Task scheduling method and system based on dynamic cognitive resource model, computer equipment and medium

Abstract

The invention relates to the technical field of computing resource management, in particular to a task scheduling method and system based on a dynamic cognitive resource model, computer equipment and a medium. The method comprises the following steps: acquiring resource monitoring data and task state data of a heterogeneous computing power environment; inputting the resource monitoring data and the task state data into a dynamic cognitive resource model, executing fine-grained prediction on the current resource demand through a fusion framework, and quantitatively calculating interference effect characteristics among tasks in the heterogeneous computing power environment; and generating an active prevention scheduling decision through a dynamic intervention module configured by a dynamic cognitive resource model based on the fine-grained prediction result and the interference effect characteristics. By means of the mode, the technical problem that an existing AI platform resource management technology is difficult to achieve collaborative optimization of resource utilization efficiency and scheduling reliability is solved, and the adaptability, efficiency and safety of task scheduling in the heterogeneous computing power environment are improved.

Description

Technical Field

[0001] This invention relates to the field of computing resource management technology, and in particular to a task scheduling method, system, computer device and medium based on a dynamic cognitive resource model. Background Technology

[0002] With the rapid development of artificial intelligence technology and the continuous deepening of its application scenarios, AI R&D platforms are placing higher demands on the efficiency and reliability of computing resource scheduling. In heterogeneous computing environments where tasks such as model training, fine-tuning, and large-scale inference are deployed in a mixed manner, traditional task scheduling schemes struggle to simultaneously meet the comprehensive requirements of high throughput, low latency, and high resource utilization, becoming a key bottleneck restricting the large-scale expansion of the AI ​​industry.

[0003] In the field of resource scheduling technology, existing methods have made some progress. Intelligent scheduling systems based on monitoring data and optimization algorithms are gradually being applied to cloud computing and edge computing scenarios due to their automation capabilities. Current mainstream scheduling schemes mostly rely on static rules or machine learning models, achieving task allocation through real-time monitoring of resource indicators such as CPU (Central Processing Unit) and memory, and combining reinforcement learning techniques to improve global optimization effects. However, the heterogeneous computing environment of AI R&D platforms is highly dynamic and complex, with nonlinear coupling relationships between tasks and resources. Existing technologies still have significant limitations in addressing these challenges. Reference document 1 (application publication number CN119597493A) discloses a method and system for intelligent evolution of distributed computing resources based on digital twins. It collects data through intelligent sensing probes and optimizes resource allocation by combining graph neural networks and a hierarchical reinforcement learning framework. However, this method is mainly aimed at general distributed computing scenarios, and its feature extraction and decision-making logic are more adapted to static resource environments. It fails to fully consider the sudden changes in task load, the causal relationships of resource competition, and the real-time requirements of business constraints in AI platforms. Specifically, existing technologies generally suffer from two prominent problems: First, resource awareness mechanisms rely on monitoring isolated indicators, failing to dynamically capture the long-range coupling relationships of heterogeneous computing power such as GPUs (Graphics Processing Units) or NPUs (Neural Processing Units), resulting in insufficient prediction accuracy. Second, interference handling and decision-making architectures lack proactive prevention capabilities, making it difficult to achieve global optimization while ensuring safety. These problems, when combined, make it difficult for existing technologies to balance scheduling efficiency and system stability in practical applications of AI R&D platforms, and cannot meet the high-reliability, low-latency scheduling requirements of high-risk operational scenarios.

[0004] Therefore, existing AI platform resource management technologies suffer from insufficient adaptability to dynamic environments and difficulty in balancing global optimization and local security, making it difficult to achieve coordinated optimization of resource utilization efficiency and scheduling reliability. Summary of the Invention

[0005] To address the aforementioned shortcomings or deficiencies, this invention provides a task scheduling method, system, computer device, and medium based on a dynamic cognitive resource model, which can solve the technical problem of existing AI platform resource management technologies being unable to achieve coordinated optimization of resource utilization efficiency and scheduling reliability.

[0006] This invention provides a task scheduling method based on a dynamic cognitive resource model, comprising: Obtain resource monitoring data and task status data from heterogeneous computing environments.

[0007] Resource monitoring data and task status data are input into a preset dynamic cognitive resource model. The current resource demand is then predicted in a fine-grained manner through the fusion framework configured by the dynamic cognitive resource model, and the interference effect characteristics between various tasks in the heterogeneous computing environment are quantified.

[0008] Based on fine-grained prediction results and disturbance effect characteristics, proactive prevention scheduling decisions are generated through a dynamic intervention module configured by a dynamic cognitive resource model.

[0009] Based on proactive prevention-oriented scheduling decisions, a constrained hierarchical decision-making architecture configured through a dynamic cognitive resource model is used to output target task scheduling strategies.

[0010] According to a second aspect, the present invention provides a task scheduling system based on a dynamic cognitive resource model, comprising: The task resource data acquisition module is used to acquire resource monitoring data and task status data of heterogeneous computing power environments.

[0011] The task resource fusion prediction module is used to input resource monitoring data and task status data into a preset dynamic cognitive resource model. Through the fusion framework configured by the dynamic cognitive resource model, it performs fine-grained prediction of the current resource demand and quantifies the interference effect characteristics between various tasks in a heterogeneous computing environment.

[0012] The dynamic scheduling decision intervention module is used to generate proactive prevention scheduling decisions based on fine-grained prediction results and interference effect characteristics, through a dynamic intervention module configured by a dynamic cognitive resource model.

[0013] The task scheduling strategy generation module is used for proactive prevention-based scheduling decisions. It outputs the target task scheduling strategy through a constrained hierarchical decision architecture configured by a dynamic cognitive resource model.

[0014] According to a third aspect, the present invention provides a computer device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform any of the task scheduling methods based on the dynamic cognitive resource model in the embodiments of the present invention.

[0015] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute any of the task scheduling methods based on a dynamic cognitive resource model in the embodiments of the present invention.

[0016] The present invention provides a task scheduling method based on a dynamic cognitive resource model, which is implemented through four core steps. These steps include: acquiring resource monitoring data and task status data from a heterogeneous computing environment to achieve full-dimensional data collection and provide basic input for dynamic scheduling; inputting the resource monitoring data and task status data into a preset dynamic cognitive resource model, and using a fusion framework configured in this model to perform fine-grained prediction of current resource demands and quantify the interference effect characteristics between tasks, thereby achieving accurate understanding of resource status and interference assessment; based on the fine-grained prediction results and interference effect characteristics, generating proactive scheduling decisions through a dynamic intervention module configured in the dynamic cognitive resource model, thereby achieving a paradigm shift from passive response to proactive intervention; and based on the proactive scheduling decisions, outputting a target task scheduling strategy through a constrained hierarchical decision architecture configured in the dynamic cognitive resource model, thereby ensuring the security and global optimization of the scheduling results.

