture topology, so that the decision-making basis for the execution control strategy not only includes differences at the numerical level, but also covers the stability at the structural level, thereby improving the accuracy and robustness of task activation and termination judgments.

[0008] As one implementation method, obtaining semantic feature data of a computing task includes: structurally encoding the computation graph of the computing task using a graph neural network model, extracting operator types, data dependency paths, parallelism constraints, and memory access patterns from the computation graph; and generating a low-dimensional dense semantic vector based on the extracted operator types, data dependency paths, parallelism constraints, and memory access patterns, as the semantic feature data.

[0009] This implementation utilizes graph neural networks to structurally encode the computation graph, a low-level execution abstraction, transforming discrete operator attributes and dependencies into continuous dense vector representations. This enables semantic feature data to accurately capture the actual resource consumption patterns of tasks at the hardware execution level, laying a data foundation for subsequent accurate matching.

[0010] As one implementation method, obtaining the resource evolution trend data of candidate computing nodes includes: obtaining multi-dimensional resource status data of the candidate computing nodes within a historical time window; inputting the multi-dimensional resource status data into a time series prediction model, and outputting the resource availability and confidence interval of the candidate computing nodes within the future time period as the resource evolution trend data.

[0011] This implementation method uses a time-series prediction model to mine the time dependence of historical resource states, expanding the basis for scheduling decisions from "static snapshots of the current moment" to "dynamic probability distributions of future time periods," enabling the scheduler to predict node computing power fluctuations and resource bottlenecks, and achieve forward-looking resource allocation.

[0012] As one implementation method, the multidimensional resource status data includes at least one of computing unit utilization, memory allocation rate, device temperature, network throughput, and task completion rate; the time-series prediction model is a long short-term memory network model based on the attention mechanism, and the resource evolution trend data output by the time-series prediction model also includes the probability of abnormal fluctuations of resources in each dimension.

[0013] This implementation further refines the physical dimension of resource status and adopts a time series model with an attention mechanism to capture the characteristics of long-term dependencies and key time steps. At the same time, it outputs the probability of abnormal fluctuations, providing risk quantification indicators for scheduling decisions and enhancing the system's adaptability in uncertain environments.

[0014] As one implementation, determining the target computing node from the candidate computing nodes based on the matching degree between the semantic feature data and the resource evolution trend data includes: generating a scheduling score matrix based on the semantic feature data and the resource evolution trend data, wherein the elements in the scheduling score matrix are weighted scores of the semantic matching degree and evolution trend consistency between task and node pairs; and determining the target computing node from the candidate computing nodes based on the weighted scores in the scheduling score matrix.

[0015] This implementation constructs a two-dimensional evaluation system that includes semantic matching degree and evolutionary trend consistency. It comprehensively considers the degree of fit between task requirements and node supply in two dimensions, namely "current adaptability" and "future sustainability", through a weighted score, thus avoiding short-sighted scheduling decisions caused by a single-dimensional evaluation.

[0016] As one implementation method, the determination of the evolution trend consistency includes: deriving a task demand trend vector based on the expected resource consumption curve of the main execution function of the computing task; obtaining the node resource evolution trend vector of the candidate computing node; calculating the similarity between the task demand trend vector and the node resource evolution trend vector to obtain the evolution trend consistency.

[0017] This implementation maps the future resource requirements of the task and the future resource supply of the node into trend vectors. By calculating the vector similarity, it achieves dynamic alignment of the supply and demand sides on the time axis, ensuring that the selected target node not only meets the requirements at the current moment, but also maintains the stability of resource supply throughout the entire task execution cycle.

[0018] As one implementation, before scheduling the computing task to the target computing node for execution, the method further includes: simulating the execution trajectory of the computing task on the target computing node; in response to the predicted delay determined based on the execution trajectory exceeding a delay threshold, or the probability of resource exhaustion being greater than a preset probability, reverting the computing task to an alternative computing node and triggering a system warning.

[0019] This implementation introduces a simulation verification phase before formal scheduling, serving as a safety net for scheduling decisions. This effectively intercepts potential scheduling failures caused by prediction deviations or sudden loads, ensuring the system's service quality and operational safety.

[0020] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the computing task scheduling method as described above.

[0021] In addition, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the computing task scheduling method as described above. Beneficial effects

[0022] The computing task scheduling method, device, and medium provided in this application fundamentally change the limitation of traditional scheduling that relies solely on static resource snapshots by acquiring semantic feature data characterizing the resource requirements and model structure of computing tasks and combining it with resource evolution trend data of candidate computing nodes over a future time period for matching and scheduling. The principle is that semantic feature data reveals the deep demand patterns of tasks at the hardware execution level, while resource evolution trend data depicts the dynamic changes in node supply capacity. This two-dimensional matching achieves precise alignment of task demand and node supply in the spatiotemporal dimensions, significantly improving the computing power utilization and task execution efficiency of heterogeneous clusters.