[0017] In this technical solution, the present invention addresses the problem of insufficient adaptability to dynamic environments described in the background technology. By acquiring multi-dimensional monitoring data from heterogeneous computing environments and utilizing a fusion framework for fine-grained prediction and interference quantification, it solves the perception blind spot defect caused by traditional schedulers based on static thresholds or single-indicator monitoring, thus improving the system's responsiveness to dynamic loads. Regarding the difficulty in balancing global optimization and local security, the invention generates proactive prevention decisions through a dynamic intervention module and combines this with the output mechanism of a constrained hierarchical decision architecture, achieving effective synergy between optimization goals and business constraints, and avoiding stability risks in scheduling decisions. Therefore, the technical solution of this invention solves the technical problem of existing AI platform resource management technologies failing to achieve coordinated optimization of resource utilization efficiency and scheduling reliability, improving the adaptability, efficiency, and security of task scheduling in heterogeneous computing environments. Attached Figure Description

[0018] Figure 1This is a flowchart of a task scheduling method based on a dynamic cognitive resource model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a task scheduling system based on a dynamic cognitive resource model according to an embodiment of the present invention; Figure 3 This is a block diagram of a computer device for implementing embodiments of the present invention. Detailed Implementation

[0019] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0020] During the development of this invention, the inventors, through extensive experiments and data analysis, revealed the intrinsic relationship between traditional static scheduling methods and the dynamic characteristics of heterogeneous computing environments: traditional static scheduling methods not only struggle to adapt to sudden changes in resource demand, but also suffer from uneven resource allocation and localized hotspots due to a lack of system-level relational understanding. Based on this relationship, the inventors innovatively proposed this technical solution, utilizing a dynamic cognitive resource model to achieve fine-grained resource prediction and disturbance quantification through a fusion framework. Combined with a constrained hierarchical decision-making architecture, this achieves synergistic optimization of resource utilization efficiency and scheduling reliability, embodying the core concept of an integrated "perception-cognition-decision-intervention" system.

[0021] Specifically, through comparative experiments, the invention team discovered three major technical flaws in traditional scheduling methods: first, a blind spot in monitoring heterogeneous computing power metrics such as GPUs or NPUs at the resource awareness level; second, a passive response mode at the task interference handling level; and third, high trial-and-error costs and difficulty in coordinating with business constraints at the decision-making architecture level. These flaws result in an overall cluster resource utilization rate of less than 70%, and an interruption rate as high as 15% for long-cycle training tasks. However, the four-step collaborative processing mechanism proposed in this invention (data acquisition—prediction quantification—decision generation—policy output) can improve resource demand prediction accuracy to 93%, reduce task latency by 30%, and achieve a dual breakthrough in resource utilization and task stability.

[0022] Therefore, according to the first aspect, this invention provides a task scheduling method based on a dynamic cognitive resource model, which can be used in the computing resource scheduling system of an AI R&D platform (hereinafter referred to as the "system"). This system can be deployed locally or run on a heterogeneous computing cluster via cloud services to complete intelligent resource allocation and scheduling for large-scale machine learning tasks.

[0023] Specifically, this system can be deployed in various hardware environments, including but not limited to: GPU clusters, NPU accelerator card arrays, CPU-GPU hybrid architecture servers, and edge computing nodes. This flexible deployment architecture allows the system to meet both the high throughput requirements of centralized training tasks and the low latency requirements of distributed inference scenarios. In terms of its operational mechanism, the system achieves end-to-end intelligent scheduling through a multi-layered perception decision pipeline.

[0024] like Figure 1 As shown, the method may include: Step S110: Obtain resource monitoring data and task status data of the heterogeneous computing environment.

[0025] Among them, heterogeneous computing environment refers to a hybrid computing platform that includes multiple computing units (such as CPU, GPU, NPU) with different architectures and performance characteristics; resource monitoring data refers to a sequence of performance indicators collected in real time from computing nodes, including but not limited to CPU utilization, GPU memory usage, and network bandwidth utilization; task status data refers to relevant information of running tasks, such as task ID (identifier), resource requirements, execution status, and priority queue.

[0026] Specifically, the system can use intelligent sensing probes deployed on each node of the distributed computing cluster to collect resource indicators at fixed time intervals using a multi-task parallel collection algorithm, and then aggregate the data to the central scheduler through a message queue.

[0027] For example, in a heterogeneous cluster consisting of 10 computing nodes (including 5 GPU nodes and 5 CPU nodes), the system collects resource usage information every 5 seconds, generates time-series data packets, each data packet being approximately 2MB in size, and transmits them to the processing center via Kafka (a distributed stream processing platform).

[0028] Step S120: Input resource monitoring data and task status data into the preset dynamic cognitive resource model, perform fine-grained prediction of current resource demand through the fusion framework configured by the dynamic cognitive resource model, and quantify the interference effect characteristics between various tasks in the heterogeneous computing environment.

[0029] Among them, the dynamic cognitive resource model refers to a core component of intelligent scheduling based on a machine learning architecture, which can dynamically learn and infer resource usage patterns; the fusion framework refers to a collaborative processing architecture that combines multimodal temporal basic models and topology-aware graph learning, used to comprehensively analyze the relationship between temporal data and tasks; fine-grained prediction refers to high-precision, short-time-scale quantitative estimation of resource demand; and interference effect characteristics refer to the intensity index of mutual influence between tasks due to resource competition, represented in the form of scalar values.

[0030] Specifically, the system can extract long-term dependency features from resource monitoring data using a multimodal temporal model (such as a Transformer-based pre-trained network), and construct a dynamic task relationship graph using a topology-aware graph learning module (such as a graph attention network). End-to-end training is then used to jointly optimize prediction and quantization tasks. The Transformer-based pre-trained network refers to a neural network model pre-trained on large-scale unlabeled data using a Transformer architecture and a self-supervised learning task. The Transformer can transform the input sequence through a self-attention mechanism, thereby extracting deep features.