[0023] Meanwhile, this application determines the execution control strategy by calculating the feature difference between the task's target state and its current running state, thus constructing a dynamic management and control mechanism for the entire task lifecycle. When the feature difference indicates that the task has deviated from its expected goal or is in an ineffective running state, the system can terminate execution in a timely manner, avoiding the idleness and waste of computing resources. Furthermore, by combining the difference assessment with changes in the model architecture topology before scheduling, the adaptability of the control strategy to complex scenarios such as model drift is further enhanced. Combined with the execution trajectory simulation and rollback warning mechanism before scheduling, this application effectively improves the robustness and service reliability of the system while ensuring scheduling accuracy, meeting the stringent requirements of highly sensitive scenarios such as real-time AI inference. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a computational task scheduling method provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. Example 1

[0027] like Figure 1 As shown, this embodiment provides a computing task scheduling method. This method is applied to a computing power cluster composed of multiple heterogeneous computing nodes, aiming to solve the problem that traditional static scheduling cannot accurately match the deep needs of tasks and dynamic changes in nodes. The method includes the following steps: Step S101: Obtain semantic feature data of the computing task. The semantic feature data represents the resource requirements and model structure characteristics of the computing task.

[0028] Specifically, the semantic feature data here does not refer to textual semantics in the field of natural language processing, but rather to low-level computational features that reflect the actual behavioral patterns of computational tasks at the hardware execution level. For example, this semantic feature data may include the types of operators in the computation graph and their dependencies, data access patterns between registers and video memory, constraints of parallel computing, and sensitivity to memory bandwidth. By transforming abstract task requirements into quantifiable low-level structural features, the system can identify the differentiated requirements of different tasks for computational units, storage levels, and communication bandwidth, thereby providing a data foundation for subsequent precise scheduling. It should be understood that although this embodiment uses computation graph structural features as an example, in other embodiments, semantic feature data may also include instruction flow features, compiler intermediate representation features, or other digital descriptions that can characterize task resource consumption patterns, as long as they can establish a mapping relationship between task logic and physical resources.

[0029] Step S102: Based on the feature difference between the target state features and the current running state features of the computing task, determine the execution control strategy of the computing task. The execution control strategy includes activating scheduling or terminating execution.

[0030] Specifically, this step establishes a dynamic management and control mechanism for the entire task lifecycle. The target state characteristics represent the expected ideal operating indicators or convergence state of the task, while the current operating state characteristics reflect the real-time execution status of the task. Feature variance is used to quantify the degree of deviation between the two, which may stem from changes in input data distribution, drift in model parameters, or fluctuations in the hardware environment. When the feature variance indicates that the task has significantly deviated from its expected target or fallen into an ineffective computational loop, the execution control strategy will determine to terminate execution, promptly sending a stop command to the node to release the occupied computing resources; conversely, if the variance is within the allowable range or indicates that the task needs to reallocate resources to adapt to the new state, then scheduling will be activated. In this way, the present invention achieves a shift from passive response to proactive intervention, effectively avoiding the idle and wasted computing power caused by task failure.

[0031] Step S103: In response to the execution control strategy of activating scheduling, obtain the resource evolution trend data of candidate computing nodes. The resource evolution trend data represents the multi-dimensional resource state prediction value of candidate computing nodes in the future time period.

[0032] Specifically, resource evolution trend data is an objective prediction of the future state of physical hardware, rather than a subjective estimate based on experience. This data is typically generated through time-series analysis of multi-dimensional hardware monitoring data collected within historical time windows, covering the changing trajectories of physical indicators such as computing unit utilization, memory allocation rate, device temperature, and network throughput over a specific future time period. For example, the system can predict that the memory utilization of a GPU node will increase from 60% to 85% within the next 30 seconds, accompanied by periodic computing power peaks. This forward-looking state awareness capability allows scheduling decisions to move beyond the current instantaneous snapshot, predicting the future supply capacity and potential bottlenecks of nodes, thereby avoiding scheduling failures or performance fluctuations caused by sudden resource changes.

[0033] Step S104: Based on the matching degree between semantic feature data and resource evolution trend data, determine the target computing node from the candidate computing nodes and schedule the computing task to be executed on the target computing node.

[0034] Specifically, this step achieves a two-dimensional alignment between task requirements and node supply in both time and space. The matching degree calculation considers not only the adaptability of task semantic features to the current resources of nodes, but more importantly, the consistency between the future resource consumption trend of the task and the future resource supply trend of nodes. For example, for an inference task about to enter a high memory consumption phase, the system will prioritize nodes with sufficient memory and no other high-load tasks competing for it in the future, even if the node's current memory utilization is not the lowest. Through this dynamic matching based on evolutionary trends, this invention overcomes the common time-space mismatch problem in traditional static scheduling, significantly improving the overall throughput efficiency and task execution stability of heterogeneous clusters.

[0035] In summary, this embodiment constructs a complete technical closed loop of "feature perception - dynamic control - trend prediction - precise matching" through steps S101 to S104. Specifically, the execution control strategy in step S102 and the scheduling matching logic in step S104 constitute a dual-parallel processing mechanism: the former is responsible for filtering out valid tasks worthy of scheduling and promptly blocking invalid computations, while the latter is responsible for finding the optimal execution node for valid tasks that is spatiotemporally adapted. Working together, they ensure the efficient operation of the computing cluster in complex and dynamic environments. Example 2

[0036] This embodiment, based on Embodiment 1, further elaborates on the method for acquiring semantic feature data in step S101 and the logic for determining the execution control strategy in step S102. By introducing graph structure encoding and topology stability verification mechanisms, the scheduling system's ability to perceive the underlying physical attributes of tasks and the robustness of its control decisions are enhanced.