[0031] For example, the system takes into account monitoring data from the past hour (containing 1000 time points), predicts the GPU utilization curve for the next 10 minutes, achieves a prediction accuracy (in terms of mean absolute error) of 5%, and calculates the interference effect strength between task A and task B to be 0.75 (range 0~1), indicating a strong competitive relationship.

[0032] Step S130: Based on the fine-grained prediction results and the characteristics of the interference effect, generate proactive prevention scheduling decisions through the dynamic intervention module configured by the dynamic cognitive resource model.

[0033] Among them, the dynamic intervention module refers to a decision generation component that integrates online causal discovery and causal reinforcement learning, used to proactively adjust resource allocation; the proactive prevention scheduling decision refers to a resource allocation scheme based on forward-looking analysis, aimed at avoiding performance degradation.

[0034] Specifically, the system can apply safety perturbations (such as adjusting task CPU quotas) through the online causal discovery unit. The system observes the system response, constructs a dynamic causal graph using an incremental algorithm, and then performs counterfactual reasoning through a causal reinforcement learning agent to output optimized scheduling actions.

[0035] For example, when the module detects that task C may cause GPU memory contention, it actively schedules it to an idle node, reducing the task completion time by 20% and avoiding an increase in overall system latency.

[0036] Step S140: Based on proactive prevention scheduling decisions, the target task scheduling strategy is output through a constrained hierarchical decision architecture configured by a dynamic cognitive resource model.

[0037] Among them, the constrained hierarchical decision architecture refers to a framework that decomposes the decision-making process into multiple layers (such as the rule engine layer, the world model layer, and the decision generation layer) and integrates business constraints; the target task scheduling strategy refers to the final generated, executable resource allocation sequence that meets the requirements of security and efficiency.

[0038] Specifically, the system can transform business requirements into the safety boundaries of Lyapunov functions through the constraint formalization layer, simulate the decision-making effects at the world model layer, and generate compliant scheduling action sequences in an autoregressive manner using the decision Transformer model.

[0039] For example, the architecture outputs a scheduling policy that ensures high-priority tasks start within 2 seconds, keeps resource utilization below 85%, and generates a JSON (JavaScript Object Notation) policy file, approximately 50KB in size, which can be directly sent to the cluster scheduler for execution.

[0040] Therefore, according to the above implementation method, the system acquires resource monitoring data and task status data from heterogeneous computing environments to achieve full-dimensional data collection, providing basic input for dynamic scheduling; the resource monitoring data and task status data are input into a preset dynamic cognitive resource model, and the fusion framework configured in this model performs fine-grained prediction of current resource requirements and quantifies the interference effect characteristics between tasks to achieve accurate cognition of resource status and interference assessment; based on the fine-grained prediction results and interference effect characteristics, the dynamic intervention module configured in the dynamic cognitive resource model generates proactive prevention scheduling decisions to achieve a paradigm shift from passive response to proactive intervention; based on the proactive prevention scheduling decisions, the constrained hierarchical decision architecture configured in the dynamic cognitive resource model outputs the target task scheduling strategy to ensure the security and global optimization of scheduling results.

[0041] Specifically, in this implementation, addressing the issue of insufficient adaptability to dynamic environments mentioned in the background technology, the solution acquires multi-dimensional monitoring data from heterogeneous computing environments and utilizes a fusion framework for fine-grained prediction and interference quantification. This resolves the perception blind spots caused by traditional schedulers relying on static thresholds or single-indicator monitoring, thus improving the system's responsiveness to dynamic loads. Regarding the difficulty in balancing global optimization and local security, the solution generates proactive preventative decisions through a dynamic intervention module and combines this with the output mechanism of a constrained hierarchical decision architecture. This achieves effective synergy between optimization objectives and business constraints, avoiding stability risks in scheduling decisions. Therefore, this implementation's technical solution addresses the technical problem of existing AI platform resource management technologies failing to achieve coordinated optimization of resource utilization efficiency and scheduling reliability, improving the adaptability, efficiency, and security of task scheduling in heterogeneous computing environments.

[0042] In some embodiments, the fusion framework is configured with a multimodal temporal base model and a topology-aware graph learning module; the step of performing fine-grained prediction of current resource demand through the fusion framework configured with a dynamic cognitive resource model includes: The multivariate time series data from resource monitoring is input into the multimodal time series basic model. The multimodal time series basic model performs time series feature extraction and outputs the hidden state sequence.

[0043] Among them, multivariate time series refers to a multidimensional monitoring data set arranged in chronological order, with each time point containing synchronous observations of multiple resource indicators; hidden state sequence refers to a low-dimensional vector representation containing temporal dynamic features extracted through a deep learning model.

[0044] Specifically, the system uses a Transformer-based encoder network to capture long-term dependencies between different resource metrics using a self-attention mechanism, and utilizes residual connections to maintain stable gradient propagation.

[0045] For example, the system takes in monitoring data from the past 30 minutes (including 6 indicators such as GPU utilization, memory usage, and network bandwidth, totaling 360 time points) and outputs a hidden state vector sequence with a dimension of 128, with the sequence length maintained at 360 time steps.

[0046] Based on the hidden state sequence and the current task scheduling event, a dynamic task relationship graph is constructed. The dynamic task relationship graph uses task instances as nodes and resource competition relationships between tasks as edges.

[0047] Among them, task scheduling events refer to system state changes that trigger updates to the task relationship graph, including new task submission, task completion, and resource preemption; resource competition relationship refers to the intensity of concurrent requests from multiple tasks for the same type of computing resources (such as GPU memory).

[0048] Specifically, the system identifies resource dependencies between tasks by parsing cluster scheduling logs in real time, and dynamically adjusts edge weights based on resource pressure characteristics in the hidden state sequence.

[0049] For example, when it is detected that task A and task B simultaneously request the video memory resources of GPU node N3, a bidirectional edge is established in the graph, with the edge weight initialized to 0.8 (range 0~1), and updated every 5 seconds according to the real-time resource contention situation.

[0050] The topology-aware graph learning module performs graph attention network computation on the dynamic task relationship graph, aggregates the features of neighbor nodes, and quantifies the intensity of interference effects between task instances.

[0051] Among them, graph attention network computation refers to the information propagation method of dynamically weighting and aggregating the features of adjacent nodes in the graph through a learnable attention mechanism; interference effect strength refers to a quantitative index (range 0~1) obtained by graph neural network computation, which characterizes the degree of mutual influence between tasks.