[0037] As one implementation method, obtaining semantic feature data of a computation task includes: structurally encoding the computation graph of the computation task using a graph neural network model, extracting operator types, data dependency paths, parallelism constraints, and memory access patterns from the computation graph; and generating low-dimensional dense semantic vectors based on the extracted operator types, data dependency paths, parallelism constraints, and memory access patterns as semantic feature data.

[0038] Specifically, a computation graph is a directed acyclic graph or a directed cyclic graph that describes the execution logic of a computation task. Its nodes represent operators, and edges represent the flow of data tensors. Because computation graphs possess non-Euclidean topological characteristics, traditional convolutional neural networks or recurrent neural networks are difficult to process directly. Therefore, this embodiment employs a graph neural network model for structural encoding. During encoding, operator types include, but are not limited to, convolution operations, matrix multiplication, activation functions, and normalization operations. These types directly determine the type requirements of computational units. Data dependency paths reflect the transmission order and hierarchical depth of tensors between different operators, determining the potential for pipelined parallelism or data parallelism. Parallelism constraints describe the separable dimensions and communication synchronization requirements within operators. Memory access patterns characterize the frequency and locality of read / write operations on on-chip cache, shared memory, or global memory. The graph neural network aggregates neighbor node information through a message-passing mechanism, mapping the aforementioned discrete graph structure attributes into a fixed-dimensional, low-dimensional, dense semantic vector. This vector is not abstract textual semantics, but rather a fingerprint of the task's resource consumption at the hardware execution level. In this way, the system can uniformly represent computational tasks with different structures and scales into the same vector space, so that the subsequent matching degree calculation has an operable mathematical basis, and at the same time avoids the problem of poor generalization caused by relying solely on manual rules to extract features.

[0039] As one implementation method, the execution control strategy for a computational task is determined based on the feature difference between the target state features and the current running state features. This includes: obtaining a first feature vector corresponding to the target state features and a second feature vector corresponding to the current running state features; calculating the similarity between the first feature vector and the second feature vector, and determining the feature difference based on the model architecture topology change of the computational task; determining the execution control strategy as activation scheduling in response to the feature difference being greater than a preset threshold; and determining the execution control strategy as termination execution in response to the feature difference not being greater than the preset threshold and the computational task being in a running state.

[0040] Specifically, the target state feature refers to the intermediate state or output indicator that the task should reach under the ideal execution path, usually determined by the task definition file or historical benchmark data; the current running state feature is the actual execution feedback collected through real-time monitoring. Both the first and second feature vectors are isomorphic vectors generated by the aforementioned graph neural network or other feature extractors. When calculating feature difference, in addition to considering the similarity of the two vectors in the numerical space (such as cosine similarity or Euclidean distance), a model architecture topology change is specifically introduced as a correction factor. The model architecture topology change is used to quantify the degree of change in the computational graph structure itself, such as adding or deleting operator nodes, changing the connection relationship of data flow, adjusting branch jump logic, etc. The reason for introducing this factor is that during actual inference or training, numerical fluctuations in the input data (such as noise, outliers) may cause a shift in the numerical value of the output feature vector, but this shift often does not mean that the essence of the task has changed; conversely, if the model structure undergoes substantial changes (such as conditional branch switching in dynamic graph execution), even if the numerical similarity is high, it may indicate that the task has entered a completely new execution stage or that abnormal drift has occurred. Therefore, the calculation logic of feature difference can be represented as a weighted combination or nonlinear fusion of similarity deviation and topology change. When the feature difference after comprehensive evaluation exceeds a preset threshold, it indicates that the current task state has significantly deviated from expectations or entered a new stage requiring resource reallocation. At this time, the activation scheduling process is triggered, i.e., steps S103 and S104 in Example 1 are executed to find a new suitable node. When the feature difference does not exceed the threshold and the task is still running, it indicates that the task may be trapped in an invalid loop, deadlock, or inefficient execution state, and continuing to occupy resources is no longer beneficial. At this time, the system will generate a termination instruction and send it to the computing node currently carrying the task through a high-speed communication link, forcibly releasing the computing power and storage resources it occupies. This dual verification mechanism, which combines numerical similarity and structural stability, effectively prevents misscheduling or missed termination caused by fluctuations in a single indicator, significantly improving the accuracy and security of cluster resource management.

[0041] It should be understood that although this embodiment uses graph neural networks as the preferred implementation for semantic feature extraction, other embodiments may also employ analysis methods based on compiler intermediate representations, statistical methods based on performance counters, or embedding methods based on instruction stream sequences, as long as they can generate digital features representing task resource requirements and model structure. Similarly, the formula for calculating feature difference is not limited to a specific weighted form; decision trees, support vector machines, or lightweight classification networks can also be used to achieve the fusion and judgment of multi-dimensional features. This invention does not limit this. Example 3

[0042] Based on step S103 of embodiment 1, this embodiment further details the specific implementation method for obtaining resource evolution trend data of candidate computing nodes, and transforms the abstract evolution trend into a time-series prediction mechanism that can be engineered and implemented.

[0043] As one implementation method, obtaining resource evolution trend data of candidate computing nodes includes: obtaining multi-dimensional resource status data of candidate computing nodes within a historical time window; inputting the multi-dimensional resource status data into a time series prediction model, and outputting the resource availability and confidence interval of candidate computing nodes in a future time period as resource evolution trend data.

[0044] Specifically, the historical time window refers to a continuous sampling period that traces backward from the current moment. Its length N can be dynamically configured according to the frequency of cluster load changes. For example, it can be set to the past 60 seconds in an inference cluster with drastic load fluctuations, while it can be extended to the past 300 seconds in a relatively stable training cluster. The future time period refers to the prediction span T extending backward from the current moment. This span is usually matched with the estimated execution time of the task to be scheduled to ensure that the prediction covers the entire lifecycle of the task. The multidimensional resource status data is not a logical statistical value at the software level, but rather physical data directly from hardware monitoring interfaces or onboard sensors, possessing objectivity and real-time characteristics. By modeling these historical physical data through a time-series prediction model, the system can capture the time-dependent patterns of resource consumption, thereby expanding the basis for scheduling decisions from static instantaneous snapshots to dynamic probability distribution predictions, enabling the scheduler to have forward-looking perception capabilities.

[0045] As a preferred implementation, the multidimensional resource status data includes at least one of computing unit utilization, memory allocation rate, device temperature, network throughput, and task completion rate; the time-series prediction model is a long short-term memory network model based on the attention mechanism, and the resource evolution trend data output by the time-series prediction model also includes the probability of abnormal fluctuations in resources of each dimension.

[0046] Specifically, computing unit utilization reflects the actual computational load of the GPU or NPU cores, memory allocation rate characterizes the usage level of video memory or main memory, device temperature reflects the load on the cooling system and potential thermal throttling risks, network throughput characterizes the consumption of communication bandwidth between nodes, and task completion rate indirectly reflects the processing progress and queuing pressure of the current load. All these data are acquired at a fixed sampling frequency through the underlying driver or hardware management controller, ensuring the physical authenticity of the input features. In terms of processing flow, the attention-based Long Short-Term Memory (LSTM) network model first uses LSTM units to extract the long-term trend and short-term fluctuation features of resource states within historical time windows. Then, through the attention mechanism, it assigns higher weights to key time steps, such as automatically focusing on peak periods or abrupt changes similar to the current load pattern in history, thereby improving the prediction accuracy of resource states within the future time period T. This model not only outputs the predicted mean of resources in each dimension at each future time point (i.e., resource availability), but also simultaneously outputs the corresponding confidence interval and probability of abnormal fluctuations. The confidence interval quantifies the range of uncertainty in the prediction result. For example, a 95% confidence interval indicates a 95% probability that the actual resource state falls within this range. The probability of abnormal fluctuations characterizes the likelihood of sudden resource preemption or hardware performance jitter that occurs in the future, exceeding normal patterns. These two indicators provide a risk quantification basis for the matching degree calculation in subsequent step S104: when the predicted resource availability of two candidate nodes is similar, the system can prioritize the node with a narrower confidence interval and a lower probability of abnormal fluctuations, thereby significantly improving the robustness and service reliability of the system while ensuring scheduling accuracy. It should be understood that although this embodiment uses a long short-term memory network model based on an attention mechanism as the preferred implementation of the temporal prediction model, other embodiments may also use a Transformer model, a gated recurrent unit network, a temporal convolutional network, or a prediction method based on a state-space model, as long as it can generate a future trend prediction containing uncertainty measures based on historical multidimensional physical resource data. This invention does not limit this. Example 4

[0047] This embodiment, based on step S104 of embodiment 1, further details the specific implementation method for determining target computing nodes based on the matching degree of semantic feature data and resource evolution trend data. This process achieves precise alignment of task requirements and node supply in the spatiotemporal dimensions by constructing a two-dimensional evaluation system.

[0048] As one implementation method, a scheduling score matrix is ​​generated based on semantic feature data and resource evolution trend data. The elements in the scheduling score matrix are weighted scores of the semantic matching degree and evolution trend consistency between task and node pairs. Based on the weighted scores in the scheduling score matrix, the target computing node is determined from the candidate computing nodes.