[0052] Specifically, the system employs a multi-head graph attention layer to perform weighted summation of the neighbor features of each task node, and then uses a multilayer perceptron to map the aggregated features into a scalar value of interference intensity.

[0053] For example, for a relationship diagram containing 50 tasks, the module calculates that the interference effect strength between task C and task D is 0.6, which means that changes in the resource requirements of task C will increase the execution delay of task D by 60%.

[0054] Based on the hidden state sequence and the intensity of the interference effect, end-to-end training processing is performed on the resource demand prediction task and the interference effect quantification task to obtain fine-grained prediction results.

[0055] Among them, end-to-end training processing refers to the training strategy of jointly optimizing the model parameters by weighted summation of the loss functions of multiple sub-tasks; fine-grained prediction results refer to complete prediction data that includes the quantitative values ​​of resource demand at future time points and confidence intervals.

[0056] Specifically, the system adopts a multi-task learning framework, which combines the mean squared error loss of the prediction task with the cross-entropy loss of the interference quantization task with a weight of 1:0.5, and updates the parameters of the multimodal temporal base model and the topology-aware graph learning module synchronously through gradient backpropagation.

[0057] For example, after 100 rounds of model training, the average absolute error of resource demand prediction on the test set is less than 3%, the classification accuracy of interference effect intensity reaches 92%, and the final output is a GPU utilization prediction curve with a granularity of seconds for the next 10 minutes.

[0058] Therefore, according to the above implementation method, the system can achieve accurate prediction of resource demand and quantitative assessment of task interference through the synergistic effect of time series modeling and graph structure analysis, providing reliable data support for subsequent scheduling decisions. Specifically, this is manifested in: using a multimodal time series basic model to capture the dynamic patterns of resource usage; modeling complex competitive relationships between tasks through a dynamic task relationship graph; and employing end-to-end training to achieve joint optimization of prediction and interference assessment. This design ensures the system's rapid response capability in the face of sudden load fluctuations while maintaining the overall resource utilization efficiency of the cluster.

[0059] In some embodiments, the fusion framework is further configured with a dynamic causal graph construction module; the step of quantifying the interference effect characteristics between various tasks in a heterogeneous computing environment includes: The dynamic cause-effect graph construction module applies a safe micro-perturbation intervention operation to the performance index data of the system in the heterogeneous computing environment. The safe micro-perturbation intervention operation refers to the small adjustment of resource allocation parameters within a preset threshold range.

[0060] Among them, safety micro-disturbance intervention is a proactive causal detection technique that observes the causal response between system variables by applying small disturbances within the system's safe operating boundary; the preset threshold range refers to the safe range for resource adjustment determined based on statistical analysis of historical operating data, and is usually controlled within the normal value. Within.

[0061] Specifically, the system fine-tunes the CPU allocation quota or memory limit parameters of specific computing nodes through the resource controller interface, while monitoring changes in other relevant performance indicators.

[0062] For example, the system temporarily adjusts the CPU quota of Task A from the original 4 cores to 4.2 cores (an increase of 5%), observes it for 30 seconds, records the delay changes of Task B, and collects a total of 100 sets of intervention response data.

[0063] After applying a safety micro-perturbation intervention, in response to the system's response data, a dynamic causal graph is constructed using an incremental causal discovery algorithm configured by the dynamic causal graph construction module. The dynamic causal graph is a directed acyclic graph structure, and the edges of the dynamic causal graph represent the causal relationships between performance indicators.

[0064] Among them, the incremental causal discovery algorithm refers to an adaptive learning algorithm that can update the causal graph structure online. It gradually optimizes the graph structure through the conditional independence test of serialization. The directed acyclic graph structure refers to a graph model in which there are directional connections between nodes and no loops are formed. Each directed edge represents the causal influence direction from the antecedent node to the consequence node.

[0065] Specifically, the system employs a method that integrates the PC algorithm constraint framework with the scoring function optimization objective. It uses the G-test statistic to perform conditional independence testing, with a significance level set at 0.01. The PC algorithm constraint framework refers to a binding methodological framework in the field of causal discovery that uses the Peter-Clark (PC) algorithm. It progressively eliminates indirect associations between variables through conditional independence tests (such as the G-test) and combines this with the scoring function optimization objective to incrementally construct a directed acyclic graph (DAG), thereby accurately identifying causal relationships. The G-test statistic is a likelihood ratio test statistic used for hypothesis testing, suitable for comparing the significance of the difference between observed frequencies and expected frequencies. The formula for calculating this statistic is: ; in, For the frequency of observation, This refers to the expected frequency. Its core principle is to assess the correlation or distributional differences between variables by comparing the likelihood ratio of actual observed data with the theoretical expected distribution. In causal discovery algorithms, the G-test progressively eliminates indirect associations through a series of conditional independence tests (e.g., testing whether variables X and Y are independent under a given condition set Z), providing statistical evidence for constructing a dynamic causal graph. Its advantages lie in its insensitivity to sample size and its applicability to high-dimensional data scenarios. For example, after 100 iterations, the algorithm finally constructs a causal graph with 15 nodes and 28 edges, where the causal strength of the edge "GPU utilization → task latency" is 0.83.

[0066] The causal effect of critical paths in the dynamic causal graph is quantified by using the causal reasoning theory information configured in the dynamic causal graph construction module, and the average direct causal effect value of inter-task interference is obtained.

[0067] Among them, causal inference theory information refers to the formalized set of rules of do-calculus-based causal inference theory (a set of mathematical rules and calculation systems based on structural causal models proposed by Judea Pearl for estimating intervention effects from observational data), used to distinguish between real causal associations and spurious associations; the average direct causal effect value refers to the quantitative indicator of the average degree of influence of intervention operations on the target variable.

[0068] Specifically, the system calculates the average direct effect (ADE) of intervention variable X on outcome variable Y using the formula: "Conduct a quantitative assessment."

[0069] For example, the average direct effect of the priority of intervention task C on the completion time of task D was calculated to be... The seconds indicate that increasing the priority of task C will reduce the completion time of task D by an average of 2.3 seconds.

[0070] The average direct causal effect value and the interference effect intensity output by the topology-aware graph learning module are weighted and fused to generate interference effect features.

[0071] Among them, weighted fusion processing refers to a data integration method that linearly combines different weights based on the reliability of each input feature; interference effect features refer to the multi-dimensional feature vector that is finally generated and comprehensively reflects the interference relationship between tasks.