[0049] Specifically, the scheduling scoring matrix is ​​a two-dimensional data structure with candidate computing tasks as rows and candidate computing nodes as columns. Each element in the matrix comprehensively reflects the suitability of assigning a specific task to a specific node. Semantic matching focuses on evaluating the compatibility between the task's current static resource requirements and the node's current hardware architecture characteristics, such as whether the task's operator type is natively supported by the node's accelerator and whether the memory access pattern conforms to the node's cache hierarchy design. Evolutionary trend consistency, on the other hand, focuses on evaluating the synchronization between the task's dynamic resource consumption trajectory during its execution cycle and the node's future resource supply trajectory. This dual-dimensional weighted mechanism is designed to overcome the limitations of single-dimensional evaluation: relying solely on semantic matching may lead to tasks being scheduled to nodes that are currently compatible but are about to become overloaded, while relying solely on trend matching may ignore compatibility differences in the underlying instruction sets. In practical applications, the weighting coefficients are not fixed and can be dynamically adjusted by the system based on the current overall load level of the cluster or the priority of the tasks. For example, when the cluster is under high load and congestion, the weight of evolution trend consistency can be appropriately increased to prioritize the stability of task execution and reduce long-tail latency; while under low load, the weight of semantic matching can be increased to pursue the ultimate throughput performance. Finally, the system selects the node corresponding to the element with the highest score in the matrix as the target computing node, or randomly selects from the set of high-scoring nodes to achieve load balancing, provided that preset service quality constraints are met.

[0050] As a preferred implementation, the determination of evolution trend consistency includes: deriving the task demand trend vector based on the expected resource consumption curve of the main execution function of the computing task; obtaining the node resource evolution trend vector of candidate computing nodes; and calculating the similarity between the task demand trend vector and the node resource evolution trend vector to obtain the evolution trend consistency. Specifically, the expected resource consumption curve is a digital representation of the resource occupancy of the task during future execution. Its source can be a performance profile file attached when the task is released, a statistical average of historical execution records, or a sampling sequence obtained through pre-running a lightweight simulator. This curve typically contains a numerical sequence of multiple dimensions such as computing unit utilization, memory usage, and bandwidth requirements changing over time. The task demand trend vector is a vector representation of this multi-dimensional time series curve mapped to a unified feature space, preserving the temporal morphological characteristics of resource consumption. Correspondingly, the node resource evolution trend vector is a vector representation of the resource availability of nodes in various dimensions within the same future time period, output by the time series prediction model described in Example 3. The process of calculating the similarity between the two is essentially finding the optimal alignment relationship between the task demand waveform and the node supply waveform on the time axis. When the peak resource consumption period of a task coincides with the predicted resource abundance period of a node, and the resource release period of the task is offset from other peak load periods of the node, the vector similarity between the two will reach its maximum value, indicating that the node is the optimal carrier for the task in the time dimension. This matching method based on trend vector similarity embodies the scheduling principle of spatiotemporal alignment and can effectively avoid performance fluctuations caused by resource contention.

[0051] It should be understood that although this embodiment uses vector similarity as a measure of evolutionary trend consistency, in other embodiments, a dynamic time warping algorithm can be used to calculate the distance between two time series sequences to adapt to scenarios where the actual execution time of the task is not completely consistent with the prediction window length; or a cross-correlation function can be used to detect the phase difference between two sequences, and nodes with a phase difference less than a preset threshold are judged as having consistent trends. Furthermore, the generation of the expected resource consumption curve is not limited to the methods listed above; any data format that can reflect the dynamic changes in the future resource requirements of the task can be used as the basis for deriving the task requirement trend vector, and this invention does not limit this. Example 5

[0052] This embodiment, based on the target computing node determined in Embodiment 1, further adds a pre-scheduling verification and fault-tolerance mechanism to address potential deviations in timing prediction or sudden environmental interference. As a preferred implementation, before scheduling the computing task to the target computing node for execution, the method further includes: simulating the execution trajectory of the computing task on the target computing node; in response to the predicted delay determined based on the execution trajectory exceeding a delay threshold, or the probability of resource exhaustion exceeding a preset probability, rescheduling the computing task to an alternative computing node and triggering a system warning.

[0053] Specifically, the simulated execution trajectory does not involve actually running the task on physical hardware, but rather a virtual pre-simulation using digital twin simulation or lightweight performance modeling techniques. The system utilizes the resource evolution trend data obtained in Example 3 to construct a virtual runtime environment for the target computing node, and maps the semantic feature data of the computing task to a load model within this environment. This allows for rapid prediction of the task's execution performance on the node over future time periods without consuming actual computing resources. This simulation process is typically completed in milliseconds, and its core purpose is to verify the feasibility of the scheduling decision based on matching degree in step S104 at the actual execution level, essentially adding a safety valve to the scheduling system.

[0054] When determining whether to trigger a rollback, the system sets two independent safety boundaries. The first boundary is the time dimension, namely whether the prediction latency exceeds a latency threshold. This latency threshold can be dynamically specified by the task's service level agreement; for example, it can be set to 50ms for real-time inference tasks, while it can be relaxed to several seconds for offline batch processing tasks. If the simulation results show that the expected completion time of the task on the target node cannot meet this time limit requirement, it is judged as a timeout risk. The second boundary is the resource dimension, namely whether the probability of resource exhaustion is greater than a preset probability. This probability value is directly derived from the abnormal fluctuation probability or confidence interval analysis results output by the time series prediction model in Example 3. For example, when the simulation shows that the probability of the target node's memory utilization rate exceeding 95% in the future execution window is greater than 15%, even if the current memory is sufficient, the system will judge that there is an extremely high OOM risk. As long as either of the above conditions is met, the system determines that the current target computing node is no longer suitable to carry out the task.