[0072] Specifically, the system employs an adaptive weight allocation mechanism, dynamically adjusting the fusion weights based on the confidence interval of the causal effect and the significance level of the graph attention weights. The weight coefficients are determined using the sigmoid function (a mathematical function with an S-shaped curve, expressed as follows). It can map any real number input to the (0,1) interval and is often used in neural networks as an activation function and as a weight normalization tool in weighted fusion processing to normalize to the [0,1] interval.

[0073] For example, the average direct causal effect value (weight 0.6) is fused with the interference intensity of the graph learning output (weight 0.4) to generate a feature vector with a dimension of 64, which is used as the input feature for downstream scheduling decisions.

[0074] Therefore, according to the above implementation method, the system can accurately identify the root causes of interference between tasks by combining proactive intervention with causal reasoning, and provide interpretable quantitative basis for scheduling decisions. Specifically, this is manifested in: safe micro-perturbation intervention ensures the stability of the system when probing causal relationships; incremental causal discovery enables real-time causal graph updates in dynamic environments; causal effect quantification provides a reliable assessment of interference intensity; and weighted fusion processing enhances the characterization capability of interference features. This design enables the system not only to identify surface correlations but also to reveal deep causal mechanisms, providing theoretical support for proactive preventative scheduling.

[0075] In some embodiments, the dynamic intervention module is configured with an online causal discovery unit and an online strategy optimization unit; based on fine-grained prediction results and interference effect characteristics, the dynamic intervention module, configured through a dynamic cognitive resource model, generates proactive prevention scheduling decisions, including: Based on fine-grained prediction results and disturbance effect characteristics, a causal-enhanced Markov decision process is constructed through a causal-enhanced decision unit to perform counterfactual reasoning operations and obtain the state transition probability distribution of the system state in response to intervention actions.

[0076] Among them, the causal-enhanced decision unit refers to an intelligent decision-making component that integrates causal graph structural information to enhance the traditional Markov decision process, used to simulate the potential impact of intervention actions; the causal-enhanced Markov decision process refers to an extended decision model that incorporates the topological features of dynamic causal graphs into the state-space representation; the counterfactual reasoning operation refers to the hypothetical analysis process of simulating "how the system state will change if a specific intervention is applied" by modifying variables in the causal graph; and the state transition probability distribution refers to the probability quantification representation of the system transitioning from the current state to a future state under a given intervention action.

[0077] Specifically, the system extracts the embedding features of the dynamic causal graph through a graph structure encoder, concatenates them with the system's instantaneous performance indicators to form an enhanced state representation, and then calculates the state transition probability caused by the intervention action through a state transition network.

[0078] For example, regarding GPU resource allocation intervention for task A, the system simulation adjusts the allocation from 8GB (gigabytes) to 9GB (gigabytes), predicts that the probability of a reduction in the completion time delay of task B is 80%, and generates a state transition probability matrix (10×10 dimension, representing the transition probability of 10 possible states).

[0079] Based on the state transition probability distribution, an online policy optimization unit learns historical scheduling policies online and uses the characteristics of disturbance effects as weight coefficients of the reward function to construct a reward calculation mechanism with the goal of reducing the total system latency and ensuring policy smoothness.

[0080] Among them, the online policy optimization unit refers to the adaptive optimization component that uses reinforcement learning algorithms to update the scheduling policy in real time; the reward calculation mechanism refers to the function design that dynamically calculates the reward value based on the system performance indicators to guide policy learning; and policy smoothness refers to the stability requirement that the change amplitude of scheduling actions should not be too large to avoid system oscillation.

[0081] Specifically, the system adopts the Proximal Policy Optimization (PPO) algorithm, updates the policy network parameters based on historical policy data, and uses interference effect characteristics (such as the intensity of inter-task interference) as weight factors in the reward function. The Proximal Policy Optimization (PPO) algorithm is a reinforcement learning algorithm used to train agents. Its core design goal is to achieve high sample utilization efficiency while ensuring training stability. This algorithm introduces a clipping mechanism to limit the policy update step size, effectively avoiding the training collapse problem caused by excessively large policy updates in traditional policy gradient methods. The reward function is designed as follows: ; Where α and β are weighting coefficients.

[0082] For example, the system collects historical scheduling data every 5 minutes, updates the policy network, and sets α=0.7 and β=0.3 in the reward function so that the system can reduce latency while ensuring that the rate of action change does not exceed 10%.

[0083] Based on counterfactual reasoning results, policy network parameters obtained through online learning, evaluation results of the reward calculation mechanism, and preset security constraints, the online policy optimization unit, after online learning and the construction of the reward calculation mechanism, outputs proactive prevention scheduling decisions.

[0084] Among them, the preset security constraints refer to the resource allocation boundaries set according to business needs, such as the maximum CPU utilization not exceeding 90%; the proactive scheduling decision refers to the resource configuration scheme generated based on forward-looking analysis, which aims to avoid performance degradation.

[0085] Specifically, the system integrates the state transition probabilities obtained from counterfactual reasoning, the output actions of the policy network, the reward value evaluation, and security constraints (such as Lyapunov function boundaries), and selects the optimal action that satisfies all conditions through the decision filtering module.

[0086] For example, the system outputs a scheduling decision: migrate task C to node N5 and limit its memory usage to no more than 16GB (gigabytes), which is expected to reduce the total system latency by 15% and keep resource utilization within a safe threshold.

[0087] Therefore, according to the above implementation method, the system can achieve a paradigm shift from passive response to proactive prevention in scheduling through the synergistic effect of causal-enhanced decision reasoning and online policy optimization. Specifically, this manifests as: using counterfactual reasoning to accurately predict intervention effects and avoid blind decision-making; dynamically adapting to environmental changes through online learning to improve the real-time performance of the strategy; and balancing performance optimization and system stability by combining reward mechanisms and security constraints. This design ensures high reliability and efficiency of scheduling decisions in complex heterogeneous environments.

[0088] In some embodiments, the causal reinforcement decision unit is configured with a graph structure encoder and a state transition network; based on fine-grained prediction results and disturbance effect characteristics, a causal reinforcement Markov decision process is constructed through the causal reinforcement decision unit to perform counterfactual reasoning operations and obtain the state transition probability distribution of the intervention action on the system state, including: The graph structure encoder extracts graph structure features from the dynamic causal graph, and encodes the graph node features and edge features into low-dimensional vector representations to obtain graph embedding feature vectors.

[0089] Among them, the graph structure encoder refers to the feature extraction component based on the graph neural network architecture, which is specifically used for distributed representation learning of graph structure data; the low-dimensional vector representation refers to the dimensionality reduction process that compresses high-dimensional sparse graph structure information into dense vectors, while retaining the main topological features of the original graph.

[0090] Specifically, the system employs a Graph Attention Network (GAT) as the encoder architecture, aggregating multi-hop neighbor node information through a multi-head attention mechanism and using layer normalization to ensure training stability. GAT (Graph Attention Network) is a graph neural network architecture based on an attention mechanism. Its core innovation lies in assigning differentiated importance to different neighbors of each node in the graph through adaptive attention weights, thereby achieving deep feature extraction from graph structure data.

[0091] For example, the system encodes a dynamic causal graph containing 50 nodes and 120 edges into a 128-dimensional feature vector. The encoding process takes about 15 milliseconds and the feature retention rate reaches 95%.

[0092] The graph embedding feature vector is concatenated with the instantaneous performance index of the system to construct an enhanced state representation, which is used as the state space of the Markov decision process for causal reinforcement.

[0093] Feature concatenation refers to a data fusion operation that connects feature vectors from different sources along a specific dimension; enhanced state representation refers to a composite feature representation that integrates graph structure information and real-time system state, providing a more comprehensive environmental perception for the decision-making process.

[0094] Specifically, the system concatenates 128-dimensional graph embedding features with system instantaneous performance metrics (including 10 metrics such as CPU utilization and memory usage) to generate a 138-dimensional enhanced state vector, and then unifies the dimensions through a fully connected layer.

[0095] For example, the system collects performance metrics every 2 seconds, concatenates them with the latest graph embedding features to generate a state representation, which serves as the input to the Markov decision process.

[0096] Based on the enhanced state representation, a state transition probability model is constructed through a state transition network. The state transition probability model is used to characterize the state evolution law of the system.

[0097] Among them, the state transition network refers to the dynamic model built on the deep neural network, which is used to predict the probability distribution of system state changes; the state transition probability model refers to the mathematical model that describes the state transition law of the system through the conditional probability distribution.

[0098] Specifically, the system employs a Long Short-Term Memory (LSTM) network as the core architecture of the state transition network. The input consists of the current state and the intervention action, and the output is the probability distribution of the next state. The model is trained using the cross-entropy loss function. LSTM refers to a special type of recurrent neural network (RNN) structure. Its core design goal is to solve the gradient vanishing or gradient exploding problems that traditional RNNs encounter when processing long-sequence data, thereby achieving effective modeling of long-term dependencies in time series.

[0099] For example, after 1,000 training sessions, the network achieved a state prediction accuracy of 92% on the test set, accurately predicting the changing trend of resource utilization under intervention actions.

[0100] Based on the state transition probability model and the counterfactual causal graph, the state transition probability distributions corresponding to multiple candidate intervention actions are obtained through parallel computation of the state transition network.

[0101] Among them, counterfactual causal graphs refer to hypothetical graph structures generated by intervening in specific nodes of dynamic causal graphs; parallel computing refers to an efficient computing mode that processes multiple candidate solutions simultaneously, greatly improving decision-making efficiency.

[0102] Specifically, the system simulates different intervention scenarios by modifying the node connections in the counterfactual causal graph, and uses the parallel computing power of the GPU (graphics processing unit) to evaluate the effects of multiple intervention actions simultaneously.

[0103] For example, the system can calculate the state transition probabilities of five candidate intervention actions (such as adjusting task priority, migrating computing nodes, etc.) in parallel, complete all calculations within 200 milliseconds, and select the optimal action to execute.

[0104] Therefore, according to the above implementation method, the system can achieve accurate counterfactual reasoning and efficient parallel decision evaluation through the deep integration of graph structure encoding and state transition modeling. Specifically, the graph structure encoder effectively extracts the topological features of the causal graph; the enhanced state representation integrates structural information and real-time state; the state transition network accurately predicts the system's evolution; and the parallel computing mechanism improves decision-making efficiency. This design enables the system to quickly evaluate the potential effects of various intervention schemes in complex heterogeneous environments, providing reliable technical support for proactive preventative scheduling. Experiments show that this implementation method can reduce the decision evaluation time from the traditional second level to the millisecond level while maintaining a prediction accuracy of over 90%.

[0105] In some embodiments, the constrained hierarchical decision architecture is configured with a constraint formalization layer, a world model layer, and a decision generation layer; based on proactive prevention-oriented scheduling decisions, the constrained hierarchical decision architecture configured through a dynamic cognitive resource model outputs a target task scheduling strategy, including: The constraint formalization layer transforms the current business constraints of the heterogeneous computing environment into safe boundary conditions of scalar functions.

[0106] In this context, business constraints refer to the system operation requirements defined by the Service Level Agreement (SLA), including task deadlines and resource availability guarantees; scalar function safety boundary conditions refer to constraints expressed in mathematical function form that define the scope of safe system operation. The SLA (Service Level Agreement) is a formally defined contractual document between the service provider and the customer, clearly quantifying key service quality indicators and agreeing on related rights and responsibilities. In this technical field, it serves as the core bridge for transforming business requirements into measurable and executable technical parameters.

[0107] Specifically, the system uses the Lyapunov function as a scalar function to transform business constraints into a safe threshold for the function value. When the Lyapunov function value remains within the threshold range, the system can be ensured to operate within a safe region.

[0108] For example, regarding task completion time constraints, the system sets a safety threshold of 50 (in milliseconds) for the Lyapunov function. When the predicted function value exceeds this threshold, a constraint alarm is triggered.

[0109] The proactive prevention scheduling decisions are input into the digital twin environment constructed by the world model layer, and the system state transition trajectory after the scheduling decisions are executed is simulated through the world model layer.

[0110] Among them, a digital twin environment refers to a virtual simulation environment trained with historical data that can accurately simulate the behavior of a real system; a system state transition trajectory refers to the continuous change path of the system state over time.

[0111] Specifically, the world model layer uses a deep neural network to construct a state transition function, taking the current state and scheduling decision as input, and outputting a system state prediction sequence for multiple future time steps.

[0112] For example, the system inputs the scheduling decision of "migrating task A to node N5" into the digital twin environment, simulates the change trajectory of system resource utilization in the next 10 minutes, and generates a state sequence containing 100 time points.

[0113] Based on the safety boundary conditions, the constraint satisfaction of the system state transition trajectory is verified through the decision generation layer.