[0055] When the rollback mechanism is triggered, the system does not directly declare scheduling failure, but automatically executes a degraded scheduling strategy. Specifically, the system will backtrack to the scheduling score matrix generated in Example 4, remove the current target computing node, select the candidate node with the second highest weighted score as the alternative computing node, and redirect the computing task to the alternative node. Simultaneously, the system will generate an early warning event containing information such as the cause of the anomaly, the original target node ID, and the prediction deviation value, and trigger the system warning through log recording, message queue push, or highlighting on the operation and maintenance dashboard, so that operation and maintenance personnel can promptly perceive cluster status anomalies or prediction model drift. It should be understood that although this embodiment uses "prediction latency" and "resource exhaustion probability" as preferred indicators for triggering the rollback, in other implementations, physical constraints such as power consumption exceeding limits and excessive temperature can be introduced as additional circuit breaker conditions. Furthermore, the selection of alternative computing nodes is not limited to the suboptimal solution in the score matrix; in extreme congestion scenarios, tasks can be suspended and waited for, or split and executed in parallel on multiple nodes. This invention does not limit this. Through this closed-loop design of "simulation verification - automatic rollback - proactive early warning", this embodiment effectively compensates for the vulnerability of pure predictive scheduling in the face of uncertainty, and significantly improves the service reliability and robustness of computing power clusters in complex dynamic environments. Example 6

[0056] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the computing task scheduling method as described in any one of embodiments 1 to 5.

[0057] Specifically, this computer device constitutes the physical execution carrier of the aforementioned computing task scheduling method, solidifying the abstract scheduling logic into a concrete hardware entity. The memory may include high-speed random access memory (RAM) and non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. In this embodiment, the memory is used not only to store the computer program code implementing the scheduling method but also to cache key data generated during runtime, such as the low-dimensional dense semantic vector generated in Embodiment 2, the weight parameters of the timing prediction model in Embodiment 3, and the scheduling scoring matrix constructed in Embodiment 4. By residing this frequently accessed data in memory, the I / O wait time of the processor when executing scheduling decisions can be significantly reduced, ensuring the response speed of real-time scheduling.

[0058] The processor is the core computing and control unit of the computer device, connected to memory and other peripherals via a system bus. The processor can be a central processing unit (CPU), or a chip with data processing capabilities such as a graphics processing unit (GPU), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC). When the processor reads and executes the computer program from memory, its internal logic circuits are configured to sequentially perform a series of actions, including acquiring semantic feature data, calculating feature differences to determine execution control strategies, acquiring resource evolution trend data, and determining target computing nodes based on matching degrees. It should be understood that although this embodiment uses a general-purpose processor as an example, other embodiments may employ a multi-core heterogeneous architecture. For example, the CPU may be responsible for scheduling logic judgment and flow control, while the GPU or NPU accelerates the inference process of graph neural network encoding and timing prediction models, thereby further improving the overall throughput performance of the scheduling system.

[0059] In addition, this computer device typically includes a communication interface connected to the processor and memory via a bus. The communication interface enables high-speed data interaction between the device and various candidate computing nodes in the computing power cluster. For example, the communication interface can be a high-performance network card supporting RDMA (Remote Direct Memory Access) or RoCE (RDMA over Converged Ethernet) protocols to ensure the acquisition of real-time resource status data from each node with microsecond-level latency and the rapid distribution of scheduling instructions to the target node. In a typical computing power cluster deployment architecture, this computer device is usually deployed independently as a main scheduling server or cluster management node, or runs as a containerized service in the cluster's control plane, responsible for coordinating the allocation and lifecycle management of all computing tasks from a global perspective. Through this integrated hardware and software design, this invention not only provides a complete methodological process but also clarifies the specific product form for the implementation of this method, providing a solid physical foundation for the practical application and rights protection of the technical solution. Example 7

[0060] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the computing task scheduling method as described in any one of embodiments 1 to 5.

[0061] Specifically, a computer-readable storage medium is a physical or logical entity that carries the software implementation of the technical solution of this invention. It exists independently of specific hardware execution devices and constitutes the basic carrier for software product distribution and licensing. In this invention, the form of the storage medium is broadly inclusive, including both tangible non-volatile storage devices and intangible signal transmission media or cloud storage space. For example, in terms of non-volatile storage, read-only memory (ROM), programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), flash memory, solid-state drive (SSD), optical discs (such as CD-ROM, DVD, Blu-ray disc), or magnetic tape can be used; in terms of volatile storage, random access memory (RAM), dynamic random access memory (DRAM), or static random access memory (SRAM) can be included, used to temporarily cache program instructions during system operation to achieve high-speed reading. Furthermore, with the development of cloud computing and edge computing technologies, the storage medium can also be embodied as a storage node in a distributed file system, a storage bucket in an object storage service, or a remote virtual disk mounted via a network protocol. Any medium that can persistently or temporarily store a sequence of binary code used to instruct the processor to perform a specific operation falls within the scope of protection of this invention.

[0062] When the computer program stored in the aforementioned medium is loaded into a processor (such as a CPU, GPU, NPU, or FPGA) and executed, its instruction sequence drives the hardware circuitry to sequentially execute the logical steps described in the foregoing embodiments. Specifically, the program instructions first control the data acquisition module to obtain the semantic feature data of the computation task, then call the difference evaluation algorithm to determine the execution control strategy, then activate the time-series prediction model to generate resource evolution trend data of candidate nodes, and finally complete the selection of target nodes and task assignment through the matching degree calculation engine. This process solidifies the abstract scheduling strategy into a repeatable machine instruction stream, enabling the scheduling mechanism based on the two-dimensional matching of semantic features and evolution trends proposed in this invention to move beyond the theoretical design stage and be transformed into a practically usable software product or service.

[0063] It should be understood that although this embodiment lists various specific forms of storage media, this is merely an illustrative example for full disclosure and is not intended to limit the scope of protection of this invention. In practical applications, this computer-readable storage medium can be sold and distributed as a standalone software installation package, or it can be pre-installed in the main scheduling server of a computing cluster and delivered along with the hardware. It can also be deployed via the Internet through online updates, container image pulls, or API service calls. Regardless of the distribution or deployment method used, as long as the program stored therein can implement the computing task scheduling method described in this invention when executed, it falls within the scope of protection of this invention. This independent protection of the storage medium effectively covers various business implementation models such as pure software sales, cloud service subscriptions, and firmware upgrades, providing a solid legal basis for comprehensive rights protection of the technical solution. Example 8

[0064] This embodiment provides a specific application scenario to further illustrate the practical effect of the present invention in real-time inference task scheduling in heterogeneous AI computing power clusters. This scenario aims to solve the inference latency jitter problem caused by GPU resource fragmentation and dynamic load changes in existing technologies. By integrating task semantic features with node resource evolution trends, it achieves accurate computing power matching and robust task management.

[0065] In this application scenario, the computing cluster consists of various GPU computing nodes of different models, including high-performance nodes suitable for massively parallel computing and low-power nodes suitable for lightweight inference. Multiple real-time AI inference services run within the cluster, such as a natural language processing model based on the Transformer architecture. When a new inference request arrives, the system first performs a step to obtain semantic feature data for the computation task. Specifically, for this Transformer inference task, the system uses a graph neural network model to structurally encode its computation graph, focusing on extracting the memory access patterns of the self-attention mechanism operator and the parallelism constraints of the matrix multiplication operator, generating a low-dimensional dense semantic vector characterizing the task's high sensitivity to GPU memory bandwidth and computational intensity. This semantic feature data accurately reflects the true resource consumption fingerprint of the inference task at the hardware level, unlike coarse-grained descriptions based solely on model names or input sizes.

[0066] Meanwhile, the system continuously monitors the running status of each candidate computing node in the cluster. For a specific GPU candidate node, the system obtains its multi-dimensional resource status data over the past 60 seconds, including GPU core utilization, memory allocation rate, and PCIe bus throughput, and inputs this data into a long short-term memory network model based on an attention mechanism. The model output shows that although the node's current memory utilization is only 40%, in the next 30 seconds, due to the upcoming peak data loading of other deployed background batch processing tasks, the node's memory bandwidth utilization is predicted to surge to over 95%, accompanied by a high probability of abnormal fluctuations. This resource evolution trend data reveals that the node does not actually have the stable ability to handle high-bandwidth-sensitive tasks in the future, thus avoiding the misjudgment caused by traditional static scheduling that only considers the current remaining memory.

[0067] During the phase of determining the execution control strategy, the system calculates the feature difference between the target state features and the current running state features of the inference task in real time. For example, when a sudden change in the length of the input text sequence causes a significant deviation between the actual computation graph's topology and the expected baseline, the feature difference exceeds a preset threshold. The system determines that the current instance is no longer suitable for the original resource configuration and then triggers the activation scheduling strategy, adding the task to the scheduling queue to find a new suitable node, instead of continuing inefficient execution on the original node or waiting for a timeout.

[0068] Subsequently, the system determines the target computing node from the candidate computing nodes based on the matching degree between semantic feature data and resource evolution trend data. When generating the scheduling score matrix, the system found that another candidate node, although its current memory utilization rate is 60%, higher than the aforementioned node, has a resource evolution trend vector over the next 30 seconds that is highly similar to the demand trend vector of the Transformer task, indicating a good supply-demand alignment on the time axis. Simultaneously, this node's hardware architecture natively supports the specific operator instruction set required by the task, resulting in a high semantic matching score. After comprehensive weighting, this node's overall score is better than nodes that appear idle at present but whose future trends are deteriorating, and therefore it is selected as the target computing node. This two-dimensional matching mechanism effectively avoids scheduling traps caused by the illusion of instantaneous resource availability, ensuring the long-tail latency stability of the inference service.