[0114] Among them, constraint satisfaction verification refers to the inspection process of checking whether the system state trajectory always satisfies the safety boundary conditions.

[0115] Specifically, the system calculates the constraint violation index by comparing the Lyapunov function values ​​at each point in the state trajectory with the safety threshold. When the function values ​​at all time points are lower than the safety threshold, the system is determined to satisfy the constraint.

[0116] For example, the verification results show that the maximum Lyapunov function value in the state trajectory is 45 (in milliseconds), which is lower than the safety threshold of 50 (in milliseconds), and the trajectory is determined to meet the constraint requirements.

[0117] Based on the system state transition trajectory after constraint satisfaction verification, a scheduling action sequence that satisfies the safety boundary conditions is generated in an autoregressive manner through the decision generation layer, and the target task scheduling strategy is output according to the scheduling action sequence.

[0118] Among them, the autoregressive method refers to a sequence generation method that predicts subsequent elements step by step based on the generated sequence elements; the scheduling action sequence refers to a set of multiple scheduling instructions arranged in chronological order.

[0119] Specifically, the decision generation layer adopts the decision Transformer model, which uses historical states, reward signals and constraints as inputs to autoregressively generate the optimal scheduling action sequence.

[0120] For example, the system generates a sequence of 5 scheduling actions: [Action 1: Allocate GPU resources, Action 2: Adjust task priority, Action 3: Migrate computing tasks, Action 4: Balance load, Action 5: Release idle resources], and finally outputs a scheduling policy file in JSON format.

[0121] Therefore, according to the above implementation method, the system can generate the optimal scheduling strategy while ensuring safety constraints through the collaborative work of a three-layer architecture. Specifically, the constraint formalization layer transforms business requirements into computable mathematical boundaries; the world model layer provides high-fidelity simulation of decision-making effects; and the decision generation layer implements sequence optimization under safety constraints. This design enables the system to simultaneously consider scheduling performance and safety in complex dynamic environments. Experiments show that this implementation method can reduce the constraint violation rate to below 1% while maintaining a scheduling efficiency of over 95%.

[0122] In some embodiments, the scalar function is a Lyapunov function, and the current business constraints of the heterogeneous computing environment are transformed into safety boundary conditions of the scalar function through a constraint formalization layer, including: The system state variables corresponding to business constraints are identified through the constraint formalization layer.

[0123] Among them, system state variables refer to the set of key performance indicators used to describe the dynamic behavior of the system, including real-time observable parameters such as resource utilization, task latency, and queue length; business constraints refer to the system operation restrictions set according to service quality requirements, such as maximum task completion time and minimum resource guarantee.

[0124] Specifically, the system parses business constraint documents through a rules engine, maps abstract constraints to specific system monitoring metrics, and establishes a correspondence table between constraints and state variables.

[0125] For example, in response to the business constraint that "task completion time shall not exceed 10 seconds", the system identifies the corresponding system state variables, including CPU utilization (required to be less than 80%), memory usage (required to be less than 90%), and network latency (required to be less than 50 milliseconds).

[0126] Lyapunov functions are constructed based on system state variables, and the safety threshold of the Lyapunov functions is set according to the strictness of the business constraints. The safety threshold is used to define the safe operating area of ​​the system state.

[0127] Among them, the stringency characteristic refers to the tolerance level of business constraints, which is quantified through statistical analysis of historical violation data. For example, high stringency constraints correspond to a smaller security threshold; the safe operating area refers to the set of legal states in the system state space that satisfy all business constraints.

[0128] Specifically, the system constructs Lyapunov functions using quadratic form functions, such as... Where x is the system state vector and P is the weight matrix; the safety threshold is set according to the constraint violation probability, such as requiring the Lyapunov function value to always be below the threshold. .

[0129] For example, for the CPU utilization state variable, the safety threshold for the Lyapunov function is set to 100 (unit: dimensionless). When the function value exceeds this threshold, the system is determined to be in a dangerous state. The threshold for highly stringent constraints (such as financial trading tasks) is set to 80, while the threshold for ordinary tasks is set to 100.

[0130] Establish a comparison mapping relationship between Lyapunov function values ​​and safety thresholds to generate safety boundary conditions. These safety boundary conditions are used to determine whether the system trajectory is within a preset safety area.

[0131] Among them, the comparison mapping relationship refers to the logical judgment rule between the function value and the threshold expressed in the form of mathematical inequalities; the safety boundary condition refers to the specific constraint expression used to monitor the system status in real time.

[0132] Specifically, the system design comparison rule is: if If the condition is met, the system is safe; otherwise, an alert is triggered. This condition is encoded as a Boolean expression and integrated into the decision logic.

[0133] For example, the system calculates the Lyapunov function value every 2 seconds, compares it with a threshold, and generates a safety status flag (0 indicates safety, 1 indicates warning), which serves as input for scheduling decisions; when three consecutive warnings are detected... When this happens, a downgrade operation will be performed automatically.

[0134] Therefore, according to the above implementation method, the system can transform abstract business constraints into computable security boundaries through Lyapunov functions, realizing real-time monitoring and proactive protection of the system state. Specifically, this manifests in: accurately mapping business constraints to state variables to ensure comprehensive constraint coverage; adaptively setting security thresholds to adapt to business requirements of varying stringency; and establishing an efficient comparison mechanism to achieve millisecond-level security state judgment. This design enables the system to identify risks in advance in dynamic environments and avoid constraint violations. Experiments show that this implementation method can reduce the constraint violation rate to below 2% while maintaining system availability at 99.5%.

[0135] Figure 2 This is a structural block diagram of a task scheduling system based on a dynamic cognitive resource model according to an embodiment of the present invention.

[0136] like Figure 2 As shown, the task scheduling system based on the dynamic cognitive resource model includes: The task resource data acquisition module 210 is used to acquire resource monitoring data and task status data of the heterogeneous computing environment.

[0137] The task resource fusion prediction module 220 is used to input resource monitoring data and task status data into a preset dynamic cognitive resource model, perform fine-grained prediction of current resource demand through the fusion framework configured by the dynamic cognitive resource model, and quantify the interference effect characteristics between various tasks in the heterogeneous computing environment.

[0138] The dynamic scheduling decision intervention module 230 is used to generate proactive prevention scheduling decisions based on fine-grained prediction results and interference effect characteristics through a dynamic intervention module configured by a dynamic cognitive resource model.