[0069] As a preferred safety fallback measure, the system also performs a simulation verification step before officially distributing the computing task to the target computing node. The system constructs a digital twin environment of the target node and simulates the execution trajectory of the inference task over the next 30 seconds. In one simulation, the system detected that due to a sudden high-priority preemptive task interference in the cluster, the target node's resource exhaustion probability in the latter half of the execution trajectory reached 18%, exceeding the preset 15% safety threshold. Although this node scored the highest in the previous matching evaluation, the system still decisively triggered a rollback mechanism, rescheduling the task to the second-highest-scoring but more stable alternative computing node, and simultaneously triggering a system alert to notify maintenance personnel to investigate the source of the sudden load. This process fully demonstrates how the present invention achieves accurate and reliable computing task scheduling in a complex and dynamic heterogeneous computing environment through a closed-loop logic of "perception-prediction-matching-verification," significantly improving the overall throughput efficiency and service availability of the cluster.

[0070] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. The graph neural network encoding method for semantic feature data, the feature difference calculation logic for execution control strategies, the temporal prediction model architecture for resource evolution trend data, the weighting mechanism of the scheduling scoring matrix, and the backoff triggering conditions for simulation verification in the foregoing embodiments are all preferred examples listed to clearly illustrate the technical solution of the present invention, and are not intended to limit the scope of protection of the present invention. Any equivalent changes, substitutions, or improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the concept of the present invention, such as using other feature extraction networks to replace graph neural networks, using different temporal analysis algorithms to obtain resource evolution trends, adjusting the matching degree evaluation dimension, or modifying the security verification threshold, should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A computational task scheduling method, characterized in that, The method includes: Obtain semantic feature data of the computing task, wherein the semantic feature data characterizes the resource requirements and model structure features of the computing task; Based on the feature difference between the target state features and the current running state features of the computing task, the execution control strategy of the computing task is determined, and the execution control strategy includes activating scheduling or terminating execution. In response to the execution control strategy of activating scheduling, resource evolution trend data of candidate computing nodes is obtained, wherein the resource evolution trend data represents the multi-dimensional resource state prediction value of the candidate computing nodes in a future time period; Based on the matching degree between the semantic feature data and the resource evolution trend data, a target computing node is determined from the candidate computing nodes, and the computing task is scheduled to be executed on the target computing node.

2. The method according to claim 1, characterized in that, The step of determining the execution control strategy for the computing task based on the feature difference between the target state features and the current running state features of the computing task includes: Obtain the first feature vector corresponding to the target state feature, and the second feature vector corresponding to the current running state feature; Calculate the similarity between the first feature vector and the second feature vector, and determine the feature difference by combining the model architecture topology change amount of the computation task; In response to the feature difference degree being greater than a preset threshold, the execution control strategy is determined to be activation scheduling; In response to the fact that the feature difference is not greater than the preset threshold and the computing task is in a running state, the execution control strategy is determined to terminate execution.

3. The method according to claim 2, characterized in that, The acquisition of semantic feature data for the computation task includes: The computation graph of the computation task is structurally encoded using a graph neural network model, and the operator type, data dependency path, parallelism constraints and memory access patterns in the computation graph are extracted. Based on the extracted operator type, data dependency path, parallelism constraint, and memory access pattern, a low-dimensional dense semantic vector is generated as the semantic feature data.

4. The method according to claim 1, characterized in that, The acquisition of resource evolution trend data for candidate computing nodes includes: Obtain the multidimensional resource status data of the candidate computing nodes within the historical time window; The multidimensional resource status data is input into the time series prediction model, and the resource availability and confidence interval of the candidate computing nodes in the future time period are output as the resource evolution trend data.

5. The method according to claim 4, characterized in that, The multidimensional resource status data includes at least one of computing unit utilization, memory allocation rate, device temperature, network throughput, and task completion rate. The time-series prediction model is a long short-term memory network model based on the attention mechanism. The resource evolution trend data output by the time-series prediction model also includes the probability of abnormal fluctuations in resources in each dimension.

6. The method according to claim 1, characterized in that, The step of determining the target computing node from the candidate computing nodes based on the matching degree between the semantic feature data and the resource evolution trend data includes: Based on the semantic feature data and the resource evolution trend data, a scheduling score matrix is ​​generated. The elements in the scheduling score matrix are weighted scores of the semantic matching degree and the consistency of the evolution trend between task and node pairs. The target computing node is determined from the candidate computing nodes based on the weighted score in the scheduling scoring matrix.

7. The method according to claim 6, characterized in that, The methods for determining the consistency of the evolutionary trend include: Based on the expected resource consumption curve of the main execution function of the computing task, the task demand trend vector is derived. Obtain the node resource evolution trend vector of the candidate computing nodes; Calculate the similarity between the task demand trend vector and the node resource evolution trend vector to obtain the consistency of the evolution trend.

8. The method according to claim 1, characterized in that, Before scheduling the computing task to the target computing node for execution, the method further includes: Simulate the execution trajectory of the computational task on the target computing node; In response to the predicted delay determined based on the execution trajectory exceeding the delay threshold, or the probability of resource exhaustion being greater than a preset probability, the computing task is rolled back to an alternative computing node, and a system warning is triggered.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images