[0139] The task scheduling strategy generation module 240 is used for proactive prevention-based scheduling decisions. It outputs the target task scheduling strategy through a constrained hierarchical decision architecture configured by a dynamic cognitive resource model.

[0140] The specific functions and examples of each module and submodule of the device in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0141] According to embodiments of the present invention, the above-described method of the present invention can be applied to a computer device and a readable storage medium.

[0142] Figure 3A schematic block diagram of an example computer device 600 that can be used to implement embodiments of the present invention is shown. The computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0143] like Figure 3 As shown, the computer device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the computer device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0144] Multiple components in computer device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows computer device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0145] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as a task scheduling method based on a dynamic cognitive resource model. For example, in some embodiments, a task scheduling method based on a dynamic cognitive resource model can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the computer device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the task scheduling method based on a dynamic cognitive resource model described above can be performed. Alternatively, in other embodiments, computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform a task scheduling method based on a dynamic cognitive resource model.

[0146] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0147] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0148] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0149] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0150] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0151] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0152] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0153] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for task scheduling based on a dynamic cognitive resource model, characterized in that, The method comprises the steps of: acquiring resource monitoring data and task state data of a heterogeneous computing power environment; inputting the resource monitoring data and the task state data into a preset dynamic cognitive resource model, performing fine-grained prediction on current resource demand through a fusion framework configured by the dynamic cognitive resource model, and quantitatively calculating interference effect characteristics between tasks in the heterogeneous computing power environment; based on the fine-grained prediction result and the interference effect characteristics, generating a proactive prevention scheduling decision through a dynamic intervention module configured by the dynamic cognitive resource model; based on the proactive prevention scheduling decision, outputting a target task scheduling strategy through a constrained hierarchical decision architecture configured by the dynamic cognitive resource model.

2. The method of claim 1, wherein, The fusion framework is configured with a multi-modal time series basic model and a topology-aware graph learning module; the step of performing fine-grained prediction on current resource demand through the fusion framework configured by the dynamic cognitive resource model comprises: inputting the multivariate time series in the resource monitoring data into the multi-modal time series basic model, performing time series feature extraction operation through the multi-modal time series basic model, and outputting a hidden state sequence; based on the hidden state sequence and the current task scheduling event, constructing a dynamic task relationship graph, the dynamic task relationship graph taking task instances as nodes and taking resource competition relationships between tasks as edges; performing graph attention network calculation on the dynamic task relationship graph through the topology-aware graph learning module, aggregating neighbor node features, and quantitatively calculating interference effect strength between task instances; based on the hidden state sequence and the interference effect strength, performing end-to-end training processing on resource demand prediction tasks and interference effect quantification tasks to obtain the fine-grained prediction result.

3. The method of claim 2, wherein, The fusion framework is also configured with a dynamic causal graph construction module; the step of quantitatively calculating interference effect characteristics between tasks in the heterogeneous computing power environment comprises: applying a safe micro-disturbance intervention operation to performance index data of system running in the heterogeneous computing power environment through the dynamic causal graph construction module, the safe micro-disturbance intervention operation referring to small-amplitude adjustment of resource allocation parameters within a preset threshold range; after applying the safe micro-disturbance intervention operation, in response to response data of the system, constructing a dynamic causal graph through an incremental causal discovery algorithm configured by the dynamic causal graph construction module, the dynamic causal graph being a directed acyclic graph structure, and the edges of the dynamic causal graph representing causal relationships between performance indexes; performing causal effect quantification calculation on a key path in the dynamic causal graph through causal reasoning theory information configured by the dynamic causal graph construction module to obtain an average direct causal effect value of task interference; performing weighted fusion processing on the average direct causal effect value and the interference effect strength output by the topology-aware graph learning module to generate the interference effect characteristics.

4. The method of claim 3, wherein, The dynamic intervention module is configured with an online causal discovery unit and an online strategy optimization unit; the step of generating a proactive prevention scheduling decision through the dynamic intervention module configured by the dynamic cognitive resource model based on the fine-grained prediction result and the interference effect characteristics comprises: construct a causal-enhanced Markov decision process through the causal-enhanced decision unit based on the fine-grained prediction result and the interference effect feature, to perform counterfactual reasoning operation, and obtain a state transition probability distribution of an intervention action on a system state; based on the state transition probability distribution, perform online learning on a historical scheduling strategy through the online policy optimization unit, and take the interference effect feature as a weight coefficient of a reward function, to construct a reward calculation mechanism aiming to reduce system total delay and guarantee policy smoothness; based on the counterfactual reasoning result, the policy network parameter obtained through online learning, the evaluation result of the reward calculation mechanism, and a preset safety constraint condition, output the proactive scheduling decision through the online policy optimization unit after online learning and construction of the reward calculation mechanism.

5. The method of claim 4, wherein, The causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the causal-enhanced decision unit is configured with a graph structure encoder and a state transition network; the ​ ​ ​ ​ 6. The method of claim 5, wherein, ​ ​ ​ ​ ​ 7. The method of claim 6, wherein, ​ The system state variable corresponding to the business constraint is identified through the constraint formalization layer; A Lyapunov function is constructed based on the system state variable, and a safety threshold of the Lyapunov function is set according to a strictness feature set by the business constraint, the safety threshold being used to define a safe operation region of the system state; A comparison mapping relationship between the Lyapunov function value and the safety threshold is established, and the safety boundary condition is generated, the safety boundary condition being used to judge whether a system trajectory is in a preset safe region.

8. A task scheduling system based on a dynamic cognitive resource model, characterized by, Comprise: A task resource data acquisition module configured to acquire resource monitoring data and task state data of a heterogeneous computing power environment; A task resource fusion prediction module configured to input the resource monitoring data and the task state data into a preset dynamic cognitive resource model, perform fine-grained prediction on current resource demand through a fusion framework configured by the dynamic cognitive resource model, and quantitatively calculate interference effect feature among each task in the heterogeneous computing power environment; A dynamic scheduling decision intervention module configured to generate a proactive prevention scheduling decision through a dynamic intervention module configured by the dynamic cognitive resource model based on the fine-grained prediction result and the interference effect feature; A task scheduling strategy generation module configured to output a target task scheduling strategy through a constrained hierarchical decision architecture configured by the dynamic cognitive resource model based on the proactive prevention scheduling decision.

9. A computer device comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, Computer instructions for causing a computer to perform the method of any one of claims 1-7.

Images